Show More
@@ -0,0 +1,69 b'' | |||
|
1 | """Tests for asyncresult.py""" | |
|
2 | ||
|
3 | #------------------------------------------------------------------------------- | |
|
4 | # Copyright (C) 2011 The IPython Development Team | |
|
5 | # | |
|
6 | # Distributed under the terms of the BSD License. The full license is in | |
|
7 | # the file COPYING, distributed as part of this software. | |
|
8 | #------------------------------------------------------------------------------- | |
|
9 | ||
|
10 | #------------------------------------------------------------------------------- | |
|
11 | # Imports | |
|
12 | #------------------------------------------------------------------------------- | |
|
13 | ||
|
14 | ||
|
15 | from IPython.zmq.parallel.error import TimeoutError | |
|
16 | ||
|
17 | from IPython.zmq.parallel.tests import add_engines | |
|
18 | from .clienttest import ClusterTestCase | |
|
19 | ||
|
20 | def setup(): | |
|
21 | add_engines(2) | |
|
22 | ||
|
23 | def wait(n): | |
|
24 | import time | |
|
25 | time.sleep(n) | |
|
26 | return n | |
|
27 | ||
|
28 | class AsyncResultTest(ClusterTestCase): | |
|
29 | ||
|
30 | def test_single_result(self): | |
|
31 | eid = self.client.ids[-1] | |
|
32 | ar = self.client[eid].apply_async(lambda : 42) | |
|
33 | self.assertEquals(ar.get(), 42) | |
|
34 | ar = self.client[[eid]].apply_async(lambda : 42) | |
|
35 | self.assertEquals(ar.get(), [42]) | |
|
36 | ar = self.client[-1:].apply_async(lambda : 42) | |
|
37 | self.assertEquals(ar.get(), [42]) | |
|
38 | ||
|
39 | def test_get_after_done(self): | |
|
40 | ar = self.client[-1].apply_async(lambda : 42) | |
|
41 | self.assertFalse(ar.ready()) | |
|
42 | ar.wait() | |
|
43 | self.assertTrue(ar.ready()) | |
|
44 | self.assertEquals(ar.get(), 42) | |
|
45 | self.assertEquals(ar.get(), 42) | |
|
46 | ||
|
47 | def test_get_before_done(self): | |
|
48 | ar = self.client[-1].apply_async(wait, 0.1) | |
|
49 | self.assertRaises(TimeoutError, ar.get, 0) | |
|
50 | ar.wait(0) | |
|
51 | self.assertFalse(ar.ready()) | |
|
52 | self.assertEquals(ar.get(), 0.1) | |
|
53 | ||
|
54 | def test_get_after_error(self): | |
|
55 | ar = self.client[-1].apply_async(lambda : 1/0) | |
|
56 | ar.wait() | |
|
57 | self.assertRaisesRemote(ZeroDivisionError, ar.get) | |
|
58 | self.assertRaisesRemote(ZeroDivisionError, ar.get) | |
|
59 | self.assertRaisesRemote(ZeroDivisionError, ar.get_dict) | |
|
60 | ||
|
61 | def test_get_dict(self): | |
|
62 | n = len(self.client) | |
|
63 | ar = self.client[:].apply_async(lambda : 5) | |
|
64 | self.assertEquals(ar.get(), [5]*n) | |
|
65 | d = ar.get_dict() | |
|
66 | self.assertEquals(sorted(d.keys()), sorted(self.client.ids)) | |
|
67 | for eid,r in d.iteritems(): | |
|
68 | self.assertEquals(r, 5) | |
|
69 |
@@ -0,0 +1,101 b'' | |||
|
1 | """Tests for dependency.py""" | |
|
2 | ||
|
3 | __docformat__ = "restructuredtext en" | |
|
4 | ||
|
5 | #------------------------------------------------------------------------------- | |
|
6 | # Copyright (C) 2011 The IPython Development Team | |
|
7 | # | |
|
8 | # Distributed under the terms of the BSD License. The full license is in | |
|
9 | # the file COPYING, distributed as part of this software. | |
|
10 | #------------------------------------------------------------------------------- | |
|
11 | ||
|
12 | #------------------------------------------------------------------------------- | |
|
13 | # Imports | |
|
14 | #------------------------------------------------------------------------------- | |
|
15 | ||
|
16 | # import | |
|
17 | import os | |
|
18 | ||
|
19 | from IPython.utils.pickleutil import can, uncan | |
|
20 | ||
|
21 | from IPython.zmq.parallel import dependency as dmod | |
|
22 | from IPython.zmq.parallel.util import interactive | |
|
23 | ||
|
24 | from IPython.zmq.parallel.tests import add_engines | |
|
25 | from .clienttest import ClusterTestCase | |
|
26 | ||
|
27 | def setup(): | |
|
28 | add_engines(1) | |
|
29 | ||
|
30 | @dmod.require('time') | |
|
31 | def wait(n): | |
|
32 | time.sleep(n) | |
|
33 | return n | |
|
34 | ||
|
35 | mixed = map(str, range(10)) | |
|
36 | completed = map(str, range(0,10,2)) | |
|
37 | failed = map(str, range(1,10,2)) | |
|
38 | ||
|
39 | class DependencyTest(ClusterTestCase): | |
|
40 | ||
|
41 | def setUp(self): | |
|
42 | ClusterTestCase.setUp(self) | |
|
43 | self.user_ns = {'__builtins__' : __builtins__} | |
|
44 | self.view = self.client.load_balanced_view() | |
|
45 | self.dview = self.client[-1] | |
|
46 | self.succeeded = set(map(str, range(0,25,2))) | |
|
47 | self.failed = set(map(str, range(1,25,2))) | |
|
48 | ||
|
49 | def assertMet(self, dep): | |
|
50 | self.assertTrue(dep.check(self.succeeded, self.failed), "Dependency should be met") | |
|
51 | ||
|
52 | def assertUnmet(self, dep): | |
|
53 | self.assertFalse(dep.check(self.succeeded, self.failed), "Dependency should not be met") | |
|
54 | ||
|
55 | def assertUnreachable(self, dep): | |
|
56 | self.assertTrue(dep.unreachable(self.succeeded, self.failed), "Dependency should be unreachable") | |
|
57 | ||
|
58 | def assertReachable(self, dep): | |
|
59 | self.assertFalse(dep.unreachable(self.succeeded, self.failed), "Dependency should be reachable") | |
|
60 | ||
|
61 | def cancan(self, f): | |
|
62 | """decorator to pass through canning into self.user_ns""" | |
|
63 | return uncan(can(f), self.user_ns) | |
|
64 | ||
|
65 | def test_require_imports(self): | |
|
66 | """test that @require imports names""" | |
|
67 | @self.cancan | |
|
68 | @dmod.require('urllib') | |
|
69 | @interactive | |
|
70 | def encode(dikt): | |
|
71 | return urllib.urlencode(dikt) | |
|
72 | # must pass through canning to properly connect namespaces | |
|
73 | self.assertEquals(encode(dict(a=5)), 'a=5') | |
|
74 | ||
|
75 | def test_success_only(self): | |
|
76 | dep = dmod.Dependency(mixed, success=True, failure=False) | |
|
77 | self.assertUnmet(dep) | |
|
78 | self.assertUnreachable(dep) | |
|
79 | dep.all=False | |
|
80 | self.assertMet(dep) | |
|
81 | self.assertReachable(dep) | |
|
82 | dep = dmod.Dependency(completed, success=True, failure=False) | |
|
83 | self.assertMet(dep) | |
|
84 | self.assertReachable(dep) | |
|
85 | dep.all=False | |
|
86 | self.assertMet(dep) | |
|
87 | self.assertReachable(dep) | |
|
88 | ||
|
89 | def test_failure_only(self): | |
|
90 | dep = dmod.Dependency(mixed, success=False, failure=True) | |
|
91 | self.assertUnmet(dep) | |
|
92 | self.assertUnreachable(dep) | |
|
93 | dep.all=False | |
|
94 | self.assertMet(dep) | |
|
95 | self.assertReachable(dep) | |
|
96 | dep = dmod.Dependency(completed, success=False, failure=True) | |
|
97 | self.assertUnmet(dep) | |
|
98 | self.assertUnreachable(dep) | |
|
99 | dep.all=False | |
|
100 | self.assertUnmet(dep) | |
|
101 | self.assertUnreachable(dep) |
@@ -0,0 +1,287 b'' | |||
|
1 | """test View objects""" | |
|
2 | #------------------------------------------------------------------------------- | |
|
3 | # Copyright (C) 2011 The IPython Development Team | |
|
4 | # | |
|
5 | # Distributed under the terms of the BSD License. The full license is in | |
|
6 | # the file COPYING, distributed as part of this software. | |
|
7 | #------------------------------------------------------------------------------- | |
|
8 | ||
|
9 | #------------------------------------------------------------------------------- | |
|
10 | # Imports | |
|
11 | #------------------------------------------------------------------------------- | |
|
12 | ||
|
13 | import time | |
|
14 | from tempfile import mktemp | |
|
15 | ||
|
16 | import zmq | |
|
17 | ||
|
18 | from IPython.zmq.parallel import client as clientmod | |
|
19 | from IPython.zmq.parallel import error | |
|
20 | from IPython.zmq.parallel.asyncresult import AsyncResult, AsyncHubResult, AsyncMapResult | |
|
21 | from IPython.zmq.parallel.view import LoadBalancedView, DirectView | |
|
22 | from IPython.zmq.parallel.util import interactive | |
|
23 | ||
|
24 | from IPython.zmq.parallel.tests import add_engines | |
|
25 | ||
|
26 | from .clienttest import ClusterTestCase, segfault, wait, skip_without | |
|
27 | ||
|
28 | def setup(): | |
|
29 | add_engines(3) | |
|
30 | ||
|
31 | class TestView(ClusterTestCase): | |
|
32 | ||
|
33 | def test_segfault_task(self): | |
|
34 | """test graceful handling of engine death (balanced)""" | |
|
35 | # self.add_engines(1) | |
|
36 | ar = self.client[-1].apply_async(segfault) | |
|
37 | self.assertRaisesRemote(error.EngineError, ar.get) | |
|
38 | eid = ar.engine_id | |
|
39 | while eid in self.client.ids: | |
|
40 | time.sleep(.01) | |
|
41 | self.client.spin() | |
|
42 | ||
|
43 | def test_segfault_mux(self): | |
|
44 | """test graceful handling of engine death (direct)""" | |
|
45 | # self.add_engines(1) | |
|
46 | eid = self.client.ids[-1] | |
|
47 | ar = self.client[eid].apply_async(segfault) | |
|
48 | self.assertRaisesRemote(error.EngineError, ar.get) | |
|
49 | eid = ar.engine_id | |
|
50 | while eid in self.client.ids: | |
|
51 | time.sleep(.01) | |
|
52 | self.client.spin() | |
|
53 | ||
|
54 | def test_push_pull(self): | |
|
55 | """test pushing and pulling""" | |
|
56 | data = dict(a=10, b=1.05, c=range(10), d={'e':(1,2),'f':'hi'}) | |
|
57 | t = self.client.ids[-1] | |
|
58 | v = self.client[t] | |
|
59 | push = v.push | |
|
60 | pull = v.pull | |
|
61 | v.block=True | |
|
62 | nengines = len(self.client) | |
|
63 | push({'data':data}) | |
|
64 | d = pull('data') | |
|
65 | self.assertEquals(d, data) | |
|
66 | self.client[:].push({'data':data}) | |
|
67 | d = self.client[:].pull('data', block=True) | |
|
68 | self.assertEquals(d, nengines*[data]) | |
|
69 | ar = push({'data':data}, block=False) | |
|
70 | self.assertTrue(isinstance(ar, AsyncResult)) | |
|
71 | r = ar.get() | |
|
72 | ar = self.client[:].pull('data', block=False) | |
|
73 | self.assertTrue(isinstance(ar, AsyncResult)) | |
|
74 | r = ar.get() | |
|
75 | self.assertEquals(r, nengines*[data]) | |
|
76 | self.client[:].push(dict(a=10,b=20)) | |
|
77 | r = self.client[:].pull(('a','b')) | |
|
78 | self.assertEquals(r, nengines*[[10,20]]) | |
|
79 | ||
|
80 | def test_push_pull_function(self): | |
|
81 | "test pushing and pulling functions" | |
|
82 | def testf(x): | |
|
83 | return 2.0*x | |
|
84 | ||
|
85 | t = self.client.ids[-1] | |
|
86 | self.client[t].block=True | |
|
87 | push = self.client[t].push | |
|
88 | pull = self.client[t].pull | |
|
89 | execute = self.client[t].execute | |
|
90 | push({'testf':testf}) | |
|
91 | r = pull('testf') | |
|
92 | self.assertEqual(r(1.0), testf(1.0)) | |
|
93 | execute('r = testf(10)') | |
|
94 | r = pull('r') | |
|
95 | self.assertEquals(r, testf(10)) | |
|
96 | ar = self.client[:].push({'testf':testf}, block=False) | |
|
97 | ar.get() | |
|
98 | ar = self.client[:].pull('testf', block=False) | |
|
99 | rlist = ar.get() | |
|
100 | for r in rlist: | |
|
101 | self.assertEqual(r(1.0), testf(1.0)) | |
|
102 | execute("def g(x): return x*x") | |
|
103 | r = pull(('testf','g')) | |
|
104 | self.assertEquals((r[0](10),r[1](10)), (testf(10), 100)) | |
|
105 | ||
|
106 | def test_push_function_globals(self): | |
|
107 | """test that pushed functions have access to globals""" | |
|
108 | @interactive | |
|
109 | def geta(): | |
|
110 | return a | |
|
111 | # self.add_engines(1) | |
|
112 | v = self.client[-1] | |
|
113 | v.block=True | |
|
114 | v['f'] = geta | |
|
115 | self.assertRaisesRemote(NameError, v.execute, 'b=f()') | |
|
116 | v.execute('a=5') | |
|
117 | v.execute('b=f()') | |
|
118 | self.assertEquals(v['b'], 5) | |
|
119 | ||
|
120 | def test_push_function_defaults(self): | |
|
121 | """test that pushed functions preserve default args""" | |
|
122 | def echo(a=10): | |
|
123 | return a | |
|
124 | v = self.client[-1] | |
|
125 | v.block=True | |
|
126 | v['f'] = echo | |
|
127 | v.execute('b=f()') | |
|
128 | self.assertEquals(v['b'], 10) | |
|
129 | ||
|
130 | def test_get_result(self): | |
|
131 | """test getting results from the Hub.""" | |
|
132 | c = clientmod.Client(profile='iptest') | |
|
133 | # self.add_engines(1) | |
|
134 | t = c.ids[-1] | |
|
135 | v = c[t] | |
|
136 | v2 = self.client[t] | |
|
137 | ar = v.apply_async(wait, 1) | |
|
138 | # give the monitor time to notice the message | |
|
139 | time.sleep(.25) | |
|
140 | ahr = v2.get_result(ar.msg_ids) | |
|
141 | self.assertTrue(isinstance(ahr, AsyncHubResult)) | |
|
142 | self.assertEquals(ahr.get(), ar.get()) | |
|
143 | ar2 = v2.get_result(ar.msg_ids) | |
|
144 | self.assertFalse(isinstance(ar2, AsyncHubResult)) | |
|
145 | c.spin() | |
|
146 | c.close() | |
|
147 | ||
|
148 | def test_run_newline(self): | |
|
149 | """test that run appends newline to files""" | |
|
150 | tmpfile = mktemp() | |
|
151 | with open(tmpfile, 'w') as f: | |
|
152 | f.write("""def g(): | |
|
153 | return 5 | |
|
154 | """) | |
|
155 | v = self.client[-1] | |
|
156 | v.run(tmpfile, block=True) | |
|
157 | self.assertEquals(v.apply_sync(lambda f: f(), clientmod.Reference('g')), 5) | |
|
158 | ||
|
159 | def test_apply_tracked(self): | |
|
160 | """test tracking for apply""" | |
|
161 | # self.add_engines(1) | |
|
162 | t = self.client.ids[-1] | |
|
163 | v = self.client[t] | |
|
164 | v.block=False | |
|
165 | def echo(n=1024*1024, **kwargs): | |
|
166 | with v.temp_flags(**kwargs): | |
|
167 | return v.apply(lambda x: x, 'x'*n) | |
|
168 | ar = echo(1, track=False) | |
|
169 | self.assertTrue(isinstance(ar._tracker, zmq.MessageTracker)) | |
|
170 | self.assertTrue(ar.sent) | |
|
171 | ar = echo(track=True) | |
|
172 | self.assertTrue(isinstance(ar._tracker, zmq.MessageTracker)) | |
|
173 | self.assertEquals(ar.sent, ar._tracker.done) | |
|
174 | ar._tracker.wait() | |
|
175 | self.assertTrue(ar.sent) | |
|
176 | ||
|
177 | def test_push_tracked(self): | |
|
178 | t = self.client.ids[-1] | |
|
179 | ns = dict(x='x'*1024*1024) | |
|
180 | v = self.client[t] | |
|
181 | ar = v.push(ns, block=False, track=False) | |
|
182 | self.assertTrue(isinstance(ar._tracker, zmq.MessageTracker)) | |
|
183 | self.assertTrue(ar.sent) | |
|
184 | ||
|
185 | ar = v.push(ns, block=False, track=True) | |
|
186 | self.assertTrue(isinstance(ar._tracker, zmq.MessageTracker)) | |
|
187 | self.assertEquals(ar.sent, ar._tracker.done) | |
|
188 | ar._tracker.wait() | |
|
189 | self.assertTrue(ar.sent) | |
|
190 | ar.get() | |
|
191 | ||
|
192 | def test_scatter_tracked(self): | |
|
193 | t = self.client.ids | |
|
194 | x='x'*1024*1024 | |
|
195 | ar = self.client[t].scatter('x', x, block=False, track=False) | |
|
196 | self.assertTrue(isinstance(ar._tracker, zmq.MessageTracker)) | |
|
197 | self.assertTrue(ar.sent) | |
|
198 | ||
|
199 | ar = self.client[t].scatter('x', x, block=False, track=True) | |
|
200 | self.assertTrue(isinstance(ar._tracker, zmq.MessageTracker)) | |
|
201 | self.assertEquals(ar.sent, ar._tracker.done) | |
|
202 | ar._tracker.wait() | |
|
203 | self.assertTrue(ar.sent) | |
|
204 | ar.get() | |
|
205 | ||
|
206 | def test_remote_reference(self): | |
|
207 | v = self.client[-1] | |
|
208 | v['a'] = 123 | |
|
209 | ra = clientmod.Reference('a') | |
|
210 | b = v.apply_sync(lambda x: x, ra) | |
|
211 | self.assertEquals(b, 123) | |
|
212 | ||
|
213 | ||
|
214 | def test_scatter_gather(self): | |
|
215 | view = self.client[:] | |
|
216 | seq1 = range(16) | |
|
217 | view.scatter('a', seq1) | |
|
218 | seq2 = view.gather('a', block=True) | |
|
219 | self.assertEquals(seq2, seq1) | |
|
220 | self.assertRaisesRemote(NameError, view.gather, 'asdf', block=True) | |
|
221 | ||
|
222 | @skip_without('numpy') | |
|
223 | def test_scatter_gather_numpy(self): | |
|
224 | import numpy | |
|
225 | from numpy.testing.utils import assert_array_equal, assert_array_almost_equal | |
|
226 | view = self.client[:] | |
|
227 | a = numpy.arange(64) | |
|
228 | view.scatter('a', a) | |
|
229 | b = view.gather('a', block=True) | |
|
230 | assert_array_equal(b, a) | |
|
231 | ||
|
232 | def test_map(self): | |
|
233 | view = self.client[:] | |
|
234 | def f(x): | |
|
235 | return x**2 | |
|
236 | data = range(16) | |
|
237 | r = view.map_sync(f, data) | |
|
238 | self.assertEquals(r, map(f, data)) | |
|
239 | ||
|
240 | def test_scatterGatherNonblocking(self): | |
|
241 | data = range(16) | |
|
242 | view = self.client[:] | |
|
243 | view.scatter('a', data, block=False) | |
|
244 | ar = view.gather('a', block=False) | |
|
245 | self.assertEquals(ar.get(), data) | |
|
246 | ||
|
247 | @skip_without('numpy') | |
|
248 | def test_scatter_gather_numpy_nonblocking(self): | |
|
249 | import numpy | |
|
250 | from numpy.testing.utils import assert_array_equal, assert_array_almost_equal | |
|
251 | a = numpy.arange(64) | |
|
252 | view = self.client[:] | |
|
253 | ar = view.scatter('a', a, block=False) | |
|
254 | self.assertTrue(isinstance(ar, AsyncResult)) | |
|
255 | amr = view.gather('a', block=False) | |
|
256 | self.assertTrue(isinstance(amr, AsyncMapResult)) | |
|
257 | assert_array_equal(amr.get(), a) | |
|
258 | ||
|
259 | def test_execute(self): | |
|
260 | view = self.client[:] | |
|
261 | # self.client.debug=True | |
|
262 | execute = view.execute | |
|
263 | ar = execute('c=30', block=False) | |
|
264 | self.assertTrue(isinstance(ar, AsyncResult)) | |
|
265 | ar = execute('d=[0,1,2]', block=False) | |
|
266 | self.client.wait(ar, 1) | |
|
267 | self.assertEquals(len(ar.get()), len(self.client)) | |
|
268 | for c in view['c']: | |
|
269 | self.assertEquals(c, 30) | |
|
270 | ||
|
271 | def test_abort(self): | |
|
272 | view = self.client[-1] | |
|
273 | ar = view.execute('import time; time.sleep(0.25)', block=False) | |
|
274 | ar2 = view.apply_async(lambda : 2) | |
|
275 | ar3 = view.apply_async(lambda : 3) | |
|
276 | view.abort(ar2) | |
|
277 | view.abort(ar3.msg_ids) | |
|
278 | self.assertRaises(error.TaskAborted, ar2.get) | |
|
279 | self.assertRaises(error.TaskAborted, ar3.get) | |
|
280 | ||
|
281 | def test_temp_flags(self): | |
|
282 | view = self.client[-1] | |
|
283 | view.block=True | |
|
284 | with view.temp_flags(block=False): | |
|
285 | self.assertFalse(view.block) | |
|
286 | self.assertTrue(view.block) | |
|
287 |
@@ -0,0 +1,97 b'' | |||
|
1 | """ | |
|
2 | An exceptionally lousy site spider | |
|
3 | Ken Kinder <ken@kenkinder.com> | |
|
4 | ||
|
5 | Updated for newparallel by Min Ragan-Kelley <benjaminrk@gmail.com> | |
|
6 | ||
|
7 | This module gives an example of how the task interface to the | |
|
8 | IPython controller works. Before running this script start the IPython controller | |
|
9 | and some engines using something like:: | |
|
10 | ||
|
11 | ipclusterz start -n 4 | |
|
12 | """ | |
|
13 | import sys | |
|
14 | from IPython.zmq.parallel import client, error | |
|
15 | import time | |
|
16 | import BeautifulSoup # this isn't necessary, but it helps throw the dependency error earlier | |
|
17 | ||
|
18 | def fetchAndParse(url, data=None): | |
|
19 | import urllib2 | |
|
20 | import urlparse | |
|
21 | import BeautifulSoup | |
|
22 | links = [] | |
|
23 | try: | |
|
24 | page = urllib2.urlopen(url, data=data) | |
|
25 | except Exception: | |
|
26 | return links | |
|
27 | else: | |
|
28 | if page.headers.type == 'text/html': | |
|
29 | doc = BeautifulSoup.BeautifulSoup(page.read()) | |
|
30 | for node in doc.findAll('a'): | |
|
31 | href = node.get('href', None) | |
|
32 | if href: | |
|
33 | links.append(urlparse.urljoin(url, href)) | |
|
34 | return links | |
|
35 | ||
|
36 | class DistributedSpider(object): | |
|
37 | ||
|
38 | # Time to wait between polling for task results. | |
|
39 | pollingDelay = 0.5 | |
|
40 | ||
|
41 | def __init__(self, site): | |
|
42 | self.client = client.Client() | |
|
43 | self.view = self.client.load_balanced_view() | |
|
44 | self.mux = self.client[:] | |
|
45 | ||
|
46 | self.allLinks = [] | |
|
47 | self.linksWorking = {} | |
|
48 | self.linksDone = {} | |
|
49 | ||
|
50 | self.site = site | |
|
51 | ||
|
52 | def visitLink(self, url): | |
|
53 | if url not in self.allLinks: | |
|
54 | self.allLinks.append(url) | |
|
55 | if url.startswith(self.site): | |
|
56 | print ' ', url | |
|
57 | self.linksWorking[url] = self.view.apply(fetchAndParse, url) | |
|
58 | ||
|
59 | def onVisitDone(self, links, url): | |
|
60 | print url, ':' | |
|
61 | self.linksDone[url] = None | |
|
62 | del self.linksWorking[url] | |
|
63 | for link in links: | |
|
64 | self.visitLink(link) | |
|
65 | ||
|
66 | def run(self): | |
|
67 | self.visitLink(self.site) | |
|
68 | while self.linksWorking: | |
|
69 | print len(self.linksWorking), 'pending...' | |
|
70 | self.synchronize() | |
|
71 | time.sleep(self.pollingDelay) | |
|
72 | ||
|
73 | def synchronize(self): | |
|
74 | for url, ar in self.linksWorking.items(): | |
|
75 | # Calling get_task_result with block=False will return None if the | |
|
76 | # task is not done yet. This provides a simple way of polling. | |
|
77 | try: | |
|
78 | links = ar.get(0) | |
|
79 | except error.TimeoutError: | |
|
80 | continue | |
|
81 | except Exception as e: | |
|
82 | self.linksDone[url] = None | |
|
83 | del self.linksWorking[url] | |
|
84 | print url, ':', e.traceback | |
|
85 | else: | |
|
86 | self.onVisitDone(links, url) | |
|
87 | ||
|
88 | def main(): | |
|
89 | if len(sys.argv) > 1: | |
|
90 | site = sys.argv[1] | |
|
91 | else: | |
|
92 | site = raw_input('Enter site to crawl: ') | |
|
93 | distributedSpider = DistributedSpider(site) | |
|
94 | distributedSpider.run() | |
|
95 | ||
|
96 | if __name__ == '__main__': | |
|
97 | main() |
@@ -0,0 +1,19 b'' | |||
|
1 | """ | |
|
2 | A Distributed Hello world | |
|
3 | Ken Kinder <ken@kenkinder.com> | |
|
4 | """ | |
|
5 | from IPython.zmq.parallel import client | |
|
6 | ||
|
7 | rc = client.Client() | |
|
8 | ||
|
9 | def sleep_and_echo(t, msg): | |
|
10 | import time | |
|
11 | time.sleep(t) | |
|
12 | return msg | |
|
13 | ||
|
14 | view = rc.load_balanced_view() | |
|
15 | ||
|
16 | world = view.apply_async(sleep_and_echo, 3, 'World!') | |
|
17 | hello = view.apply_async(sleep_and_echo, 2, 'Hello') | |
|
18 | print "Submitted tasks:", hello.msg_ids, world.msg_ids | |
|
19 | print hello.get(), world.get() |
@@ -1,143 +1,153 b'' | |||
|
1 | 1 | # encoding: utf-8 |
|
2 | 2 | |
|
3 | 3 | """Pickle related utilities. Perhaps this should be called 'can'.""" |
|
4 | 4 | |
|
5 | 5 | __docformat__ = "restructuredtext en" |
|
6 | 6 | |
|
7 | 7 | #------------------------------------------------------------------------------- |
|
8 | 8 | # Copyright (C) 2008 The IPython Development Team |
|
9 | 9 | # |
|
10 | 10 | # Distributed under the terms of the BSD License. The full license is in |
|
11 | 11 | # the file COPYING, distributed as part of this software. |
|
12 | 12 | #------------------------------------------------------------------------------- |
|
13 | 13 | |
|
14 | 14 | #------------------------------------------------------------------------------- |
|
15 | 15 | # Imports |
|
16 | 16 | #------------------------------------------------------------------------------- |
|
17 | 17 | |
|
18 | from types import FunctionType | |
|
19 | 18 | import copy |
|
20 | ||
|
21 | from IPython.zmq.parallel.dependency import dependent | |
|
19 | import sys | |
|
20 | from types import FunctionType | |
|
22 | 21 | |
|
23 | 22 | import codeutil |
|
24 | 23 | |
|
25 | 24 | #------------------------------------------------------------------------------- |
|
26 | 25 | # Classes |
|
27 | 26 | #------------------------------------------------------------------------------- |
|
28 | 27 | |
|
29 | 28 | |
|
30 | 29 | class CannedObject(object): |
|
31 | 30 | def __init__(self, obj, keys=[]): |
|
32 | 31 | self.keys = keys |
|
33 | 32 | self.obj = copy.copy(obj) |
|
34 | 33 | for key in keys: |
|
35 | 34 | setattr(self.obj, key, can(getattr(obj, key))) |
|
36 | 35 | |
|
37 | 36 | |
|
38 | 37 | def getObject(self, g=None): |
|
39 | 38 | if g is None: |
|
40 | 39 | g = globals() |
|
41 | 40 | for key in self.keys: |
|
42 | 41 | setattr(self.obj, key, uncan(getattr(self.obj, key), g)) |
|
43 | 42 | return self.obj |
|
44 | 43 | |
|
45 | 44 | class Reference(CannedObject): |
|
46 | 45 | """object for wrapping a remote reference by name.""" |
|
47 | 46 | def __init__(self, name): |
|
48 | 47 | if not isinstance(name, basestring): |
|
49 | 48 | raise TypeError("illegal name: %r"%name) |
|
50 | 49 | self.name = name |
|
51 | 50 | |
|
52 | 51 | def __repr__(self): |
|
53 | 52 | return "<Reference: %r>"%self.name |
|
54 | 53 | |
|
55 | 54 | def getObject(self, g=None): |
|
56 | 55 | if g is None: |
|
57 | 56 | g = globals() |
|
58 | 57 | try: |
|
59 | 58 | return g[self.name] |
|
60 | 59 | except KeyError: |
|
61 | 60 | raise NameError("name %r is not defined"%self.name) |
|
62 | 61 | |
|
63 | 62 | |
|
64 | 63 | class CannedFunction(CannedObject): |
|
65 | 64 | |
|
66 | 65 | def __init__(self, f): |
|
67 | 66 | self._checkType(f) |
|
68 | 67 | self.code = f.func_code |
|
69 | 68 | self.defaults = f.func_defaults |
|
69 | self.module = f.__module__ or '__main__' | |
|
70 | 70 | self.__name__ = f.__name__ |
|
71 | 71 | |
|
72 | 72 | def _checkType(self, obj): |
|
73 | 73 | assert isinstance(obj, FunctionType), "Not a function type" |
|
74 | 74 | |
|
75 | 75 | def getObject(self, g=None): |
|
76 | # try to load function back into its module: | |
|
77 | if not self.module.startswith('__'): | |
|
78 | try: | |
|
79 | __import__(self.module) | |
|
80 | except ImportError: | |
|
81 | pass | |
|
82 | else: | |
|
83 | g = sys.modules[self.module].__dict__ | |
|
84 | ||
|
76 | 85 | if g is None: |
|
77 | 86 | g = globals() |
|
78 | 87 | newFunc = FunctionType(self.code, g, self.__name__, self.defaults) |
|
79 | 88 | return newFunc |
|
80 | 89 | |
|
81 | 90 | #------------------------------------------------------------------------------- |
|
82 | 91 | # Functions |
|
83 | 92 | #------------------------------------------------------------------------------- |
|
84 | 93 | |
|
85 | ||
|
86 | 94 | def can(obj): |
|
95 | # import here to prevent module-level circular imports | |
|
96 | from IPython.zmq.parallel.dependency import dependent | |
|
87 | 97 | if isinstance(obj, dependent): |
|
88 | 98 | keys = ('f','df') |
|
89 | 99 | return CannedObject(obj, keys=keys) |
|
90 | 100 | elif isinstance(obj, FunctionType): |
|
91 | 101 | return CannedFunction(obj) |
|
92 | 102 | elif isinstance(obj,dict): |
|
93 | 103 | return canDict(obj) |
|
94 | 104 | elif isinstance(obj, (list,tuple)): |
|
95 | 105 | return canSequence(obj) |
|
96 | 106 | else: |
|
97 | 107 | return obj |
|
98 | 108 | |
|
99 | 109 | def canDict(obj): |
|
100 | 110 | if isinstance(obj, dict): |
|
101 | 111 | newobj = {} |
|
102 | 112 | for k, v in obj.iteritems(): |
|
103 | 113 | newobj[k] = can(v) |
|
104 | 114 | return newobj |
|
105 | 115 | else: |
|
106 | 116 | return obj |
|
107 | 117 | |
|
108 | 118 | def canSequence(obj): |
|
109 | 119 | if isinstance(obj, (list, tuple)): |
|
110 | 120 | t = type(obj) |
|
111 | 121 | return t([can(i) for i in obj]) |
|
112 | 122 | else: |
|
113 | 123 | return obj |
|
114 | 124 | |
|
115 | 125 | def uncan(obj, g=None): |
|
116 | 126 | if isinstance(obj, CannedObject): |
|
117 | 127 | return obj.getObject(g) |
|
118 | 128 | elif isinstance(obj,dict): |
|
119 | 129 | return uncanDict(obj, g) |
|
120 | 130 | elif isinstance(obj, (list,tuple)): |
|
121 | 131 | return uncanSequence(obj, g) |
|
122 | 132 | else: |
|
123 | 133 | return obj |
|
124 | 134 | |
|
125 | 135 | def uncanDict(obj, g=None): |
|
126 | 136 | if isinstance(obj, dict): |
|
127 | 137 | newobj = {} |
|
128 | 138 | for k, v in obj.iteritems(): |
|
129 | 139 | newobj[k] = uncan(v,g) |
|
130 | 140 | return newobj |
|
131 | 141 | else: |
|
132 | 142 | return obj |
|
133 | 143 | |
|
134 | 144 | def uncanSequence(obj, g=None): |
|
135 | 145 | if isinstance(obj, (list, tuple)): |
|
136 | 146 | t = type(obj) |
|
137 | 147 | return t([uncan(i,g) for i in obj]) |
|
138 | 148 | else: |
|
139 | 149 | return obj |
|
140 | 150 | |
|
141 | 151 | |
|
142 | 152 | def rebindFunctionGlobals(f, glbls): |
|
143 | 153 | return FunctionType(f.func_code, glbls) |
@@ -1,322 +1,340 b'' | |||
|
1 | 1 | """AsyncResult objects for the client""" |
|
2 | 2 | #----------------------------------------------------------------------------- |
|
3 | 3 | # Copyright (C) 2010-2011 The IPython Development Team |
|
4 | 4 | # |
|
5 | 5 | # Distributed under the terms of the BSD License. The full license is in |
|
6 | 6 | # the file COPYING, distributed as part of this software. |
|
7 | 7 | #----------------------------------------------------------------------------- |
|
8 | 8 | |
|
9 | 9 | #----------------------------------------------------------------------------- |
|
10 | 10 | # Imports |
|
11 | 11 | #----------------------------------------------------------------------------- |
|
12 | 12 | |
|
13 | 13 | import time |
|
14 | 14 | |
|
15 | from zmq import MessageTracker | |
|
16 | ||
|
15 | 17 | from IPython.external.decorator import decorator |
|
16 | 18 | from . import error |
|
17 | 19 | |
|
18 | 20 | #----------------------------------------------------------------------------- |
|
19 | 21 | # Classes |
|
20 | 22 | #----------------------------------------------------------------------------- |
|
21 | 23 | |
|
24 | # global empty tracker that's always done: | |
|
25 | finished_tracker = MessageTracker() | |
|
26 | ||
|
22 | 27 | @decorator |
|
23 | 28 | def check_ready(f, self, *args, **kwargs): |
|
24 | 29 | """Call spin() to sync state prior to calling the method.""" |
|
25 | 30 | self.wait(0) |
|
26 | 31 | if not self._ready: |
|
27 | 32 | raise error.TimeoutError("result not ready") |
|
28 | 33 | return f(self, *args, **kwargs) |
|
29 | 34 | |
|
30 | 35 | class AsyncResult(object): |
|
31 | 36 | """Class for representing results of non-blocking calls. |
|
32 | 37 | |
|
33 | 38 | Provides the same interface as :py:class:`multiprocessing.pool.AsyncResult`. |
|
34 | 39 | """ |
|
35 | 40 | |
|
36 | 41 | msg_ids = None |
|
37 | 42 | _targets = None |
|
38 | 43 | _tracker = None |
|
44 | _single_result = False | |
|
39 | 45 | |
|
40 | 46 | def __init__(self, client, msg_ids, fname='unknown', targets=None, tracker=None): |
|
41 | self._client = client | |
|
42 | 47 | if isinstance(msg_ids, basestring): |
|
48 | # always a list | |
|
43 | 49 | msg_ids = [msg_ids] |
|
50 | if tracker is None: | |
|
51 | # default to always done | |
|
52 | tracker = finished_tracker | |
|
53 | self._client = client | |
|
44 | 54 | self.msg_ids = msg_ids |
|
45 | 55 | self._fname=fname |
|
46 | 56 | self._targets = targets |
|
47 | 57 | self._tracker = tracker |
|
48 | 58 | self._ready = False |
|
49 | 59 | self._success = None |
|
50 |
|
|
|
60 | if len(msg_ids) == 1: | |
|
61 | self._single_result = not isinstance(targets, (list, tuple)) | |
|
62 | else: | |
|
63 | self._single_result = False | |
|
51 | 64 | |
|
52 | 65 | def __repr__(self): |
|
53 | 66 | if self._ready: |
|
54 | 67 | return "<%s: finished>"%(self.__class__.__name__) |
|
55 | 68 | else: |
|
56 | 69 | return "<%s: %s>"%(self.__class__.__name__,self._fname) |
|
57 | 70 | |
|
58 | 71 | |
|
59 | 72 | def _reconstruct_result(self, res): |
|
60 | 73 | """Reconstruct our result from actual result list (always a list) |
|
61 | 74 | |
|
62 | 75 | Override me in subclasses for turning a list of results |
|
63 | 76 | into the expected form. |
|
64 | 77 | """ |
|
65 | 78 | if self._single_result: |
|
66 | 79 | return res[0] |
|
67 | 80 | else: |
|
68 | 81 | return res |
|
69 | 82 | |
|
70 | 83 | def get(self, timeout=-1): |
|
71 | 84 | """Return the result when it arrives. |
|
72 | 85 | |
|
73 | 86 | If `timeout` is not ``None`` and the result does not arrive within |
|
74 | 87 | `timeout` seconds then ``TimeoutError`` is raised. If the |
|
75 | 88 | remote call raised an exception then that exception will be reraised |
|
76 | 89 | by get() inside a `RemoteError`. |
|
77 | 90 | """ |
|
78 | 91 | if not self.ready(): |
|
79 | 92 | self.wait(timeout) |
|
80 | 93 | |
|
81 | 94 | if self._ready: |
|
82 | 95 | if self._success: |
|
83 | 96 | return self._result |
|
84 | 97 | else: |
|
85 | 98 | raise self._exception |
|
86 | 99 | else: |
|
87 | 100 | raise error.TimeoutError("Result not ready.") |
|
88 | 101 | |
|
89 | 102 | def ready(self): |
|
90 | 103 | """Return whether the call has completed.""" |
|
91 | 104 | if not self._ready: |
|
92 | 105 | self.wait(0) |
|
93 | 106 | return self._ready |
|
94 | 107 | |
|
95 | 108 | def wait(self, timeout=-1): |
|
96 | 109 | """Wait until the result is available or until `timeout` seconds pass. |
|
97 | 110 | |
|
98 | 111 | This method always returns None. |
|
99 | 112 | """ |
|
100 | 113 | if self._ready: |
|
101 | 114 | return |
|
102 |
self._ready = self._client. |
|
|
115 | self._ready = self._client.wait(self.msg_ids, timeout) | |
|
103 | 116 | if self._ready: |
|
104 | 117 | try: |
|
105 | 118 | results = map(self._client.results.get, self.msg_ids) |
|
106 | 119 | self._result = results |
|
107 | 120 | if self._single_result: |
|
108 | 121 | r = results[0] |
|
109 | 122 | if isinstance(r, Exception): |
|
110 | 123 | raise r |
|
111 | 124 | else: |
|
112 | 125 | results = error.collect_exceptions(results, self._fname) |
|
113 | 126 | self._result = self._reconstruct_result(results) |
|
114 | 127 | except Exception, e: |
|
115 | 128 | self._exception = e |
|
116 | 129 | self._success = False |
|
117 | 130 | else: |
|
118 | 131 | self._success = True |
|
119 | 132 | finally: |
|
120 | 133 | self._metadata = map(self._client.metadata.get, self.msg_ids) |
|
121 | 134 | |
|
122 | 135 | |
|
123 | 136 | def successful(self): |
|
124 | 137 | """Return whether the call completed without raising an exception. |
|
125 | 138 | |
|
126 | 139 | Will raise ``AssertionError`` if the result is not ready. |
|
127 | 140 | """ |
|
128 | 141 | assert self.ready() |
|
129 | 142 | return self._success |
|
130 | 143 | |
|
131 | 144 | #---------------------------------------------------------------- |
|
132 | 145 | # Extra methods not in mp.pool.AsyncResult |
|
133 | 146 | #---------------------------------------------------------------- |
|
134 | 147 | |
|
135 | 148 | def get_dict(self, timeout=-1): |
|
136 | 149 | """Get the results as a dict, keyed by engine_id. |
|
137 | 150 | |
|
138 | 151 | timeout behavior is described in `get()`. |
|
139 | 152 | """ |
|
140 | 153 | |
|
141 | 154 | results = self.get(timeout) |
|
142 | 155 | engine_ids = [ md['engine_id'] for md in self._metadata ] |
|
143 | 156 | bycount = sorted(engine_ids, key=lambda k: engine_ids.count(k)) |
|
144 | 157 | maxcount = bycount.count(bycount[-1]) |
|
145 | 158 | if maxcount > 1: |
|
146 | 159 | raise ValueError("Cannot build dict, %i jobs ran on engine #%i"%( |
|
147 | 160 | maxcount, bycount[-1])) |
|
148 | 161 | |
|
149 | 162 | return dict(zip(engine_ids,results)) |
|
150 | 163 | |
|
151 | 164 | @property |
|
152 | @check_ready | |
|
153 | 165 | def result(self): |
|
154 | 166 | """result property wrapper for `get(timeout=0)`.""" |
|
155 |
return self. |
|
|
167 | return self.get() | |
|
156 | 168 | |
|
157 | 169 | # abbreviated alias: |
|
158 | 170 | r = result |
|
159 | 171 | |
|
160 | 172 | @property |
|
161 | 173 | @check_ready |
|
162 | 174 | def metadata(self): |
|
163 | 175 | """property for accessing execution metadata.""" |
|
164 | 176 | if self._single_result: |
|
165 | 177 | return self._metadata[0] |
|
166 | 178 | else: |
|
167 | 179 | return self._metadata |
|
168 | 180 | |
|
169 | 181 | @property |
|
170 | 182 | def result_dict(self): |
|
171 | 183 | """result property as a dict.""" |
|
172 |
return self.get_dict( |
|
|
184 | return self.get_dict() | |
|
173 | 185 | |
|
174 | 186 | def __dict__(self): |
|
175 | 187 | return self.get_dict(0) |
|
176 | 188 | |
|
177 | 189 | def abort(self): |
|
178 | 190 | """abort my tasks.""" |
|
179 | 191 | assert not self.ready(), "Can't abort, I am already done!" |
|
180 | 192 | return self.client.abort(self.msg_ids, targets=self._targets, block=True) |
|
181 | 193 | |
|
182 | 194 | @property |
|
183 | 195 | def sent(self): |
|
184 | """check whether my messages have been sent""" | |
|
185 | if self._tracker is None: | |
|
186 | return True | |
|
187 | else: | |
|
196 | """check whether my messages have been sent.""" | |
|
188 | 197 |
|
|
189 | 198 | |
|
199 | def wait_for_send(self, timeout=-1): | |
|
200 | """wait for pyzmq send to complete. | |
|
201 | ||
|
202 | This is necessary when sending arrays that you intend to edit in-place. | |
|
203 | `timeout` is in seconds, and will raise TimeoutError if it is reached | |
|
204 | before the send completes. | |
|
205 | """ | |
|
206 | return self._tracker.wait(timeout) | |
|
207 | ||
|
190 | 208 | #------------------------------------- |
|
191 | 209 | # dict-access |
|
192 | 210 | #------------------------------------- |
|
193 | 211 | |
|
194 | 212 | @check_ready |
|
195 | 213 | def __getitem__(self, key): |
|
196 | 214 | """getitem returns result value(s) if keyed by int/slice, or metadata if key is str. |
|
197 | 215 | """ |
|
198 | 216 | if isinstance(key, int): |
|
199 | 217 | return error.collect_exceptions([self._result[key]], self._fname)[0] |
|
200 | 218 | elif isinstance(key, slice): |
|
201 | 219 | return error.collect_exceptions(self._result[key], self._fname) |
|
202 | 220 | elif isinstance(key, basestring): |
|
203 | 221 | values = [ md[key] for md in self._metadata ] |
|
204 | 222 | if self._single_result: |
|
205 | 223 | return values[0] |
|
206 | 224 | else: |
|
207 | 225 | return values |
|
208 | 226 | else: |
|
209 | 227 | raise TypeError("Invalid key type %r, must be 'int','slice', or 'str'"%type(key)) |
|
210 | 228 | |
|
211 | 229 | @check_ready |
|
212 | 230 | def __getattr__(self, key): |
|
213 | 231 | """getattr maps to getitem for convenient attr access to metadata.""" |
|
214 | 232 | if key not in self._metadata[0].keys(): |
|
215 | 233 | raise AttributeError("%r object has no attribute %r"%( |
|
216 | 234 | self.__class__.__name__, key)) |
|
217 | 235 | return self.__getitem__(key) |
|
218 | 236 | |
|
219 | 237 | # asynchronous iterator: |
|
220 | 238 | def __iter__(self): |
|
221 | 239 | if self._single_result: |
|
222 | 240 | raise TypeError("AsyncResults with a single result are not iterable.") |
|
223 | 241 | try: |
|
224 | 242 | rlist = self.get(0) |
|
225 | 243 | except error.TimeoutError: |
|
226 | 244 | # wait for each result individually |
|
227 | 245 | for msg_id in self.msg_ids: |
|
228 | 246 | ar = AsyncResult(self._client, msg_id, self._fname) |
|
229 | 247 | yield ar.get() |
|
230 | 248 | else: |
|
231 | 249 | # already done |
|
232 | 250 | for r in rlist: |
|
233 | 251 | yield r |
|
234 | 252 | |
|
235 | 253 | |
|
236 | 254 | |
|
237 | 255 | class AsyncMapResult(AsyncResult): |
|
238 | 256 | """Class for representing results of non-blocking gathers. |
|
239 | 257 | |
|
240 | 258 | This will properly reconstruct the gather. |
|
241 | 259 | """ |
|
242 | 260 | |
|
243 | 261 | def __init__(self, client, msg_ids, mapObject, fname=''): |
|
244 | 262 | AsyncResult.__init__(self, client, msg_ids, fname=fname) |
|
245 | 263 | self._mapObject = mapObject |
|
246 | 264 | self._single_result = False |
|
247 | 265 | |
|
248 | 266 | def _reconstruct_result(self, res): |
|
249 | 267 | """Perform the gather on the actual results.""" |
|
250 | 268 | return self._mapObject.joinPartitions(res) |
|
251 | 269 | |
|
252 | 270 | # asynchronous iterator: |
|
253 | 271 | def __iter__(self): |
|
254 | 272 | try: |
|
255 | 273 | rlist = self.get(0) |
|
256 | 274 | except error.TimeoutError: |
|
257 | 275 | # wait for each result individually |
|
258 | 276 | for msg_id in self.msg_ids: |
|
259 | 277 | ar = AsyncResult(self._client, msg_id, self._fname) |
|
260 | 278 | rlist = ar.get() |
|
261 | 279 | try: |
|
262 | 280 | for r in rlist: |
|
263 | 281 | yield r |
|
264 | 282 | except TypeError: |
|
265 | 283 | # flattened, not a list |
|
266 | 284 | # this could get broken by flattened data that returns iterables |
|
267 | 285 | # but most calls to map do not expose the `flatten` argument |
|
268 | 286 | yield rlist |
|
269 | 287 | else: |
|
270 | 288 | # already done |
|
271 | 289 | for r in rlist: |
|
272 | 290 | yield r |
|
273 | 291 | |
|
274 | 292 | |
|
275 | 293 | class AsyncHubResult(AsyncResult): |
|
276 | 294 | """Class to wrap pending results that must be requested from the Hub. |
|
277 | 295 | |
|
278 | 296 | Note that waiting/polling on these objects requires polling the Hubover the network, |
|
279 | 297 | so use `AsyncHubResult.wait()` sparingly. |
|
280 | 298 | """ |
|
281 | 299 | |
|
282 | 300 | def wait(self, timeout=-1): |
|
283 | 301 | """wait for result to complete.""" |
|
284 | 302 | start = time.time() |
|
285 | 303 | if self._ready: |
|
286 | 304 | return |
|
287 | 305 | local_ids = filter(lambda msg_id: msg_id in self._client.outstanding, self.msg_ids) |
|
288 |
local_ready = self._client. |
|
|
306 | local_ready = self._client.wait(local_ids, timeout) | |
|
289 | 307 | if local_ready: |
|
290 | 308 | remote_ids = filter(lambda msg_id: msg_id not in self._client.results, self.msg_ids) |
|
291 | 309 | if not remote_ids: |
|
292 | 310 | self._ready = True |
|
293 | 311 | else: |
|
294 | 312 | rdict = self._client.result_status(remote_ids, status_only=False) |
|
295 | 313 | pending = rdict['pending'] |
|
296 | 314 | while pending and (timeout < 0 or time.time() < start+timeout): |
|
297 | 315 | rdict = self._client.result_status(remote_ids, status_only=False) |
|
298 | 316 | pending = rdict['pending'] |
|
299 | 317 | if pending: |
|
300 | 318 | time.sleep(0.1) |
|
301 | 319 | if not pending: |
|
302 | 320 | self._ready = True |
|
303 | 321 | if self._ready: |
|
304 | 322 | try: |
|
305 | 323 | results = map(self._client.results.get, self.msg_ids) |
|
306 | 324 | self._result = results |
|
307 | 325 | if self._single_result: |
|
308 | 326 | r = results[0] |
|
309 | 327 | if isinstance(r, Exception): |
|
310 | 328 | raise r |
|
311 | 329 | else: |
|
312 | 330 | results = error.collect_exceptions(results, self._fname) |
|
313 | 331 | self._result = self._reconstruct_result(results) |
|
314 | 332 | except Exception, e: |
|
315 | 333 | self._exception = e |
|
316 | 334 | self._success = False |
|
317 | 335 | else: |
|
318 | 336 | self._success = True |
|
319 | 337 | finally: |
|
320 | 338 | self._metadata = map(self._client.metadata.get, self.msg_ids) |
|
321 | 339 | |
|
322 | 340 | __all__ = ['AsyncResult', 'AsyncMapResult', 'AsyncHubResult'] No newline at end of file |
This diff has been collapsed as it changes many lines, (632 lines changed) Show them Hide them | |||
@@ -1,1591 +1,1343 b'' | |||
|
1 |
"""A semi-synchronous Client for the ZMQ c |
|
|
1 | """A semi-synchronous Client for the ZMQ cluster""" | |
|
2 | 2 | #----------------------------------------------------------------------------- |
|
3 | 3 | # Copyright (C) 2010 The IPython Development Team |
|
4 | 4 | # |
|
5 | 5 | # Distributed under the terms of the BSD License. The full license is in |
|
6 | 6 | # the file COPYING, distributed as part of this software. |
|
7 | 7 | #----------------------------------------------------------------------------- |
|
8 | 8 | |
|
9 | 9 | #----------------------------------------------------------------------------- |
|
10 | 10 | # Imports |
|
11 | 11 | #----------------------------------------------------------------------------- |
|
12 | 12 | |
|
13 | 13 | import os |
|
14 | 14 | import json |
|
15 | 15 | import time |
|
16 | 16 | import warnings |
|
17 | 17 | from datetime import datetime |
|
18 | 18 | from getpass import getpass |
|
19 | 19 | from pprint import pprint |
|
20 | 20 | |
|
21 | 21 | pjoin = os.path.join |
|
22 | 22 | |
|
23 | 23 | import zmq |
|
24 | 24 | # from zmq.eventloop import ioloop, zmqstream |
|
25 | 25 | |
|
26 | 26 | from IPython.utils.path import get_ipython_dir |
|
27 | 27 | from IPython.utils.pickleutil import Reference |
|
28 | 28 | from IPython.utils.traitlets import (HasTraits, Int, Instance, CUnicode, |
|
29 | 29 | Dict, List, Bool, Str, Set) |
|
30 | 30 | from IPython.external.decorator import decorator |
|
31 | 31 | from IPython.external.ssh import tunnel |
|
32 | 32 | |
|
33 | 33 | from . import error |
|
34 | from . import map as Map | |
|
35 | 34 | from . import util |
|
36 | 35 | from . import streamsession as ss |
|
37 | 36 | from .asyncresult import AsyncResult, AsyncMapResult, AsyncHubResult |
|
38 | 37 | from .clusterdir import ClusterDir, ClusterDirError |
|
39 | 38 | from .dependency import Dependency, depend, require, dependent |
|
40 | 39 | from .remotefunction import remote, parallel, ParallelFunction, RemoteFunction |
|
41 | from .util import ReverseDict, validate_url, disambiguate_url | |
|
42 | 40 | from .view import DirectView, LoadBalancedView |
|
43 | 41 | |
|
44 | 42 | #-------------------------------------------------------------------------- |
|
45 | # helpers for implementing old MEC API via client.apply | |
|
46 | #-------------------------------------------------------------------------- | |
|
47 | ||
|
48 | def _push(user_ns, **ns): | |
|
49 | """helper method for implementing `client.push` via `client.apply`""" | |
|
50 | user_ns.update(ns) | |
|
51 | ||
|
52 | def _pull(user_ns, keys): | |
|
53 | """helper method for implementing `client.pull` via `client.apply`""" | |
|
54 | if isinstance(keys, (list,tuple, set)): | |
|
55 | for key in keys: | |
|
56 | if not user_ns.has_key(key): | |
|
57 | raise NameError("name '%s' is not defined"%key) | |
|
58 | return map(user_ns.get, keys) | |
|
59 | else: | |
|
60 | if not user_ns.has_key(keys): | |
|
61 | raise NameError("name '%s' is not defined"%keys) | |
|
62 | return user_ns.get(keys) | |
|
63 | ||
|
64 | def _clear(user_ns): | |
|
65 | """helper method for implementing `client.clear` via `client.apply`""" | |
|
66 | user_ns.clear() | |
|
67 | ||
|
68 | def _execute(user_ns, code): | |
|
69 | """helper method for implementing `client.execute` via `client.apply`""" | |
|
70 | exec code in user_ns | |
|
71 | ||
|
72 | ||
|
73 | #-------------------------------------------------------------------------- | |
|
74 | 43 | # Decorators for Client methods |
|
75 | 44 | #-------------------------------------------------------------------------- |
|
76 | 45 | |
|
77 | 46 | @decorator |
|
78 | def spinfirst(f, self, *args, **kwargs): | |
|
47 | def spin_first(f, self, *args, **kwargs): | |
|
79 | 48 | """Call spin() to sync state prior to calling the method.""" |
|
80 | 49 | self.spin() |
|
81 | 50 | return f(self, *args, **kwargs) |
|
82 | 51 | |
|
83 | 52 | @decorator |
|
84 | def defaultblock(f, self, *args, **kwargs): | |
|
53 | def default_block(f, self, *args, **kwargs): | |
|
85 | 54 | """Default to self.block; preserve self.block.""" |
|
86 | 55 | block = kwargs.get('block',None) |
|
87 | 56 | block = self.block if block is None else block |
|
88 | 57 | saveblock = self.block |
|
89 | 58 | self.block = block |
|
90 | 59 | try: |
|
91 | 60 | ret = f(self, *args, **kwargs) |
|
92 | 61 | finally: |
|
93 | 62 | self.block = saveblock |
|
94 | 63 | return ret |
|
95 | 64 | |
|
96 | 65 | |
|
97 | 66 | #-------------------------------------------------------------------------- |
|
98 | 67 | # Classes |
|
99 | 68 | #-------------------------------------------------------------------------- |
|
100 | 69 | |
|
101 | 70 | class Metadata(dict): |
|
102 | 71 | """Subclass of dict for initializing metadata values. |
|
103 | 72 | |
|
104 | 73 | Attribute access works on keys. |
|
105 | 74 | |
|
106 | 75 | These objects have a strict set of keys - errors will raise if you try |
|
107 | 76 | to add new keys. |
|
108 | 77 | """ |
|
109 | 78 | def __init__(self, *args, **kwargs): |
|
110 | 79 | dict.__init__(self) |
|
111 | 80 | md = {'msg_id' : None, |
|
112 | 81 | 'submitted' : None, |
|
113 | 82 | 'started' : None, |
|
114 | 83 | 'completed' : None, |
|
115 | 84 | 'received' : None, |
|
116 | 85 | 'engine_uuid' : None, |
|
117 | 86 | 'engine_id' : None, |
|
118 | 87 | 'follow' : None, |
|
119 | 88 | 'after' : None, |
|
120 | 89 | 'status' : None, |
|
121 | 90 | |
|
122 | 91 | 'pyin' : None, |
|
123 | 92 | 'pyout' : None, |
|
124 | 93 | 'pyerr' : None, |
|
125 | 94 | 'stdout' : '', |
|
126 | 95 | 'stderr' : '', |
|
127 | 96 | } |
|
128 | 97 | self.update(md) |
|
129 | 98 | self.update(dict(*args, **kwargs)) |
|
130 | 99 | |
|
131 | 100 | def __getattr__(self, key): |
|
132 | 101 | """getattr aliased to getitem""" |
|
133 | 102 | if key in self.iterkeys(): |
|
134 | 103 | return self[key] |
|
135 | 104 | else: |
|
136 | 105 | raise AttributeError(key) |
|
137 | 106 | |
|
138 | 107 | def __setattr__(self, key, value): |
|
139 | 108 | """setattr aliased to setitem, with strict""" |
|
140 | 109 | if key in self.iterkeys(): |
|
141 | 110 | self[key] = value |
|
142 | 111 | else: |
|
143 | 112 | raise AttributeError(key) |
|
144 | 113 | |
|
145 | 114 | def __setitem__(self, key, value): |
|
146 | 115 | """strict static key enforcement""" |
|
147 | 116 | if key in self.iterkeys(): |
|
148 | 117 | dict.__setitem__(self, key, value) |
|
149 | 118 | else: |
|
150 | 119 | raise KeyError(key) |
|
151 | 120 | |
|
152 | 121 | |
|
153 | 122 | class Client(HasTraits): |
|
154 |
"""A semi-synchronous client to the IPython ZMQ c |
|
|
123 | """A semi-synchronous client to the IPython ZMQ cluster | |
|
155 | 124 | |
|
156 | 125 | Parameters |
|
157 | 126 | ---------- |
|
158 | 127 | |
|
159 | 128 | url_or_file : bytes; zmq url or path to ipcontroller-client.json |
|
160 | 129 | Connection information for the Hub's registration. If a json connector |
|
161 | 130 | file is given, then likely no further configuration is necessary. |
|
162 | 131 | [Default: use profile] |
|
163 | 132 | profile : bytes |
|
164 | 133 | The name of the Cluster profile to be used to find connector information. |
|
165 | 134 | [Default: 'default'] |
|
166 | 135 | context : zmq.Context |
|
167 | 136 | Pass an existing zmq.Context instance, otherwise the client will create its own. |
|
168 | 137 | username : bytes |
|
169 | 138 | set username to be passed to the Session object |
|
170 | 139 | debug : bool |
|
171 | 140 | flag for lots of message printing for debug purposes |
|
172 | 141 | |
|
173 | 142 | #-------------- ssh related args ---------------- |
|
174 | 143 | # These are args for configuring the ssh tunnel to be used |
|
175 | 144 | # credentials are used to forward connections over ssh to the Controller |
|
176 | 145 | # Note that the ip given in `addr` needs to be relative to sshserver |
|
177 | 146 | # The most basic case is to leave addr as pointing to localhost (127.0.0.1), |
|
178 | 147 | # and set sshserver as the same machine the Controller is on. However, |
|
179 | 148 | # the only requirement is that sshserver is able to see the Controller |
|
180 | 149 | # (i.e. is within the same trusted network). |
|
181 | 150 | |
|
182 | 151 | sshserver : str |
|
183 | 152 | A string of the form passed to ssh, i.e. 'server.tld' or 'user@server.tld:port' |
|
184 | 153 | If keyfile or password is specified, and this is not, it will default to |
|
185 | 154 | the ip given in addr. |
|
186 | 155 | sshkey : str; path to public ssh key file |
|
187 | 156 | This specifies a key to be used in ssh login, default None. |
|
188 | 157 | Regular default ssh keys will be used without specifying this argument. |
|
189 | 158 | password : str |
|
190 | 159 | Your ssh password to sshserver. Note that if this is left None, |
|
191 | 160 | you will be prompted for it if passwordless key based login is unavailable. |
|
192 | 161 | paramiko : bool |
|
193 | 162 | flag for whether to use paramiko instead of shell ssh for tunneling. |
|
194 | 163 | [default: True on win32, False else] |
|
195 | 164 | |
|
196 |
|
|
|
197 |
|
|
|
198 |
|
|
|
199 |
|
|
|
200 |
|
|
|
165 | ------- exec authentication args ------- | |
|
166 | If even localhost is untrusted, you can have some protection against | |
|
167 | unauthorized execution by using a key. Messages are still sent | |
|
168 | as cleartext, so if someone can snoop your loopback traffic this will | |
|
169 | not help against malicious attacks. | |
|
201 | 170 | |
|
202 | 171 | exec_key : str |
|
203 | 172 | an authentication key or file containing a key |
|
204 | 173 | default: None |
|
205 | 174 | |
|
206 | 175 | |
|
207 | 176 | Attributes |
|
208 | 177 | ---------- |
|
209 | 178 | |
|
210 |
ids : s |
|
|
179 | ids : list of int engine IDs | |
|
211 | 180 | requesting the ids attribute always synchronizes |
|
212 | 181 | the registration state. To request ids without synchronization, |
|
213 | 182 | use semi-private _ids attributes. |
|
214 | 183 | |
|
215 | 184 | history : list of msg_ids |
|
216 | 185 | a list of msg_ids, keeping track of all the execution |
|
217 | 186 | messages you have submitted in order. |
|
218 | 187 | |
|
219 | 188 | outstanding : set of msg_ids |
|
220 | 189 | a set of msg_ids that have been submitted, but whose |
|
221 | 190 | results have not yet been received. |
|
222 | 191 | |
|
223 | 192 | results : dict |
|
224 | 193 | a dict of all our results, keyed by msg_id |
|
225 | 194 | |
|
226 | 195 | block : bool |
|
227 | 196 | determines default behavior when block not specified |
|
228 | 197 | in execution methods |
|
229 | 198 | |
|
230 | 199 | Methods |
|
231 | 200 | ------- |
|
232 | 201 | |
|
233 | 202 | spin |
|
234 | 203 | flushes incoming results and registration state changes |
|
235 | 204 | control methods spin, and requesting `ids` also ensures up to date |
|
236 | 205 | |
|
237 |
|
|
|
206 | wait | |
|
238 | 207 | wait on one or more msg_ids |
|
239 | 208 | |
|
240 | 209 | execution methods |
|
241 | 210 | apply |
|
242 | 211 | legacy: execute, run |
|
243 | 212 | |
|
213 | data movement | |
|
214 | push, pull, scatter, gather | |
|
215 | ||
|
244 | 216 | query methods |
|
245 | queue_status, get_result, purge | |
|
217 | queue_status, get_result, purge, result_status | |
|
246 | 218 | |
|
247 | 219 | control methods |
|
248 | 220 | abort, shutdown |
|
249 | 221 | |
|
250 | 222 | """ |
|
251 | 223 | |
|
252 | 224 | |
|
253 | 225 | block = Bool(False) |
|
254 | 226 | outstanding = Set() |
|
255 | 227 | results = Instance('collections.defaultdict', (dict,)) |
|
256 | 228 | metadata = Instance('collections.defaultdict', (Metadata,)) |
|
257 | 229 | history = List() |
|
258 | 230 | debug = Bool(False) |
|
259 | 231 | profile=CUnicode('default') |
|
260 | 232 | |
|
261 | 233 | _outstanding_dict = Instance('collections.defaultdict', (set,)) |
|
262 | 234 | _ids = List() |
|
263 | 235 | _connected=Bool(False) |
|
264 | 236 | _ssh=Bool(False) |
|
265 | 237 | _context = Instance('zmq.Context') |
|
266 | 238 | _config = Dict() |
|
267 | _engines=Instance(ReverseDict, (), {}) | |
|
239 | _engines=Instance(util.ReverseDict, (), {}) | |
|
268 | 240 | # _hub_socket=Instance('zmq.Socket') |
|
269 | 241 | _query_socket=Instance('zmq.Socket') |
|
270 | 242 | _control_socket=Instance('zmq.Socket') |
|
271 | 243 | _iopub_socket=Instance('zmq.Socket') |
|
272 | 244 | _notification_socket=Instance('zmq.Socket') |
|
273 |
_ |
|
|
274 | _mux_ident=Str() | |
|
275 | _task_ident=Str() | |
|
245 | _mux_socket=Instance('zmq.Socket') | |
|
246 | _task_socket=Instance('zmq.Socket') | |
|
276 | 247 | _task_scheme=Str() |
|
277 | 248 | _balanced_views=Dict() |
|
278 | 249 | _direct_views=Dict() |
|
279 | 250 | _closed = False |
|
251 | _ignored_control_replies=Int(0) | |
|
252 | _ignored_hub_replies=Int(0) | |
|
280 | 253 | |
|
281 | 254 | def __init__(self, url_or_file=None, profile='default', cluster_dir=None, ipython_dir=None, |
|
282 | 255 | context=None, username=None, debug=False, exec_key=None, |
|
283 | 256 | sshserver=None, sshkey=None, password=None, paramiko=None, |
|
257 | timeout=10 | |
|
284 | 258 | ): |
|
285 | 259 | super(Client, self).__init__(debug=debug, profile=profile) |
|
286 | 260 | if context is None: |
|
287 | 261 | context = zmq.Context.instance() |
|
288 | 262 | self._context = context |
|
289 | 263 | |
|
290 | 264 | |
|
291 | 265 | self._setup_cluster_dir(profile, cluster_dir, ipython_dir) |
|
292 | 266 | if self._cd is not None: |
|
293 | 267 | if url_or_file is None: |
|
294 | 268 | url_or_file = pjoin(self._cd.security_dir, 'ipcontroller-client.json') |
|
295 |
assert url_or_file is not None, "I can't find enough information to connect to a |
|
|
269 | assert url_or_file is not None, "I can't find enough information to connect to a hub!"\ | |
|
296 | 270 | " Please specify at least one of url_or_file or profile." |
|
297 | 271 | |
|
298 | 272 | try: |
|
299 | validate_url(url_or_file) | |
|
273 | util.validate_url(url_or_file) | |
|
300 | 274 | except AssertionError: |
|
301 | 275 | if not os.path.exists(url_or_file): |
|
302 | 276 | if self._cd: |
|
303 | 277 | url_or_file = os.path.join(self._cd.security_dir, url_or_file) |
|
304 | 278 | assert os.path.exists(url_or_file), "Not a valid connection file or url: %r"%url_or_file |
|
305 | 279 | with open(url_or_file) as f: |
|
306 | 280 | cfg = json.loads(f.read()) |
|
307 | 281 | else: |
|
308 | 282 | cfg = {'url':url_or_file} |
|
309 | 283 | |
|
310 | 284 | # sync defaults from args, json: |
|
311 | 285 | if sshserver: |
|
312 | 286 | cfg['ssh'] = sshserver |
|
313 | 287 | if exec_key: |
|
314 | 288 | cfg['exec_key'] = exec_key |
|
315 | 289 | exec_key = cfg['exec_key'] |
|
316 | 290 | sshserver=cfg['ssh'] |
|
317 | 291 | url = cfg['url'] |
|
318 | 292 | location = cfg.setdefault('location', None) |
|
319 | cfg['url'] = disambiguate_url(cfg['url'], location) | |
|
293 | cfg['url'] = util.disambiguate_url(cfg['url'], location) | |
|
320 | 294 | url = cfg['url'] |
|
321 | 295 | |
|
322 | 296 | self._config = cfg |
|
323 | 297 | |
|
324 | 298 | self._ssh = bool(sshserver or sshkey or password) |
|
325 | 299 | if self._ssh and sshserver is None: |
|
326 | 300 | # default to ssh via localhost |
|
327 | 301 | sshserver = url.split('://')[1].split(':')[0] |
|
328 | 302 | if self._ssh and password is None: |
|
329 | 303 | if tunnel.try_passwordless_ssh(sshserver, sshkey, paramiko): |
|
330 | 304 | password=False |
|
331 | 305 | else: |
|
332 | 306 | password = getpass("SSH Password for %s: "%sshserver) |
|
333 | 307 | ssh_kwargs = dict(keyfile=sshkey, password=password, paramiko=paramiko) |
|
334 | 308 | if exec_key is not None and os.path.isfile(exec_key): |
|
335 | 309 | arg = 'keyfile' |
|
336 | 310 | else: |
|
337 | 311 | arg = 'key' |
|
338 | 312 | key_arg = {arg:exec_key} |
|
339 | 313 | if username is None: |
|
340 | 314 | self.session = ss.StreamSession(**key_arg) |
|
341 | 315 | else: |
|
342 | 316 | self.session = ss.StreamSession(username, **key_arg) |
|
343 | 317 | self._query_socket = self._context.socket(zmq.XREQ) |
|
344 | 318 | self._query_socket.setsockopt(zmq.IDENTITY, self.session.session) |
|
345 | 319 | if self._ssh: |
|
346 | 320 | tunnel.tunnel_connection(self._query_socket, url, sshserver, **ssh_kwargs) |
|
347 | 321 | else: |
|
348 | 322 | self._query_socket.connect(url) |
|
349 | 323 | |
|
350 | 324 | self.session.debug = self.debug |
|
351 | 325 | |
|
352 | 326 | self._notification_handlers = {'registration_notification' : self._register_engine, |
|
353 | 327 | 'unregistration_notification' : self._unregister_engine, |
|
328 | 'shutdown_notification' : lambda msg: self.close(), | |
|
354 | 329 | } |
|
355 | 330 | self._queue_handlers = {'execute_reply' : self._handle_execute_reply, |
|
356 | 331 | 'apply_reply' : self._handle_apply_reply} |
|
357 | self._connect(sshserver, ssh_kwargs) | |
|
332 | self._connect(sshserver, ssh_kwargs, timeout) | |
|
358 | 333 | |
|
359 | 334 | def __del__(self): |
|
360 | 335 | """cleanup sockets, but _not_ context.""" |
|
361 | 336 | self.close() |
|
362 | 337 | |
|
363 | 338 | def _setup_cluster_dir(self, profile, cluster_dir, ipython_dir): |
|
364 | 339 | if ipython_dir is None: |
|
365 | 340 | ipython_dir = get_ipython_dir() |
|
366 | 341 | if cluster_dir is not None: |
|
367 | 342 | try: |
|
368 | 343 | self._cd = ClusterDir.find_cluster_dir(cluster_dir) |
|
369 | 344 | return |
|
370 | 345 | except ClusterDirError: |
|
371 | 346 | pass |
|
372 | 347 | elif profile is not None: |
|
373 | 348 | try: |
|
374 | 349 | self._cd = ClusterDir.find_cluster_dir_by_profile( |
|
375 | 350 | ipython_dir, profile) |
|
376 | 351 | return |
|
377 | 352 | except ClusterDirError: |
|
378 | 353 | pass |
|
379 | 354 | self._cd = None |
|
380 | 355 | |
|
381 | @property | |
|
382 | def ids(self): | |
|
383 | """Always up-to-date ids property.""" | |
|
384 | self._flush_notifications() | |
|
385 | # always copy: | |
|
386 | return list(self._ids) | |
|
387 | ||
|
388 | def close(self): | |
|
389 | if self._closed: | |
|
390 | return | |
|
391 | snames = filter(lambda n: n.endswith('socket'), dir(self)) | |
|
392 | for socket in map(lambda name: getattr(self, name), snames): | |
|
393 | if isinstance(socket, zmq.Socket) and not socket.closed: | |
|
394 | socket.close() | |
|
395 | self._closed = True | |
|
396 | ||
|
397 | 356 | def _update_engines(self, engines): |
|
398 | 357 | """Update our engines dict and _ids from a dict of the form: {id:uuid}.""" |
|
399 | 358 | for k,v in engines.iteritems(): |
|
400 | 359 | eid = int(k) |
|
401 | 360 | self._engines[eid] = bytes(v) # force not unicode |
|
402 | 361 | self._ids.append(eid) |
|
403 | 362 | self._ids = sorted(self._ids) |
|
404 | 363 | if sorted(self._engines.keys()) != range(len(self._engines)) and \ |
|
405 |
self._task_scheme == 'pure' and self._task_ |
|
|
364 | self._task_scheme == 'pure' and self._task_socket: | |
|
406 | 365 | self._stop_scheduling_tasks() |
|
407 | 366 | |
|
408 | 367 | def _stop_scheduling_tasks(self): |
|
409 | 368 | """Stop scheduling tasks because an engine has been unregistered |
|
410 | 369 | from a pure ZMQ scheduler. |
|
411 | 370 | """ |
|
412 |
self._task_ |
|
|
413 |
|
|
|
414 | # self._task_socket = None | |
|
371 | self._task_socket.close() | |
|
372 | self._task_socket = None | |
|
415 | 373 | msg = "An engine has been unregistered, and we are using pure " +\ |
|
416 | 374 | "ZMQ task scheduling. Task farming will be disabled." |
|
417 | 375 | if self.outstanding: |
|
418 | 376 | msg += " If you were running tasks when this happened, " +\ |
|
419 | 377 | "some `outstanding` msg_ids may never resolve." |
|
420 | 378 | warnings.warn(msg, RuntimeWarning) |
|
421 | 379 | |
|
422 | 380 | def _build_targets(self, targets): |
|
423 | 381 | """Turn valid target IDs or 'all' into two lists: |
|
424 | 382 | (int_ids, uuids). |
|
425 | 383 | """ |
|
426 | 384 | if targets is None: |
|
427 | 385 | targets = self._ids |
|
428 | 386 | elif isinstance(targets, str): |
|
429 | 387 | if targets.lower() == 'all': |
|
430 | 388 | targets = self._ids |
|
431 | 389 | else: |
|
432 | 390 | raise TypeError("%r not valid str target, must be 'all'"%(targets)) |
|
433 | 391 | elif isinstance(targets, int): |
|
434 | 392 | targets = [targets] |
|
435 | 393 | return [self._engines[t] for t in targets], list(targets) |
|
436 | 394 | |
|
437 | def _connect(self, sshserver, ssh_kwargs): | |
|
438 |
"""setup all our socket connections to the c |
|
|
395 | def _connect(self, sshserver, ssh_kwargs, timeout): | |
|
396 | """setup all our socket connections to the cluster. This is called from | |
|
439 | 397 | __init__.""" |
|
440 | 398 | |
|
441 | 399 | # Maybe allow reconnecting? |
|
442 | 400 | if self._connected: |
|
443 | 401 | return |
|
444 | 402 | self._connected=True |
|
445 | 403 | |
|
446 | 404 | def connect_socket(s, url): |
|
447 | url = disambiguate_url(url, self._config['location']) | |
|
405 | url = util.disambiguate_url(url, self._config['location']) | |
|
448 | 406 | if self._ssh: |
|
449 | 407 | return tunnel.tunnel_connection(s, url, sshserver, **ssh_kwargs) |
|
450 | 408 | else: |
|
451 | 409 | return s.connect(url) |
|
452 | 410 | |
|
453 | 411 | self.session.send(self._query_socket, 'connection_request') |
|
412 | r,w,x = zmq.select([self._query_socket],[],[], timeout) | |
|
413 | if not r: | |
|
414 | raise error.TimeoutError("Hub connection request timed out") | |
|
454 | 415 | idents,msg = self.session.recv(self._query_socket,mode=0) |
|
455 | 416 | if self.debug: |
|
456 | 417 | pprint(msg) |
|
457 | 418 | msg = ss.Message(msg) |
|
458 | 419 | content = msg.content |
|
459 | 420 | self._config['registration'] = dict(content) |
|
460 | 421 | if content.status == 'ok': |
|
461 | self._apply_socket = self._context.socket(zmq.XREP) | |
|
462 | self._apply_socket.setsockopt(zmq.IDENTITY, self.session.session) | |
|
463 | 422 | if content.mux: |
|
464 |
|
|
|
465 | self._mux_ident = 'mux' | |
|
466 |
connect_socket(self._ |
|
|
423 | self._mux_socket = self._context.socket(zmq.XREQ) | |
|
424 | self._mux_socket.setsockopt(zmq.IDENTITY, self.session.session) | |
|
425 | connect_socket(self._mux_socket, content.mux) | |
|
467 | 426 | if content.task: |
|
468 | 427 | self._task_scheme, task_addr = content.task |
|
469 |
|
|
|
470 |
|
|
|
471 |
connect_socket(self._ |
|
|
472 | self._task_ident = 'task' | |
|
428 | self._task_socket = self._context.socket(zmq.XREQ) | |
|
429 | self._task_socket.setsockopt(zmq.IDENTITY, self.session.session) | |
|
430 | connect_socket(self._task_socket, task_addr) | |
|
473 | 431 | if content.notification: |
|
474 | 432 | self._notification_socket = self._context.socket(zmq.SUB) |
|
475 | 433 | connect_socket(self._notification_socket, content.notification) |
|
476 | 434 | self._notification_socket.setsockopt(zmq.SUBSCRIBE, b'') |
|
477 | 435 | # if content.query: |
|
478 | 436 | # self._query_socket = self._context.socket(zmq.XREQ) |
|
479 | 437 | # self._query_socket.setsockopt(zmq.IDENTITY, self.session.session) |
|
480 | 438 | # connect_socket(self._query_socket, content.query) |
|
481 | 439 | if content.control: |
|
482 | 440 | self._control_socket = self._context.socket(zmq.XREQ) |
|
483 | 441 | self._control_socket.setsockopt(zmq.IDENTITY, self.session.session) |
|
484 | 442 | connect_socket(self._control_socket, content.control) |
|
485 | 443 | if content.iopub: |
|
486 | 444 | self._iopub_socket = self._context.socket(zmq.SUB) |
|
487 | 445 | self._iopub_socket.setsockopt(zmq.SUBSCRIBE, b'') |
|
488 | 446 | self._iopub_socket.setsockopt(zmq.IDENTITY, self.session.session) |
|
489 | 447 | connect_socket(self._iopub_socket, content.iopub) |
|
490 | 448 | self._update_engines(dict(content.engines)) |
|
491 | # give XREP apply_socket some time to connect | |
|
492 | time.sleep(0.25) | |
|
493 | 449 | else: |
|
494 | 450 | self._connected = False |
|
495 | 451 | raise Exception("Failed to connect!") |
|
496 | 452 | |
|
497 | 453 | #-------------------------------------------------------------------------- |
|
498 | 454 | # handlers and callbacks for incoming messages |
|
499 | 455 | #-------------------------------------------------------------------------- |
|
500 | 456 | |
|
501 | 457 | def _unwrap_exception(self, content): |
|
502 | """unwrap exception, and remap engineid to int.""" | |
|
458 | """unwrap exception, and remap engine_id to int.""" | |
|
503 | 459 | e = error.unwrap_exception(content) |
|
504 | 460 | # print e.traceback |
|
505 | 461 | if e.engine_info: |
|
506 | 462 | e_uuid = e.engine_info['engine_uuid'] |
|
507 | 463 | eid = self._engines[e_uuid] |
|
508 | 464 | e.engine_info['engine_id'] = eid |
|
509 | 465 | return e |
|
510 | 466 | |
|
511 | 467 | def _extract_metadata(self, header, parent, content): |
|
512 | 468 | md = {'msg_id' : parent['msg_id'], |
|
513 | 469 | 'received' : datetime.now(), |
|
514 | 470 | 'engine_uuid' : header.get('engine', None), |
|
515 | 471 | 'follow' : parent.get('follow', []), |
|
516 | 472 | 'after' : parent.get('after', []), |
|
517 | 473 | 'status' : content['status'], |
|
518 | 474 | } |
|
519 | 475 | |
|
520 | 476 | if md['engine_uuid'] is not None: |
|
521 | 477 | md['engine_id'] = self._engines.get(md['engine_uuid'], None) |
|
522 | 478 | |
|
523 | 479 | if 'date' in parent: |
|
524 | 480 | md['submitted'] = datetime.strptime(parent['date'], util.ISO8601) |
|
525 | 481 | if 'started' in header: |
|
526 | 482 | md['started'] = datetime.strptime(header['started'], util.ISO8601) |
|
527 | 483 | if 'date' in header: |
|
528 | 484 | md['completed'] = datetime.strptime(header['date'], util.ISO8601) |
|
529 | 485 | return md |
|
530 | 486 | |
|
531 | 487 | def _register_engine(self, msg): |
|
532 | 488 | """Register a new engine, and update our connection info.""" |
|
533 | 489 | content = msg['content'] |
|
534 | 490 | eid = content['id'] |
|
535 | 491 | d = {eid : content['queue']} |
|
536 | 492 | self._update_engines(d) |
|
537 | 493 | |
|
538 | 494 | def _unregister_engine(self, msg): |
|
539 | 495 | """Unregister an engine that has died.""" |
|
540 | 496 | content = msg['content'] |
|
541 | 497 | eid = int(content['id']) |
|
542 | 498 | if eid in self._ids: |
|
543 | 499 | self._ids.remove(eid) |
|
544 | 500 | uuid = self._engines.pop(eid) |
|
545 | 501 | |
|
546 | 502 | self._handle_stranded_msgs(eid, uuid) |
|
547 | 503 | |
|
548 |
if self._task_ |
|
|
504 | if self._task_socket and self._task_scheme == 'pure': | |
|
549 | 505 | self._stop_scheduling_tasks() |
|
550 | 506 | |
|
551 | 507 | def _handle_stranded_msgs(self, eid, uuid): |
|
552 | 508 | """Handle messages known to be on an engine when the engine unregisters. |
|
553 | 509 | |
|
554 | 510 | It is possible that this will fire prematurely - that is, an engine will |
|
555 | 511 | go down after completing a result, and the client will be notified |
|
556 | 512 | of the unregistration and later receive the successful result. |
|
557 | 513 | """ |
|
558 | 514 | |
|
559 | 515 | outstanding = self._outstanding_dict[uuid] |
|
560 | 516 | |
|
561 | 517 | for msg_id in list(outstanding): |
|
562 | 518 | if msg_id in self.results: |
|
563 | 519 | # we already |
|
564 | 520 | continue |
|
565 | 521 | try: |
|
566 | 522 | raise error.EngineError("Engine %r died while running task %r"%(eid, msg_id)) |
|
567 | 523 | except: |
|
568 | 524 | content = error.wrap_exception() |
|
569 | 525 | # build a fake message: |
|
570 | 526 | parent = {} |
|
571 | 527 | header = {} |
|
572 | 528 | parent['msg_id'] = msg_id |
|
573 | 529 | header['engine'] = uuid |
|
574 | 530 | header['date'] = datetime.now().strftime(util.ISO8601) |
|
575 | 531 | msg = dict(parent_header=parent, header=header, content=content) |
|
576 | 532 | self._handle_apply_reply(msg) |
|
577 | 533 | |
|
578 | 534 | def _handle_execute_reply(self, msg): |
|
579 | 535 | """Save the reply to an execute_request into our results. |
|
580 | 536 | |
|
581 | 537 | execute messages are never actually used. apply is used instead. |
|
582 | 538 | """ |
|
583 | 539 | |
|
584 | 540 | parent = msg['parent_header'] |
|
585 | 541 | msg_id = parent['msg_id'] |
|
586 | 542 | if msg_id not in self.outstanding: |
|
587 | 543 | if msg_id in self.history: |
|
588 | 544 | print ("got stale result: %s"%msg_id) |
|
589 | 545 | else: |
|
590 | 546 | print ("got unknown result: %s"%msg_id) |
|
591 | 547 | else: |
|
592 | 548 | self.outstanding.remove(msg_id) |
|
593 | 549 | self.results[msg_id] = self._unwrap_exception(msg['content']) |
|
594 | 550 | |
|
595 | 551 | def _handle_apply_reply(self, msg): |
|
596 | 552 | """Save the reply to an apply_request into our results.""" |
|
597 | 553 | parent = msg['parent_header'] |
|
598 | 554 | msg_id = parent['msg_id'] |
|
599 | 555 | if msg_id not in self.outstanding: |
|
600 | 556 | if msg_id in self.history: |
|
601 | 557 | print ("got stale result: %s"%msg_id) |
|
602 | 558 | print self.results[msg_id] |
|
603 | 559 | print msg |
|
604 | 560 | else: |
|
605 | 561 | print ("got unknown result: %s"%msg_id) |
|
606 | 562 | else: |
|
607 | 563 | self.outstanding.remove(msg_id) |
|
608 | 564 | content = msg['content'] |
|
609 | 565 | header = msg['header'] |
|
610 | 566 | |
|
611 | 567 | # construct metadata: |
|
612 | 568 | md = self.metadata[msg_id] |
|
613 | 569 | md.update(self._extract_metadata(header, parent, content)) |
|
614 | 570 | # is this redundant? |
|
615 | 571 | self.metadata[msg_id] = md |
|
616 | 572 | |
|
617 | 573 | e_outstanding = self._outstanding_dict[md['engine_uuid']] |
|
618 | 574 | if msg_id in e_outstanding: |
|
619 | 575 | e_outstanding.remove(msg_id) |
|
620 | 576 | |
|
621 | 577 | # construct result: |
|
622 | 578 | if content['status'] == 'ok': |
|
623 | 579 | self.results[msg_id] = util.unserialize_object(msg['buffers'])[0] |
|
624 | 580 | elif content['status'] == 'aborted': |
|
625 |
self.results[msg_id] = error.Aborted |
|
|
581 | self.results[msg_id] = error.TaskAborted(msg_id) | |
|
626 | 582 | elif content['status'] == 'resubmitted': |
|
627 | 583 | # TODO: handle resubmission |
|
628 | 584 | pass |
|
629 | 585 | else: |
|
630 | 586 | self.results[msg_id] = self._unwrap_exception(content) |
|
631 | 587 | |
|
632 | 588 | def _flush_notifications(self): |
|
633 | 589 | """Flush notifications of engine registrations waiting |
|
634 | 590 | in ZMQ queue.""" |
|
635 | 591 | msg = self.session.recv(self._notification_socket, mode=zmq.NOBLOCK) |
|
636 | 592 | while msg is not None: |
|
637 | 593 | if self.debug: |
|
638 | 594 | pprint(msg) |
|
639 | 595 | msg = msg[-1] |
|
640 | 596 | msg_type = msg['msg_type'] |
|
641 | 597 | handler = self._notification_handlers.get(msg_type, None) |
|
642 | 598 | if handler is None: |
|
643 | 599 | raise Exception("Unhandled message type: %s"%msg.msg_type) |
|
644 | 600 | else: |
|
645 | 601 | handler(msg) |
|
646 | 602 | msg = self.session.recv(self._notification_socket, mode=zmq.NOBLOCK) |
|
647 | 603 | |
|
648 | 604 | def _flush_results(self, sock): |
|
649 | 605 | """Flush task or queue results waiting in ZMQ queue.""" |
|
650 | 606 | msg = self.session.recv(sock, mode=zmq.NOBLOCK) |
|
651 | 607 | while msg is not None: |
|
652 | 608 | if self.debug: |
|
653 | 609 | pprint(msg) |
|
654 | 610 | msg = msg[-1] |
|
655 | 611 | msg_type = msg['msg_type'] |
|
656 | 612 | handler = self._queue_handlers.get(msg_type, None) |
|
657 | 613 | if handler is None: |
|
658 | 614 | raise Exception("Unhandled message type: %s"%msg.msg_type) |
|
659 | 615 | else: |
|
660 | 616 | handler(msg) |
|
661 | 617 | msg = self.session.recv(sock, mode=zmq.NOBLOCK) |
|
662 | 618 | |
|
663 | 619 | def _flush_control(self, sock): |
|
664 | 620 | """Flush replies from the control channel waiting |
|
665 | 621 | in the ZMQ queue. |
|
666 | 622 | |
|
667 | 623 | Currently: ignore them.""" |
|
624 | if self._ignored_control_replies <= 0: | |
|
625 | return | |
|
668 | 626 | msg = self.session.recv(sock, mode=zmq.NOBLOCK) |
|
669 | 627 | while msg is not None: |
|
628 | self._ignored_control_replies -= 1 | |
|
670 | 629 | if self.debug: |
|
671 | 630 | pprint(msg) |
|
672 | 631 | msg = self.session.recv(sock, mode=zmq.NOBLOCK) |
|
673 | 632 | |
|
633 | def _flush_ignored_control(self): | |
|
634 | """flush ignored control replies""" | |
|
635 | while self._ignored_control_replies > 0: | |
|
636 | self.session.recv(self._control_socket) | |
|
637 | self._ignored_control_replies -= 1 | |
|
638 | ||
|
639 | def _flush_ignored_hub_replies(self): | |
|
640 | msg = self.session.recv(self._query_socket, mode=zmq.NOBLOCK) | |
|
641 | while msg is not None: | |
|
642 | msg = self.session.recv(self._query_socket, mode=zmq.NOBLOCK) | |
|
643 | ||
|
674 | 644 | def _flush_iopub(self, sock): |
|
675 | 645 | """Flush replies from the iopub channel waiting |
|
676 | 646 | in the ZMQ queue. |
|
677 | 647 | """ |
|
678 | 648 | msg = self.session.recv(sock, mode=zmq.NOBLOCK) |
|
679 | 649 | while msg is not None: |
|
680 | 650 | if self.debug: |
|
681 | 651 | pprint(msg) |
|
682 | 652 | msg = msg[-1] |
|
683 | 653 | parent = msg['parent_header'] |
|
684 | 654 | msg_id = parent['msg_id'] |
|
685 | 655 | content = msg['content'] |
|
686 | 656 | header = msg['header'] |
|
687 | 657 | msg_type = msg['msg_type'] |
|
688 | 658 | |
|
689 | 659 | # init metadata: |
|
690 | 660 | md = self.metadata[msg_id] |
|
691 | 661 | |
|
692 | 662 | if msg_type == 'stream': |
|
693 | 663 | name = content['name'] |
|
694 | 664 | s = md[name] or '' |
|
695 | 665 | md[name] = s + content['data'] |
|
696 | 666 | elif msg_type == 'pyerr': |
|
697 | 667 | md.update({'pyerr' : self._unwrap_exception(content)}) |
|
698 | 668 | else: |
|
699 | 669 | md.update({msg_type : content['data']}) |
|
700 | 670 | |
|
701 | 671 | # reduntant? |
|
702 | 672 | self.metadata[msg_id] = md |
|
703 | 673 | |
|
704 | 674 | msg = self.session.recv(sock, mode=zmq.NOBLOCK) |
|
705 | 675 | |
|
706 | 676 | #-------------------------------------------------------------------------- |
|
707 | 677 | # len, getitem |
|
708 | 678 | #-------------------------------------------------------------------------- |
|
709 | 679 | |
|
710 | 680 | def __len__(self): |
|
711 | 681 | """len(client) returns # of engines.""" |
|
712 | 682 | return len(self.ids) |
|
713 | 683 | |
|
714 | 684 | def __getitem__(self, key): |
|
715 | 685 | """index access returns DirectView multiplexer objects |
|
716 | 686 | |
|
717 | 687 | Must be int, slice, or list/tuple/xrange of ints""" |
|
718 | 688 | if not isinstance(key, (int, slice, tuple, list, xrange)): |
|
719 | 689 | raise TypeError("key by int/slice/iterable of ints only, not %s"%(type(key))) |
|
720 | 690 | else: |
|
721 | return self.view(key, balanced=False) | |
|
691 | return self._get_view(key, balanced=False) | |
|
722 | 692 | |
|
723 | 693 | #-------------------------------------------------------------------------- |
|
724 | 694 | # Begin public methods |
|
725 | 695 | #-------------------------------------------------------------------------- |
|
726 | 696 | |
|
697 | @property | |
|
698 | def ids(self): | |
|
699 | """Always up-to-date ids property.""" | |
|
700 | self._flush_notifications() | |
|
701 | # always copy: | |
|
702 | return list(self._ids) | |
|
703 | ||
|
704 | def close(self): | |
|
705 | if self._closed: | |
|
706 | return | |
|
707 | snames = filter(lambda n: n.endswith('socket'), dir(self)) | |
|
708 | for socket in map(lambda name: getattr(self, name), snames): | |
|
709 | if isinstance(socket, zmq.Socket) and not socket.closed: | |
|
710 | socket.close() | |
|
711 | self._closed = True | |
|
712 | ||
|
727 | 713 | def spin(self): |
|
728 | 714 | """Flush any registration notifications and execution results |
|
729 | 715 | waiting in the ZMQ queue. |
|
730 | 716 | """ |
|
731 | 717 | if self._notification_socket: |
|
732 | 718 | self._flush_notifications() |
|
733 |
if self._ |
|
|
734 |
self._flush_results(self._ |
|
|
719 | if self._mux_socket: | |
|
720 | self._flush_results(self._mux_socket) | |
|
721 | if self._task_socket: | |
|
722 | self._flush_results(self._task_socket) | |
|
735 | 723 | if self._control_socket: |
|
736 | 724 | self._flush_control(self._control_socket) |
|
737 | 725 | if self._iopub_socket: |
|
738 | 726 | self._flush_iopub(self._iopub_socket) |
|
727 | if self._query_socket: | |
|
728 | self._flush_ignored_hub_replies() | |
|
739 | 729 | |
|
740 |
def |
|
|
730 | def wait(self, jobs=None, timeout=-1): | |
|
741 | 731 | """waits on one or more `jobs`, for up to `timeout` seconds. |
|
742 | 732 | |
|
743 | 733 | Parameters |
|
744 | 734 | ---------- |
|
745 | 735 | |
|
746 | 736 | jobs : int, str, or list of ints and/or strs, or one or more AsyncResult objects |
|
747 | 737 | ints are indices to self.history |
|
748 | 738 | strs are msg_ids |
|
749 | 739 | default: wait on all outstanding messages |
|
750 | 740 | timeout : float |
|
751 | 741 | a time in seconds, after which to give up. |
|
752 | 742 | default is -1, which means no timeout |
|
753 | 743 | |
|
754 | 744 | Returns |
|
755 | 745 | ------- |
|
756 | 746 | |
|
757 | 747 | True : when all msg_ids are done |
|
758 | 748 | False : timeout reached, some msg_ids still outstanding |
|
759 | 749 | """ |
|
760 | 750 | tic = time.time() |
|
761 | 751 | if jobs is None: |
|
762 | 752 | theids = self.outstanding |
|
763 | 753 | else: |
|
764 | 754 | if isinstance(jobs, (int, str, AsyncResult)): |
|
765 | 755 | jobs = [jobs] |
|
766 | 756 | theids = set() |
|
767 | 757 | for job in jobs: |
|
768 | 758 | if isinstance(job, int): |
|
769 | 759 | # index access |
|
770 | 760 | job = self.history[job] |
|
771 | 761 | elif isinstance(job, AsyncResult): |
|
772 | 762 | map(theids.add, job.msg_ids) |
|
773 | 763 | continue |
|
774 | 764 | theids.add(job) |
|
775 | 765 | if not theids.intersection(self.outstanding): |
|
776 | 766 | return True |
|
777 | 767 | self.spin() |
|
778 | 768 | while theids.intersection(self.outstanding): |
|
779 | 769 | if timeout >= 0 and ( time.time()-tic ) > timeout: |
|
780 | 770 | break |
|
781 | 771 | time.sleep(1e-3) |
|
782 | 772 | self.spin() |
|
783 | 773 | return len(theids.intersection(self.outstanding)) == 0 |
|
784 | 774 | |
|
785 | 775 | #-------------------------------------------------------------------------- |
|
786 | 776 | # Control methods |
|
787 | 777 | #-------------------------------------------------------------------------- |
|
788 | 778 | |
|
789 | @spinfirst | |
|
790 | @defaultblock | |
|
779 | @spin_first | |
|
780 | @default_block | |
|
791 | 781 | def clear(self, targets=None, block=None): |
|
792 | 782 | """Clear the namespace in target(s).""" |
|
793 | 783 | targets = self._build_targets(targets)[0] |
|
794 | 784 | for t in targets: |
|
795 | 785 | self.session.send(self._control_socket, 'clear_request', content={}, ident=t) |
|
796 | 786 | error = False |
|
797 | 787 | if self.block: |
|
788 | self._flush_ignored_control() | |
|
798 | 789 | for i in range(len(targets)): |
|
799 | 790 | idents,msg = self.session.recv(self._control_socket,0) |
|
800 | 791 | if self.debug: |
|
801 | 792 | pprint(msg) |
|
802 | 793 | if msg['content']['status'] != 'ok': |
|
803 | 794 | error = self._unwrap_exception(msg['content']) |
|
795 | else: | |
|
796 | self._ignored_control_replies += len(targets) | |
|
804 | 797 | if error: |
|
805 | 798 | raise error |
|
806 | 799 | |
|
807 | 800 | |
|
808 | @spinfirst | |
|
809 | @defaultblock | |
|
801 | @spin_first | |
|
802 | @default_block | |
|
810 | 803 | def abort(self, jobs=None, targets=None, block=None): |
|
811 | 804 | """Abort specific jobs from the execution queues of target(s). |
|
812 | 805 | |
|
813 | 806 | This is a mechanism to prevent jobs that have already been submitted |
|
814 | 807 | from executing. |
|
815 | 808 | |
|
816 | 809 | Parameters |
|
817 | 810 | ---------- |
|
818 | 811 | |
|
819 | 812 | jobs : msg_id, list of msg_ids, or AsyncResult |
|
820 | 813 | The jobs to be aborted |
|
821 | 814 | |
|
822 | 815 | |
|
823 | 816 | """ |
|
824 | 817 | targets = self._build_targets(targets)[0] |
|
825 | 818 | msg_ids = [] |
|
826 | 819 | if isinstance(jobs, (basestring,AsyncResult)): |
|
827 | 820 | jobs = [jobs] |
|
828 | 821 | bad_ids = filter(lambda obj: not isinstance(obj, (basestring, AsyncResult)), jobs) |
|
829 | 822 | if bad_ids: |
|
830 | 823 | raise TypeError("Invalid msg_id type %r, expected str or AsyncResult"%bad_ids[0]) |
|
831 | 824 | for j in jobs: |
|
832 | 825 | if isinstance(j, AsyncResult): |
|
833 | 826 | msg_ids.extend(j.msg_ids) |
|
834 | 827 | else: |
|
835 | 828 | msg_ids.append(j) |
|
836 | 829 | content = dict(msg_ids=msg_ids) |
|
837 | 830 | for t in targets: |
|
838 | 831 | self.session.send(self._control_socket, 'abort_request', |
|
839 | 832 | content=content, ident=t) |
|
840 | 833 | error = False |
|
841 | 834 | if self.block: |
|
835 | self._flush_ignored_control() | |
|
842 | 836 | for i in range(len(targets)): |
|
843 | 837 | idents,msg = self.session.recv(self._control_socket,0) |
|
844 | 838 | if self.debug: |
|
845 | 839 | pprint(msg) |
|
846 | 840 | if msg['content']['status'] != 'ok': |
|
847 | 841 | error = self._unwrap_exception(msg['content']) |
|
842 | else: | |
|
843 | self._ignored_control_replies += len(targets) | |
|
848 | 844 | if error: |
|
849 | 845 | raise error |
|
850 | 846 | |
|
851 | @spinfirst | |
|
852 | @defaultblock | |
|
853 |
def shutdown(self, targets=None, restart=False, |
|
|
854 |
"""Terminates one or more engine processes, optionally including the |
|
|
855 |
if |
|
|
847 | @spin_first | |
|
848 | @default_block | |
|
849 | def shutdown(self, targets=None, restart=False, hub=False, block=None): | |
|
850 | """Terminates one or more engine processes, optionally including the hub.""" | |
|
851 | if hub: | |
|
856 | 852 | targets = 'all' |
|
857 | 853 | targets = self._build_targets(targets)[0] |
|
858 | 854 | for t in targets: |
|
859 | 855 | self.session.send(self._control_socket, 'shutdown_request', |
|
860 | 856 | content={'restart':restart},ident=t) |
|
861 | 857 | error = False |
|
862 |
if block or |
|
|
858 | if block or hub: | |
|
859 | self._flush_ignored_control() | |
|
863 | 860 | for i in range(len(targets)): |
|
864 | 861 | idents,msg = self.session.recv(self._control_socket,0) |
|
865 | 862 | if self.debug: |
|
866 | 863 | pprint(msg) |
|
867 | 864 | if msg['content']['status'] != 'ok': |
|
868 | 865 | error = self._unwrap_exception(msg['content']) |
|
866 | else: | |
|
867 | self._ignored_control_replies += len(targets) | |
|
869 | 868 | |
|
870 |
if |
|
|
869 | if hub: | |
|
871 | 870 | time.sleep(0.25) |
|
872 | 871 | self.session.send(self._query_socket, 'shutdown_request') |
|
873 | 872 | idents,msg = self.session.recv(self._query_socket, 0) |
|
874 | 873 | if self.debug: |
|
875 | 874 | pprint(msg) |
|
876 | 875 | if msg['content']['status'] != 'ok': |
|
877 | 876 | error = self._unwrap_exception(msg['content']) |
|
878 | 877 | |
|
879 | 878 | if error: |
|
880 | 879 | raise error |
|
881 | 880 | |
|
882 | 881 | #-------------------------------------------------------------------------- |
|
883 | 882 | # Execution methods |
|
884 | 883 | #-------------------------------------------------------------------------- |
|
885 | 884 | |
|
886 | @defaultblock | |
|
887 | def execute(self, code, targets='all', block=None): | |
|
885 | @default_block | |
|
886 | def _execute(self, code, targets='all', block=None): | |
|
888 | 887 | """Executes `code` on `targets` in blocking or nonblocking manner. |
|
889 | 888 | |
|
890 | 889 | ``execute`` is always `bound` (affects engine namespace) |
|
891 | 890 | |
|
892 | 891 | Parameters |
|
893 | 892 | ---------- |
|
894 | 893 | |
|
895 | 894 | code : str |
|
896 | 895 | the code string to be executed |
|
897 | 896 | targets : int/str/list of ints/strs |
|
898 | 897 | the engines on which to execute |
|
899 | 898 | default : all |
|
900 | 899 | block : bool |
|
901 | 900 | whether or not to wait until done to return |
|
902 | 901 | default: self.block |
|
903 | 902 | """ |
|
904 | result = self.apply(_execute, (code,), targets=targets, block=block, bound=True, balanced=False) | |
|
905 | if not block: | |
|
906 | return result | |
|
907 | ||
|
908 | def run(self, filename, targets='all', block=None): | |
|
909 | """Execute contents of `filename` on engine(s). | |
|
910 | ||
|
911 | This simply reads the contents of the file and calls `execute`. | |
|
912 | ||
|
913 | Parameters | |
|
914 | ---------- | |
|
915 | ||
|
916 | filename : str | |
|
917 | The path to the file | |
|
918 | targets : int/str/list of ints/strs | |
|
919 | the engines on which to execute | |
|
920 | default : all | |
|
921 | block : bool | |
|
922 | whether or not to wait until done | |
|
923 | default: self.block | |
|
924 | ||
|
925 | """ | |
|
926 | with open(filename, 'r') as f: | |
|
927 | # add newline in case of trailing indented whitespace | |
|
928 | # which will cause SyntaxError | |
|
929 | code = f.read()+'\n' | |
|
930 | return self.execute(code, targets=targets, block=block) | |
|
903 | return self[targets].execute(code, block=block) | |
|
931 | 904 | |
|
932 | 905 | def _maybe_raise(self, result): |
|
933 | 906 | """wrapper for maybe raising an exception if apply failed.""" |
|
934 | 907 | if isinstance(result, error.RemoteError): |
|
935 | 908 | raise result |
|
936 | 909 | |
|
937 | 910 | return result |
|
938 | 911 | |
|
939 | def _build_dependency(self, dep): | |
|
940 | """helper for building jsonable dependencies from various input forms""" | |
|
941 | if isinstance(dep, Dependency): | |
|
942 | return dep.as_dict() | |
|
943 | elif isinstance(dep, AsyncResult): | |
|
944 | return dep.msg_ids | |
|
945 | elif dep is None: | |
|
946 | return [] | |
|
947 | else: | |
|
948 | # pass to Dependency constructor | |
|
949 | return list(Dependency(dep)) | |
|
950 | ||
|
951 | @defaultblock | |
|
952 | def apply(self, f, args=None, kwargs=None, bound=False, block=None, | |
|
953 | targets=None, balanced=None, | |
|
954 | after=None, follow=None, timeout=None, | |
|
955 | track=False): | |
|
956 | """Call `f(*args, **kwargs)` on a remote engine(s), returning the result. | |
|
957 | ||
|
958 | This is the central execution command for the client. | |
|
959 | ||
|
960 | Parameters | |
|
961 | ---------- | |
|
962 | ||
|
963 | f : function | |
|
964 | The fuction to be called remotely | |
|
965 | args : tuple/list | |
|
966 | The positional arguments passed to `f` | |
|
967 | kwargs : dict | |
|
968 | The keyword arguments passed to `f` | |
|
969 | bound : bool (default: False) | |
|
970 | Whether to pass the Engine(s) Namespace as the first argument to `f`. | |
|
971 | block : bool (default: self.block) | |
|
972 | Whether to wait for the result, or return immediately. | |
|
973 | False: | |
|
974 | returns AsyncResult | |
|
975 | True: | |
|
976 | returns actual result(s) of f(*args, **kwargs) | |
|
977 | if multiple targets: | |
|
978 | list of results, matching `targets` | |
|
979 | track : bool | |
|
980 | whether to track non-copying sends. | |
|
981 | [default False] | |
|
982 | ||
|
983 | targets : int,list of ints, 'all', None | |
|
984 | Specify the destination of the job. | |
|
985 | if None: | |
|
986 | Submit via Task queue for load-balancing. | |
|
987 | if 'all': | |
|
988 | Run on all active engines | |
|
989 | if list: | |
|
990 | Run on each specified engine | |
|
991 | if int: | |
|
992 | Run on single engine | |
|
993 | Note: | |
|
994 | that if `balanced=True`, and `targets` is specified, | |
|
995 | then the load-balancing will be limited to balancing | |
|
996 | among `targets`. | |
|
997 | ||
|
998 | balanced : bool, default None | |
|
999 | whether to load-balance. This will default to True | |
|
1000 | if targets is unspecified, or False if targets is specified. | |
|
1001 | ||
|
1002 | If `balanced` and `targets` are both specified, the task will | |
|
1003 | be assigne to *one* of the targets by the scheduler. | |
|
1004 | ||
|
1005 | The following arguments are only used when balanced is True: | |
|
1006 | ||
|
1007 | after : Dependency or collection of msg_ids | |
|
1008 | Only for load-balanced execution (targets=None) | |
|
1009 | Specify a list of msg_ids as a time-based dependency. | |
|
1010 | This job will only be run *after* the dependencies | |
|
1011 | have been met. | |
|
1012 | ||
|
1013 | follow : Dependency or collection of msg_ids | |
|
1014 | Only for load-balanced execution (targets=None) | |
|
1015 | Specify a list of msg_ids as a location-based dependency. | |
|
1016 | This job will only be run on an engine where this dependency | |
|
1017 | is met. | |
|
1018 | ||
|
1019 | timeout : float/int or None | |
|
1020 | Only for load-balanced execution (targets=None) | |
|
1021 | Specify an amount of time (in seconds) for the scheduler to | |
|
1022 | wait for dependencies to be met before failing with a | |
|
1023 | DependencyTimeout. | |
|
912 | def send_apply_message(self, socket, f, args=None, kwargs=None, subheader=None, track=False, | |
|
913 | ident=None): | |
|
914 | """construct and send an apply message via a socket. | |
|
1024 | 915 | |
|
1025 | Returns | |
|
1026 | ------- | |
|
1027 | ||
|
1028 | if block is False: | |
|
1029 | return AsyncResult wrapping msg_ids | |
|
1030 | output of AsyncResult.get() is identical to that of `apply(...block=True)` | |
|
1031 | else: | |
|
1032 | if single target (or balanced): | |
|
1033 | return result of `f(*args, **kwargs)` | |
|
1034 | else: | |
|
1035 | return list of results, matching `targets` | |
|
916 | This is the principal method with which all engine execution is performed by views. | |
|
1036 | 917 | """ |
|
918 | ||
|
1037 | 919 | assert not self._closed, "cannot use me anymore, I'm closed!" |
|
1038 | 920 | # defaults: |
|
1039 | block = block if block is not None else self.block | |
|
1040 | 921 | args = args if args is not None else [] |
|
1041 | 922 | kwargs = kwargs if kwargs is not None else {} |
|
923 | subheader = subheader if subheader is not None else {} | |
|
1042 | 924 | |
|
1043 | if not self._ids: | |
|
1044 | # flush notification socket if no engines yet | |
|
1045 | any_ids = self.ids | |
|
1046 | if not any_ids: | |
|
1047 | raise error.NoEnginesRegistered("Can't execute without any connected engines.") | |
|
1048 | ||
|
1049 | if balanced is None: | |
|
1050 | if targets is None: | |
|
1051 | # default to balanced if targets unspecified | |
|
1052 | balanced = True | |
|
1053 | else: | |
|
1054 | # otherwise default to multiplexing | |
|
1055 | balanced = False | |
|
1056 | ||
|
1057 | if targets is None and balanced is False: | |
|
1058 | # default to all if *not* balanced, and targets is unspecified | |
|
1059 | targets = 'all' | |
|
1060 | ||
|
1061 | # enforce types of f,args,kwrags | |
|
925 | # validate arguments | |
|
1062 | 926 | if not callable(f): |
|
1063 | 927 | raise TypeError("f must be callable, not %s"%type(f)) |
|
1064 | 928 | if not isinstance(args, (tuple, list)): |
|
1065 | 929 | raise TypeError("args must be tuple or list, not %s"%type(args)) |
|
1066 | 930 | if not isinstance(kwargs, dict): |
|
1067 | 931 | raise TypeError("kwargs must be dict, not %s"%type(kwargs)) |
|
932 | if not isinstance(subheader, dict): | |
|
933 | raise TypeError("subheader must be dict, not %s"%type(subheader)) | |
|
1068 | 934 | |
|
1069 | options = dict(bound=bound, block=block, targets=targets, track=track) | |
|
1070 | ||
|
1071 | if balanced: | |
|
1072 | return self._apply_balanced(f, args, kwargs, timeout=timeout, | |
|
1073 | after=after, follow=follow, **options) | |
|
1074 | elif follow or after or timeout: | |
|
1075 | msg = "follow, after, and timeout args are only used for" | |
|
1076 | msg += " load-balanced execution." | |
|
1077 | raise ValueError(msg) | |
|
1078 | else: | |
|
1079 | return self._apply_direct(f, args, kwargs, **options) | |
|
1080 | ||
|
1081 | def _apply_balanced(self, f, args, kwargs, bound=None, block=None, targets=None, | |
|
1082 | after=None, follow=None, timeout=None, track=None): | |
|
1083 | """call f(*args, **kwargs) remotely in a load-balanced manner. | |
|
1084 | ||
|
1085 | This is a private method, see `apply` for details. | |
|
1086 | Not to be called directly! | |
|
1087 | """ | |
|
1088 | ||
|
1089 | loc = locals() | |
|
1090 | for name in ('bound', 'block', 'track'): | |
|
1091 | assert loc[name] is not None, "kwarg %r must be specified!"%name | |
|
1092 | ||
|
1093 | if not self._task_ident: | |
|
1094 | msg = "Task farming is disabled" | |
|
1095 | if self._task_scheme == 'pure': | |
|
1096 | msg += " because the pure ZMQ scheduler cannot handle" | |
|
1097 | msg += " disappearing engines." | |
|
1098 | raise RuntimeError(msg) | |
|
1099 | ||
|
1100 | if self._task_scheme == 'pure': | |
|
1101 | # pure zmq scheme doesn't support dependencies | |
|
1102 | msg = "Pure ZMQ scheduler doesn't support dependencies" | |
|
1103 | if (follow or after): | |
|
1104 | # hard fail on DAG dependencies | |
|
1105 | raise RuntimeError(msg) | |
|
1106 | if isinstance(f, dependent): | |
|
1107 | # soft warn on functional dependencies | |
|
1108 | warnings.warn(msg, RuntimeWarning) | |
|
1109 | ||
|
1110 | # defaults: | |
|
1111 | args = args if args is not None else [] | |
|
1112 | kwargs = kwargs if kwargs is not None else {} | |
|
1113 | ||
|
1114 | if targets: | |
|
1115 | idents,_ = self._build_targets(targets) | |
|
1116 | else: | |
|
1117 | idents = [] | |
|
935 | if not self._ids: | |
|
936 | # flush notification socket if no engines yet | |
|
937 | any_ids = self.ids | |
|
938 | if not any_ids: | |
|
939 | raise error.NoEnginesRegistered("Can't execute without any connected engines.") | |
|
940 | # enforce types of f,args,kwargs | |
|
1118 | 941 | |
|
1119 | after = self._build_dependency(after) | |
|
1120 | follow = self._build_dependency(follow) | |
|
1121 | subheader = dict(after=after, follow=follow, timeout=timeout, targets=idents) | |
|
1122 | 942 | bufs = util.pack_apply_message(f,args,kwargs) |
|
1123 | content = dict(bound=bound) | |
|
1124 | 943 | |
|
1125 |
msg = self.session.send( |
|
|
1126 |
|
|
|
1127 | msg_id = msg['msg_id'] | |
|
1128 | self.outstanding.add(msg_id) | |
|
1129 | self.history.append(msg_id) | |
|
1130 | self.metadata[msg_id]['submitted'] = datetime.now() | |
|
1131 | tracker = None if track is False else msg['tracker'] | |
|
1132 | ar = AsyncResult(self, [msg_id], fname=f.__name__, targets=targets, tracker=tracker) | |
|
1133 | if block: | |
|
1134 | try: | |
|
1135 | return ar.get() | |
|
1136 | except KeyboardInterrupt: | |
|
1137 | return ar | |
|
1138 | else: | |
|
1139 | return ar | |
|
1140 | ||
|
1141 | def _apply_direct(self, f, args, kwargs, bound=None, block=None, targets=None, | |
|
1142 | track=None): | |
|
1143 | """Then underlying method for applying functions to specific engines | |
|
1144 | via the MUX queue. | |
|
1145 | ||
|
1146 | This is a private method, see `apply` for details. | |
|
1147 | Not to be called directly! | |
|
1148 | """ | |
|
1149 | ||
|
1150 | if not self._mux_ident: | |
|
1151 | msg = "Multiplexing is disabled" | |
|
1152 | raise RuntimeError(msg) | |
|
1153 | ||
|
1154 | loc = locals() | |
|
1155 | for name in ('bound', 'block', 'targets', 'track'): | |
|
1156 | assert loc[name] is not None, "kwarg %r must be specified!"%name | |
|
944 | msg = self.session.send(socket, "apply_request", buffers=bufs, ident=ident, | |
|
945 | subheader=subheader, track=track) | |
|
1157 | 946 | |
|
1158 | idents,targets = self._build_targets(targets) | |
|
1159 | ||
|
1160 | subheader = {} | |
|
1161 | content = dict(bound=bound) | |
|
1162 | bufs = util.pack_apply_message(f,args,kwargs) | |
|
1163 | ||
|
1164 | msg_ids = [] | |
|
1165 | trackers = [] | |
|
1166 | for ident in idents: | |
|
1167 | msg = self.session.send(self._apply_socket, "apply_request", | |
|
1168 | content=content, buffers=bufs, ident=[self._mux_ident, ident], subheader=subheader, | |
|
1169 | track=track) | |
|
1170 | if track: | |
|
1171 | trackers.append(msg['tracker']) | |
|
1172 | 947 |
|
|
1173 | 948 |
|
|
949 | if ident: | |
|
950 | # possibly routed to a specific engine | |
|
951 | if isinstance(ident, list): | |
|
952 | ident = ident[-1] | |
|
953 | if ident in self._engines.values(): | |
|
954 | # save for later, in case of engine death | |
|
1174 | 955 | self._outstanding_dict[ident].add(msg_id) |
|
1175 | 956 |
|
|
1176 | msg_ids.append(msg_id) | |
|
1177 | ||
|
1178 | tracker = None if track is False else zmq.MessageTracker(*trackers) | |
|
1179 | ar = AsyncResult(self, msg_ids, fname=f.__name__, targets=targets, tracker=tracker) | |
|
957 | self.metadata[msg_id]['submitted'] = datetime.now() | |
|
1180 | 958 | |
|
1181 | if block: | |
|
1182 | try: | |
|
1183 | return ar.get() | |
|
1184 | except KeyboardInterrupt: | |
|
1185 | return ar | |
|
1186 | else: | |
|
1187 | return ar | |
|
959 | return msg | |
|
1188 | 960 | |
|
1189 | 961 | #-------------------------------------------------------------------------- |
|
1190 | 962 | # construct a View object |
|
1191 | 963 | #-------------------------------------------------------------------------- |
|
1192 | 964 | |
|
1193 | @defaultblock | |
|
1194 | def remote(self, bound=False, block=None, targets=None, balanced=None): | |
|
1195 | """Decorator for making a RemoteFunction""" | |
|
1196 | return remote(self, bound=bound, targets=targets, block=block, balanced=balanced) | |
|
1197 | ||
|
1198 | @defaultblock | |
|
1199 | def parallel(self, dist='b', bound=False, block=None, targets=None, balanced=None): | |
|
1200 | """Decorator for making a ParallelFunction""" | |
|
1201 | return parallel(self, bound=bound, targets=targets, block=block, balanced=balanced) | |
|
1202 | ||
|
1203 | 965 | def _cache_view(self, targets, balanced): |
|
1204 | 966 | """save views, so subsequent requests don't create new objects.""" |
|
1205 | 967 | if balanced: |
|
968 | # validate whether we can run | |
|
969 | if not self._task_socket: | |
|
970 | msg = "Task farming is disabled" | |
|
971 | if self._task_scheme == 'pure': | |
|
972 | msg += " because the pure ZMQ scheduler cannot handle" | |
|
973 | msg += " disappearing engines." | |
|
974 | raise RuntimeError(msg) | |
|
975 | socket = self._task_socket | |
|
1206 | 976 | view_class = LoadBalancedView |
|
1207 | 977 | view_cache = self._balanced_views |
|
1208 | 978 | else: |
|
979 | socket = self._mux_socket | |
|
1209 | 980 | view_class = DirectView |
|
1210 | 981 | view_cache = self._direct_views |
|
1211 | 982 | |
|
1212 | 983 | # use str, since often targets will be a list |
|
1213 | 984 | key = str(targets) |
|
1214 | 985 | if key not in view_cache: |
|
1215 | view_cache[key] = view_class(client=self, targets=targets) | |
|
986 | view_cache[key] = view_class(client=self, socket=socket, targets=targets) | |
|
1216 | 987 | |
|
1217 | 988 | return view_cache[key] |
|
1218 | 989 | |
|
1219 |
def view(self, targets=None |
|
|
990 | def load_balanced_view(self, targets=None): | |
|
991 | """construct a DirectView object. | |
|
992 | ||
|
993 | If no arguments are specified, create a LoadBalancedView | |
|
994 | using all engines. | |
|
995 | ||
|
996 | Parameters | |
|
997 | ---------- | |
|
998 | ||
|
999 | targets: list,slice,int,etc. [default: use all engines] | |
|
1000 | The subset of engines across which to load-balance | |
|
1001 | """ | |
|
1002 | return self._get_view(targets, balanced=True) | |
|
1003 | ||
|
1004 | def direct_view(self, targets='all'): | |
|
1005 | """construct a DirectView object. | |
|
1006 | ||
|
1007 | If no targets are specified, create a DirectView | |
|
1008 | using all engines. | |
|
1009 | ||
|
1010 | Parameters | |
|
1011 | ---------- | |
|
1012 | ||
|
1013 | targets: list,slice,int,etc. [default: use all engines] | |
|
1014 | The engines to use for the View | |
|
1015 | """ | |
|
1016 | return self._get_view(targets, balanced=False) | |
|
1017 | ||
|
1018 | def _get_view(self, targets, balanced): | |
|
1220 | 1019 | """Method for constructing View objects. |
|
1221 | 1020 | |
|
1222 | 1021 | If no arguments are specified, create a LoadBalancedView |
|
1223 | 1022 | using all engines. If only `targets` specified, it will |
|
1224 | 1023 | be a DirectView. This method is the underlying implementation |
|
1225 | 1024 | of ``client.__getitem__``. |
|
1226 | 1025 | |
|
1227 | 1026 | Parameters |
|
1228 | 1027 | ---------- |
|
1229 | 1028 | |
|
1230 | 1029 | targets: list,slice,int,etc. [default: use all engines] |
|
1231 | 1030 | The engines to use for the View |
|
1232 | 1031 | balanced : bool [default: False if targets specified, True else] |
|
1233 | 1032 | whether to build a LoadBalancedView or a DirectView |
|
1234 | 1033 | |
|
1235 | 1034 | """ |
|
1236 | 1035 | |
|
1237 | balanced = (targets is None) if balanced is None else balanced | |
|
1238 | ||
|
1239 | if targets is None: | |
|
1036 | if targets in (None,'all'): | |
|
1240 | 1037 | if balanced: |
|
1241 | 1038 | return self._cache_view(None,True) |
|
1242 | 1039 | else: |
|
1243 | 1040 | targets = slice(None) |
|
1244 | 1041 | |
|
1245 | 1042 | if isinstance(targets, int): |
|
1246 | 1043 | if targets < 0: |
|
1247 | 1044 | targets = self.ids[targets] |
|
1248 | 1045 | if targets not in self.ids: |
|
1249 | 1046 | raise IndexError("No such engine: %i"%targets) |
|
1250 | 1047 | return self._cache_view(targets, balanced) |
|
1251 | 1048 | |
|
1252 | 1049 | if isinstance(targets, slice): |
|
1253 | 1050 | indices = range(len(self.ids))[targets] |
|
1254 | 1051 | ids = sorted(self._ids) |
|
1255 | 1052 | targets = [ ids[i] for i in indices ] |
|
1256 | 1053 | |
|
1257 | 1054 | if isinstance(targets, (tuple, list, xrange)): |
|
1258 | 1055 | _,targets = self._build_targets(list(targets)) |
|
1259 | 1056 | return self._cache_view(targets, balanced) |
|
1260 | 1057 | else: |
|
1261 | 1058 | raise TypeError("targets by int/slice/collection of ints only, not %s"%(type(targets))) |
|
1262 | 1059 | |
|
1263 | 1060 | #-------------------------------------------------------------------------- |
|
1264 | # Data movement | |
|
1061 | # Data movement (TO BE REMOVED) | |
|
1265 | 1062 | #-------------------------------------------------------------------------- |
|
1266 | 1063 | |
|
1267 | @defaultblock | |
|
1268 | def push(self, ns, targets='all', block=None, track=False): | |
|
1064 | @default_block | |
|
1065 | def _push(self, ns, targets='all', block=None, track=False): | |
|
1269 | 1066 | """Push the contents of `ns` into the namespace on `target`""" |
|
1270 | 1067 | if not isinstance(ns, dict): |
|
1271 | 1068 | raise TypeError("Must be a dict, not %s"%type(ns)) |
|
1272 | result = self.apply(_push, kwargs=ns, targets=targets, block=block, bound=True, balanced=False, track=track) | |
|
1069 | result = self.apply(util._push, kwargs=ns, targets=targets, block=block, bound=True, balanced=False, track=track) | |
|
1273 | 1070 | if not block: |
|
1274 | 1071 | return result |
|
1275 | 1072 | |
|
1276 | @defaultblock | |
|
1277 | def pull(self, keys, targets='all', block=None): | |
|
1073 | @default_block | |
|
1074 | def _pull(self, keys, targets='all', block=None): | |
|
1278 | 1075 | """Pull objects from `target`'s namespace by `keys`""" |
|
1279 | 1076 | if isinstance(keys, basestring): |
|
1280 | 1077 | pass |
|
1281 | 1078 | elif isinstance(keys, (list,tuple,set)): |
|
1282 | 1079 | for key in keys: |
|
1283 | 1080 | if not isinstance(key, basestring): |
|
1284 | 1081 | raise TypeError("keys must be str, not type %r"%type(key)) |
|
1285 | 1082 | else: |
|
1286 | 1083 | raise TypeError("keys must be strs, not %r"%keys) |
|
1287 | result = self.apply(_pull, (keys,), targets=targets, block=block, bound=True, balanced=False) | |
|
1084 | result = self.apply(util._pull, (keys,), targets=targets, block=block, bound=True, balanced=False) | |
|
1288 | 1085 | return result |
|
1289 | 1086 | |
|
1290 | @defaultblock | |
|
1291 | def scatter(self, key, seq, dist='b', flatten=False, targets='all', block=None, track=False): | |
|
1292 | """ | |
|
1293 | Partition a Python sequence and send the partitions to a set of engines. | |
|
1294 | """ | |
|
1295 | targets = self._build_targets(targets)[-1] | |
|
1296 | mapObject = Map.dists[dist]() | |
|
1297 | nparts = len(targets) | |
|
1298 | msg_ids = [] | |
|
1299 | trackers = [] | |
|
1300 | for index, engineid in enumerate(targets): | |
|
1301 | partition = mapObject.getPartition(seq, index, nparts) | |
|
1302 | if flatten and len(partition) == 1: | |
|
1303 | r = self.push({key: partition[0]}, targets=engineid, block=False, track=track) | |
|
1304 | else: | |
|
1305 | r = self.push({key: partition}, targets=engineid, block=False, track=track) | |
|
1306 | msg_ids.extend(r.msg_ids) | |
|
1307 | if track: | |
|
1308 | trackers.append(r._tracker) | |
|
1309 | ||
|
1310 | if track: | |
|
1311 | tracker = zmq.MessageTracker(*trackers) | |
|
1312 | else: | |
|
1313 | tracker = None | |
|
1314 | ||
|
1315 | r = AsyncResult(self, msg_ids, fname='scatter', targets=targets, tracker=tracker) | |
|
1316 | if block: | |
|
1317 | r.wait() | |
|
1318 | else: | |
|
1319 | return r | |
|
1320 | ||
|
1321 | @defaultblock | |
|
1322 | def gather(self, key, dist='b', targets='all', block=None): | |
|
1323 | """ | |
|
1324 | Gather a partitioned sequence on a set of engines as a single local seq. | |
|
1325 | """ | |
|
1326 | ||
|
1327 | targets = self._build_targets(targets)[-1] | |
|
1328 | mapObject = Map.dists[dist]() | |
|
1329 | msg_ids = [] | |
|
1330 | for index, engineid in enumerate(targets): | |
|
1331 | msg_ids.extend(self.pull(key, targets=engineid,block=False).msg_ids) | |
|
1332 | ||
|
1333 | r = AsyncMapResult(self, msg_ids, mapObject, fname='gather') | |
|
1334 | if block: | |
|
1335 | return r.get() | |
|
1336 | else: | |
|
1337 | return r | |
|
1338 | ||
|
1339 | 1087 | #-------------------------------------------------------------------------- |
|
1340 | 1088 | # Query methods |
|
1341 | 1089 | #-------------------------------------------------------------------------- |
|
1342 | 1090 | |
|
1343 | @spinfirst | |
|
1344 | @defaultblock | |
|
1091 | @spin_first | |
|
1092 | @default_block | |
|
1345 | 1093 | def get_result(self, indices_or_msg_ids=None, block=None): |
|
1346 | 1094 | """Retrieve a result by msg_id or history index, wrapped in an AsyncResult object. |
|
1347 | 1095 | |
|
1348 | 1096 | If the client already has the results, no request to the Hub will be made. |
|
1349 | 1097 | |
|
1350 | 1098 | This is a convenient way to construct AsyncResult objects, which are wrappers |
|
1351 | 1099 | that include metadata about execution, and allow for awaiting results that |
|
1352 | 1100 | were not submitted by this Client. |
|
1353 | 1101 | |
|
1354 | 1102 | It can also be a convenient way to retrieve the metadata associated with |
|
1355 | 1103 | blocking execution, since it always retrieves |
|
1356 | 1104 | |
|
1357 | 1105 | Examples |
|
1358 | 1106 | -------- |
|
1359 | 1107 | :: |
|
1360 | 1108 | |
|
1361 | 1109 | In [10]: r = client.apply() |
|
1362 | 1110 | |
|
1363 | 1111 | Parameters |
|
1364 | 1112 | ---------- |
|
1365 | 1113 | |
|
1366 | 1114 | indices_or_msg_ids : integer history index, str msg_id, or list of either |
|
1367 | 1115 | The indices or msg_ids of indices to be retrieved |
|
1368 | 1116 | |
|
1369 | 1117 | block : bool |
|
1370 | 1118 | Whether to wait for the result to be done |
|
1371 | 1119 | |
|
1372 | 1120 | Returns |
|
1373 | 1121 | ------- |
|
1374 | 1122 | |
|
1375 | 1123 | AsyncResult |
|
1376 | 1124 | A single AsyncResult object will always be returned. |
|
1377 | 1125 | |
|
1378 | 1126 | AsyncHubResult |
|
1379 | 1127 | A subclass of AsyncResult that retrieves results from the Hub |
|
1380 | 1128 | |
|
1381 | 1129 | """ |
|
1382 | 1130 | if indices_or_msg_ids is None: |
|
1383 | 1131 | indices_or_msg_ids = -1 |
|
1384 | 1132 | |
|
1385 | 1133 | if not isinstance(indices_or_msg_ids, (list,tuple)): |
|
1386 | 1134 | indices_or_msg_ids = [indices_or_msg_ids] |
|
1387 | 1135 | |
|
1388 | 1136 | theids = [] |
|
1389 | 1137 | for id in indices_or_msg_ids: |
|
1390 | 1138 | if isinstance(id, int): |
|
1391 | 1139 | id = self.history[id] |
|
1392 | 1140 | if not isinstance(id, str): |
|
1393 | 1141 | raise TypeError("indices must be str or int, not %r"%id) |
|
1394 | 1142 | theids.append(id) |
|
1395 | 1143 | |
|
1396 | 1144 | local_ids = filter(lambda msg_id: msg_id in self.history or msg_id in self.results, theids) |
|
1397 | 1145 | remote_ids = filter(lambda msg_id: msg_id not in local_ids, theids) |
|
1398 | 1146 | |
|
1399 | 1147 | if remote_ids: |
|
1400 | 1148 | ar = AsyncHubResult(self, msg_ids=theids) |
|
1401 | 1149 | else: |
|
1402 | 1150 | ar = AsyncResult(self, msg_ids=theids) |
|
1403 | 1151 | |
|
1404 | 1152 | if block: |
|
1405 | 1153 | ar.wait() |
|
1406 | 1154 | |
|
1407 | 1155 | return ar |
|
1408 | 1156 | |
|
1409 | @spinfirst | |
|
1157 | @spin_first | |
|
1410 | 1158 | def result_status(self, msg_ids, status_only=True): |
|
1411 | 1159 | """Check on the status of the result(s) of the apply request with `msg_ids`. |
|
1412 | 1160 | |
|
1413 | 1161 | If status_only is False, then the actual results will be retrieved, else |
|
1414 | 1162 | only the status of the results will be checked. |
|
1415 | 1163 | |
|
1416 | 1164 | Parameters |
|
1417 | 1165 | ---------- |
|
1418 | 1166 | |
|
1419 | 1167 | msg_ids : list of msg_ids |
|
1420 | 1168 | if int: |
|
1421 | 1169 | Passed as index to self.history for convenience. |
|
1422 | 1170 | status_only : bool (default: True) |
|
1423 | 1171 | if False: |
|
1424 | 1172 | Retrieve the actual results of completed tasks. |
|
1425 | 1173 | |
|
1426 | 1174 | Returns |
|
1427 | 1175 | ------- |
|
1428 | 1176 | |
|
1429 | 1177 | results : dict |
|
1430 | 1178 | There will always be the keys 'pending' and 'completed', which will |
|
1431 | 1179 | be lists of msg_ids that are incomplete or complete. If `status_only` |
|
1432 | 1180 | is False, then completed results will be keyed by their `msg_id`. |
|
1433 | 1181 | """ |
|
1434 | 1182 | if not isinstance(msg_ids, (list,tuple)): |
|
1435 | 1183 | msg_ids = [msg_ids] |
|
1436 | 1184 | |
|
1437 | 1185 | theids = [] |
|
1438 | 1186 | for msg_id in msg_ids: |
|
1439 | 1187 | if isinstance(msg_id, int): |
|
1440 | 1188 | msg_id = self.history[msg_id] |
|
1441 | 1189 | if not isinstance(msg_id, basestring): |
|
1442 | 1190 | raise TypeError("msg_ids must be str, not %r"%msg_id) |
|
1443 | 1191 | theids.append(msg_id) |
|
1444 | 1192 | |
|
1445 | 1193 | completed = [] |
|
1446 | 1194 | local_results = {} |
|
1447 | 1195 | |
|
1448 | 1196 | # comment this block out to temporarily disable local shortcut: |
|
1449 | 1197 | for msg_id in theids: |
|
1450 | 1198 | if msg_id in self.results: |
|
1451 | 1199 | completed.append(msg_id) |
|
1452 | 1200 | local_results[msg_id] = self.results[msg_id] |
|
1453 | 1201 | theids.remove(msg_id) |
|
1454 | 1202 | |
|
1455 | 1203 | if theids: # some not locally cached |
|
1456 | 1204 | content = dict(msg_ids=theids, status_only=status_only) |
|
1457 | 1205 | msg = self.session.send(self._query_socket, "result_request", content=content) |
|
1458 | 1206 | zmq.select([self._query_socket], [], []) |
|
1459 | 1207 | idents,msg = self.session.recv(self._query_socket, zmq.NOBLOCK) |
|
1460 | 1208 | if self.debug: |
|
1461 | 1209 | pprint(msg) |
|
1462 | 1210 | content = msg['content'] |
|
1463 | 1211 | if content['status'] != 'ok': |
|
1464 | 1212 | raise self._unwrap_exception(content) |
|
1465 | 1213 | buffers = msg['buffers'] |
|
1466 | 1214 | else: |
|
1467 | 1215 | content = dict(completed=[],pending=[]) |
|
1468 | 1216 | |
|
1469 | 1217 | content['completed'].extend(completed) |
|
1470 | 1218 | |
|
1471 | 1219 | if status_only: |
|
1472 | 1220 | return content |
|
1473 | 1221 | |
|
1474 | 1222 | failures = [] |
|
1475 | 1223 | # load cached results into result: |
|
1476 | 1224 | content.update(local_results) |
|
1477 | 1225 | # update cache with results: |
|
1478 | 1226 | for msg_id in sorted(theids): |
|
1479 | 1227 | if msg_id in content['completed']: |
|
1480 | 1228 | rec = content[msg_id] |
|
1481 | 1229 | parent = rec['header'] |
|
1482 | 1230 | header = rec['result_header'] |
|
1483 | 1231 | rcontent = rec['result_content'] |
|
1484 | 1232 | iodict = rec['io'] |
|
1485 | 1233 | if isinstance(rcontent, str): |
|
1486 | 1234 | rcontent = self.session.unpack(rcontent) |
|
1487 | 1235 | |
|
1488 | 1236 | md = self.metadata[msg_id] |
|
1489 | 1237 | md.update(self._extract_metadata(header, parent, rcontent)) |
|
1490 | 1238 | md.update(iodict) |
|
1491 | 1239 | |
|
1492 | 1240 | if rcontent['status'] == 'ok': |
|
1493 | 1241 | res,buffers = util.unserialize_object(buffers) |
|
1494 | 1242 | else: |
|
1495 | 1243 | print rcontent |
|
1496 | 1244 | res = self._unwrap_exception(rcontent) |
|
1497 | 1245 | failures.append(res) |
|
1498 | 1246 | |
|
1499 | 1247 | self.results[msg_id] = res |
|
1500 | 1248 | content[msg_id] = res |
|
1501 | 1249 | |
|
1502 | 1250 | if len(theids) == 1 and failures: |
|
1503 | 1251 | raise failures[0] |
|
1504 | 1252 | |
|
1505 | 1253 | error.collect_exceptions(failures, "result_status") |
|
1506 | 1254 | return content |
|
1507 | 1255 | |
|
1508 | @spinfirst | |
|
1256 | @spin_first | |
|
1509 | 1257 | def queue_status(self, targets='all', verbose=False): |
|
1510 | 1258 | """Fetch the status of engine queues. |
|
1511 | 1259 | |
|
1512 | 1260 | Parameters |
|
1513 | 1261 | ---------- |
|
1514 | 1262 | |
|
1515 | 1263 | targets : int/str/list of ints/strs |
|
1516 | 1264 | the engines whose states are to be queried. |
|
1517 | 1265 | default : all |
|
1518 | 1266 | verbose : bool |
|
1519 | 1267 | Whether to return lengths only, or lists of ids for each element |
|
1520 | 1268 | """ |
|
1521 |
|
|
|
1522 |
content = dict(targets= |
|
|
1269 | engine_ids = self._build_targets(targets)[1] | |
|
1270 | content = dict(targets=engine_ids, verbose=verbose) | |
|
1523 | 1271 | self.session.send(self._query_socket, "queue_request", content=content) |
|
1524 | 1272 | idents,msg = self.session.recv(self._query_socket, 0) |
|
1525 | 1273 | if self.debug: |
|
1526 | 1274 | pprint(msg) |
|
1527 | 1275 | content = msg['content'] |
|
1528 | 1276 | status = content.pop('status') |
|
1529 | 1277 | if status != 'ok': |
|
1530 | 1278 | raise self._unwrap_exception(content) |
|
1531 |
|
|
|
1279 | content = util.rekey(content) | |
|
1280 | if isinstance(targets, int): | |
|
1281 | return content[targets] | |
|
1282 | else: | |
|
1283 | return content | |
|
1532 | 1284 | |
|
1533 | @spinfirst | |
|
1285 | @spin_first | |
|
1534 | 1286 | def purge_results(self, jobs=[], targets=[]): |
|
1535 |
"""Tell the |
|
|
1287 | """Tell the Hub to forget results. | |
|
1536 | 1288 | |
|
1537 | 1289 | Individual results can be purged by msg_id, or the entire |
|
1538 | 1290 | history of specific targets can be purged. |
|
1539 | 1291 | |
|
1540 | 1292 | Parameters |
|
1541 | 1293 | ---------- |
|
1542 | 1294 | |
|
1543 |
jobs : str or list of str |
|
|
1295 | jobs : str or list of str or AsyncResult objects | |
|
1544 | 1296 | the msg_ids whose results should be forgotten. |
|
1545 | 1297 | targets : int/str/list of ints/strs |
|
1546 | 1298 | The targets, by uuid or int_id, whose entire history is to be purged. |
|
1547 |
Use `targets='all'` to scrub everything from the |
|
|
1299 | Use `targets='all'` to scrub everything from the Hub's memory. | |
|
1548 | 1300 | |
|
1549 | 1301 | default : None |
|
1550 | 1302 | """ |
|
1551 | 1303 | if not targets and not jobs: |
|
1552 | 1304 | raise ValueError("Must specify at least one of `targets` and `jobs`") |
|
1553 | 1305 | if targets: |
|
1554 | 1306 | targets = self._build_targets(targets)[1] |
|
1555 | 1307 | |
|
1556 | 1308 | # construct msg_ids from jobs |
|
1557 | 1309 | msg_ids = [] |
|
1558 | 1310 | if isinstance(jobs, (basestring,AsyncResult)): |
|
1559 | 1311 | jobs = [jobs] |
|
1560 | 1312 | bad_ids = filter(lambda obj: not isinstance(obj, (basestring, AsyncResult)), jobs) |
|
1561 | 1313 | if bad_ids: |
|
1562 | 1314 | raise TypeError("Invalid msg_id type %r, expected str or AsyncResult"%bad_ids[0]) |
|
1563 | 1315 | for j in jobs: |
|
1564 | 1316 | if isinstance(j, AsyncResult): |
|
1565 | 1317 | msg_ids.extend(j.msg_ids) |
|
1566 | 1318 | else: |
|
1567 | 1319 | msg_ids.append(j) |
|
1568 | 1320 | |
|
1569 | 1321 | content = dict(targets=targets, msg_ids=msg_ids) |
|
1570 | 1322 | self.session.send(self._query_socket, "purge_request", content=content) |
|
1571 | 1323 | idents, msg = self.session.recv(self._query_socket, 0) |
|
1572 | 1324 | if self.debug: |
|
1573 | 1325 | pprint(msg) |
|
1574 | 1326 | content = msg['content'] |
|
1575 | 1327 | if content['status'] != 'ok': |
|
1576 | 1328 | raise self._unwrap_exception(content) |
|
1577 | 1329 | |
|
1578 | 1330 | |
|
1579 | 1331 | __all__ = [ 'Client', |
|
1580 | 1332 | 'depend', |
|
1581 | 1333 | 'require', |
|
1582 | 1334 | 'remote', |
|
1583 | 1335 | 'parallel', |
|
1584 | 1336 | 'RemoteFunction', |
|
1585 | 1337 | 'ParallelFunction', |
|
1586 | 1338 | 'DirectView', |
|
1587 | 1339 | 'LoadBalancedView', |
|
1588 | 1340 | 'AsyncResult', |
|
1589 | 1341 | 'AsyncMapResult', |
|
1590 | 1342 | 'Reference' |
|
1591 | 1343 | ] |
@@ -1,159 +1,184 b'' | |||
|
1 | 1 | """Dependency utilities""" |
|
2 | 2 | #----------------------------------------------------------------------------- |
|
3 | 3 | # Copyright (C) 2010-2011 The IPython Development Team |
|
4 | 4 | # |
|
5 | 5 | # Distributed under the terms of the BSD License. The full license is in |
|
6 | 6 | # the file COPYING, distributed as part of this software. |
|
7 | 7 | #----------------------------------------------------------------------------- |
|
8 | 8 | |
|
9 | from IPython.external.decorator import decorator | |
|
10 | ||
|
11 | 9 | from .asyncresult import AsyncResult |
|
12 | 10 | from .error import UnmetDependency |
|
13 | ||
|
11 | from .util import interactive | |
|
14 | 12 | |
|
15 | 13 | class depend(object): |
|
16 | 14 | """Dependency decorator, for use with tasks. |
|
17 | 15 | |
|
18 | 16 | `@depend` lets you define a function for engine dependencies |
|
19 | 17 | just like you use `apply` for tasks. |
|
20 | 18 | |
|
21 | 19 | |
|
22 | 20 | Examples |
|
23 | 21 | -------- |
|
24 | 22 | :: |
|
25 | 23 | |
|
26 | 24 | @depend(df, a,b, c=5) |
|
27 | 25 | def f(m,n,p) |
|
28 | 26 | |
|
29 | 27 | view.apply(f, 1,2,3) |
|
30 | 28 | |
|
31 | 29 | will call df(a,b,c=5) on the engine, and if it returns False or |
|
32 | 30 | raises an UnmetDependency error, then the task will not be run |
|
33 | 31 | and another engine will be tried. |
|
34 | 32 | """ |
|
35 | 33 | def __init__(self, f, *args, **kwargs): |
|
36 | 34 | self.f = f |
|
37 | 35 | self.args = args |
|
38 | 36 | self.kwargs = kwargs |
|
39 | 37 | |
|
40 | 38 | def __call__(self, f): |
|
41 | 39 | return dependent(f, self.f, *self.args, **self.kwargs) |
|
42 | 40 | |
|
43 | 41 | class dependent(object): |
|
44 | 42 | """A function that depends on another function. |
|
45 | 43 | This is an object to prevent the closure used |
|
46 | 44 | in traditional decorators, which are not picklable. |
|
47 | 45 | """ |
|
48 | 46 | |
|
49 | 47 | def __init__(self, f, df, *dargs, **dkwargs): |
|
50 | 48 | self.f = f |
|
51 | 49 | self.func_name = getattr(f, '__name__', 'f') |
|
52 | 50 | self.df = df |
|
53 | 51 | self.dargs = dargs |
|
54 | 52 | self.dkwargs = dkwargs |
|
55 | 53 | |
|
56 | 54 | def __call__(self, *args, **kwargs): |
|
55 | # if hasattr(self.f, 'func_globals') and hasattr(self.df, 'func_globals'): | |
|
56 | # self.df.func_globals = self.f.func_globals | |
|
57 | 57 | if self.df(*self.dargs, **self.dkwargs) is False: |
|
58 | 58 | raise UnmetDependency() |
|
59 | 59 | return self.f(*args, **kwargs) |
|
60 | 60 | |
|
61 | 61 | @property |
|
62 | 62 | def __name__(self): |
|
63 | 63 | return self.func_name |
|
64 | 64 | |
|
65 | @interactive | |
|
65 | 66 | def _require(*names): |
|
66 | 67 | """Helper for @require decorator.""" |
|
68 | from IPython.zmq.parallel.error import UnmetDependency | |
|
69 | user_ns = globals() | |
|
67 | 70 | for name in names: |
|
71 | if name in user_ns: | |
|
72 | continue | |
|
68 | 73 | try: |
|
69 |
|
|
|
74 | exec 'import %s'%name in user_ns | |
|
70 | 75 | except ImportError: |
|
71 | return False | |
|
76 | raise UnmetDependency(name) | |
|
72 | 77 | return True |
|
73 | 78 | |
|
74 | 79 | def require(*names): |
|
75 | 80 | """Simple decorator for requiring names to be importable. |
|
76 | 81 | |
|
77 | 82 | Examples |
|
78 | 83 | -------- |
|
79 | 84 | |
|
80 | 85 | In [1]: @require('numpy') |
|
81 | 86 | ...: def norm(a): |
|
82 | 87 | ...: import numpy |
|
83 | 88 | ...: return numpy.linalg.norm(a,2) |
|
84 | 89 | """ |
|
85 | 90 | return depend(_require, *names) |
|
86 | 91 | |
|
87 | 92 | class Dependency(set): |
|
88 | 93 | """An object for representing a set of msg_id dependencies. |
|
89 | 94 | |
|
90 | 95 | Subclassed from set(). |
|
91 | 96 | |
|
92 | 97 | Parameters |
|
93 | 98 | ---------- |
|
94 | 99 | dependencies: list/set of msg_ids or AsyncResult objects or output of Dependency.as_dict() |
|
95 | 100 | The msg_ids to depend on |
|
96 | 101 | all : bool [default True] |
|
97 | 102 | Whether the dependency should be considered met when *all* depending tasks have completed |
|
98 | 103 | or only when *any* have been completed. |
|
99 |
success |
|
|
100 |
Whether to consider |
|
|
101 | If `all=success_only=True`, then this task will fail with an ImpossibleDependency | |
|
104 | success : bool [default True] | |
|
105 | Whether to consider successes as fulfilling dependencies. | |
|
106 | failure : bool [default False] | |
|
107 | Whether to consider failures as fulfilling dependencies. | |
|
108 | ||
|
109 | If `all=success=True` and `failure=False`, then the task will fail with an ImpossibleDependency | |
|
102 | 110 | as soon as the first depended-upon task fails. |
|
103 | 111 | """ |
|
104 | 112 | |
|
105 | 113 | all=True |
|
106 |
success |
|
|
114 | success=True | |
|
115 | failure=True | |
|
107 | 116 | |
|
108 |
def __init__(self, dependencies=[], all=True, success |
|
|
117 | def __init__(self, dependencies=[], all=True, success=True, failure=False): | |
|
109 | 118 | if isinstance(dependencies, dict): |
|
110 | 119 | # load from dict |
|
111 | 120 | all = dependencies.get('all', True) |
|
112 |
success |
|
|
121 | success = dependencies.get('success', success) | |
|
122 | failure = dependencies.get('failure', failure) | |
|
113 | 123 | dependencies = dependencies.get('dependencies', []) |
|
114 | 124 | ids = [] |
|
115 | if isinstance(dependencies, AsyncResult): | |
|
116 | ids.extend(AsyncResult.msg_ids) | |
|
117 | else: | |
|
125 | ||
|
126 | # extract ids from various sources: | |
|
127 | if isinstance(dependencies, (basestring, AsyncResult)): | |
|
128 | dependencies = [dependencies] | |
|
118 | 129 |
|
|
119 | 130 |
|
|
120 | 131 |
|
|
121 | 132 |
|
|
122 | 133 |
|
|
123 | 134 |
|
|
124 | 135 |
|
|
136 | ||
|
125 | 137 | set.__init__(self, ids) |
|
126 | 138 | self.all = all |
|
127 | self.success_only=success_only | |
|
139 | if not (success or failure): | |
|
140 | raise ValueError("Must depend on at least one of successes or failures!") | |
|
141 | self.success=success | |
|
142 | self.failure = failure | |
|
128 | 143 | |
|
129 | 144 | def check(self, completed, failed=None): |
|
130 | if failed is not None and not self.success_only: | |
|
131 | completed = completed.union(failed) | |
|
145 | """check whether our dependencies have been met.""" | |
|
132 | 146 | if len(self) == 0: |
|
133 | 147 | return True |
|
148 | against = set() | |
|
149 | if self.success: | |
|
150 | against = completed | |
|
151 | if failed is not None and self.failure: | |
|
152 | against = against.union(failed) | |
|
134 | 153 | if self.all: |
|
135 |
return self.issubset( |
|
|
154 | return self.issubset(against) | |
|
136 | 155 | else: |
|
137 |
return not self.isdisjoint( |
|
|
156 | return not self.isdisjoint(against) | |
|
138 | 157 | |
|
139 | def unreachable(self, failed): | |
|
140 | if len(self) == 0 or len(failed) == 0 or not self.success_only: | |
|
158 | def unreachable(self, completed, failed=None): | |
|
159 | """return whether this dependency has become impossible.""" | |
|
160 | if len(self) == 0: | |
|
141 | 161 | return False |
|
142 | # print self, self.success_only, self.all, failed | |
|
162 | against = set() | |
|
163 | if not self.success: | |
|
164 | against = completed | |
|
165 | if failed is not None and not self.failure: | |
|
166 | against = against.union(failed) | |
|
143 | 167 | if self.all: |
|
144 |
return not self.isdisjoint( |
|
|
168 | return not self.isdisjoint(against) | |
|
145 | 169 | else: |
|
146 |
return self.issubset( |
|
|
170 | return self.issubset(against) | |
|
147 | 171 | |
|
148 | 172 | |
|
149 | 173 | def as_dict(self): |
|
150 | 174 | """Represent this dependency as a dict. For json compatibility.""" |
|
151 | 175 | return dict( |
|
152 | 176 | dependencies=list(self), |
|
153 | 177 | all=self.all, |
|
154 |
success |
|
|
178 | success=self.success, | |
|
179 | failure=self.failure | |
|
155 | 180 | ) |
|
156 | 181 | |
|
157 | 182 | |
|
158 | 183 | __all__ = ['depend', 'require', 'dependent', 'Dependency'] |
|
159 | 184 |
@@ -1,1039 +1,1035 b'' | |||
|
1 | 1 | #!/usr/bin/env python |
|
2 | 2 | """The IPython Controller Hub with 0MQ |
|
3 | 3 | This is the master object that handles connections from engines and clients, |
|
4 | 4 | and monitors traffic through the various queues. |
|
5 | 5 | """ |
|
6 | 6 | #----------------------------------------------------------------------------- |
|
7 | 7 | # Copyright (C) 2010 The IPython Development Team |
|
8 | 8 | # |
|
9 | 9 | # Distributed under the terms of the BSD License. The full license is in |
|
10 | 10 | # the file COPYING, distributed as part of this software. |
|
11 | 11 | #----------------------------------------------------------------------------- |
|
12 | 12 | |
|
13 | 13 | #----------------------------------------------------------------------------- |
|
14 | 14 | # Imports |
|
15 | 15 | #----------------------------------------------------------------------------- |
|
16 | 16 | from __future__ import print_function |
|
17 | 17 | |
|
18 | 18 | import sys |
|
19 | 19 | import time |
|
20 | 20 | from datetime import datetime |
|
21 | 21 | |
|
22 | 22 | import zmq |
|
23 | 23 | from zmq.eventloop import ioloop |
|
24 | 24 | from zmq.eventloop.zmqstream import ZMQStream |
|
25 | 25 | |
|
26 | 26 | # internal: |
|
27 | 27 | from IPython.utils.importstring import import_item |
|
28 | 28 | from IPython.utils.traitlets import HasTraits, Instance, Int, CStr, Str, Dict, Set, List, Bool |
|
29 | 29 | |
|
30 | 30 | from .entry_point import select_random_ports |
|
31 | 31 | from .factory import RegistrationFactory, LoggingFactory |
|
32 | 32 | |
|
33 | 33 | from . import error |
|
34 | 34 | from .heartmonitor import HeartMonitor |
|
35 | 35 | from .util import validate_url_container, ISO8601 |
|
36 | 36 | |
|
37 | 37 | #----------------------------------------------------------------------------- |
|
38 | 38 | # Code |
|
39 | 39 | #----------------------------------------------------------------------------- |
|
40 | 40 | |
|
41 | 41 | def _passer(*args, **kwargs): |
|
42 | 42 | return |
|
43 | 43 | |
|
44 | 44 | def _printer(*args, **kwargs): |
|
45 | 45 | print (args) |
|
46 | 46 | print (kwargs) |
|
47 | 47 | |
|
48 | 48 | def init_record(msg): |
|
49 | 49 | """Initialize a TaskRecord based on a request.""" |
|
50 | 50 | header = msg['header'] |
|
51 | 51 | return { |
|
52 | 52 | 'msg_id' : header['msg_id'], |
|
53 | 53 | 'header' : header, |
|
54 | 54 | 'content': msg['content'], |
|
55 | 55 | 'buffers': msg['buffers'], |
|
56 | 56 | 'submitted': datetime.strptime(header['date'], ISO8601), |
|
57 | 57 | 'client_uuid' : None, |
|
58 | 58 | 'engine_uuid' : None, |
|
59 | 59 | 'started': None, |
|
60 | 60 | 'completed': None, |
|
61 | 61 | 'resubmitted': None, |
|
62 | 62 | 'result_header' : None, |
|
63 | 63 | 'result_content' : None, |
|
64 | 64 | 'result_buffers' : None, |
|
65 | 65 | 'queue' : None, |
|
66 | 66 | 'pyin' : None, |
|
67 | 67 | 'pyout': None, |
|
68 | 68 | 'pyerr': None, |
|
69 | 69 | 'stdout': '', |
|
70 | 70 | 'stderr': '', |
|
71 | 71 | } |
|
72 | 72 | |
|
73 | 73 | |
|
74 | 74 | class EngineConnector(HasTraits): |
|
75 | 75 | """A simple object for accessing the various zmq connections of an object. |
|
76 | 76 | Attributes are: |
|
77 | 77 | id (int): engine ID |
|
78 | 78 | uuid (str): uuid (unused?) |
|
79 | 79 | queue (str): identity of queue's XREQ socket |
|
80 | 80 | registration (str): identity of registration XREQ socket |
|
81 | 81 | heartbeat (str): identity of heartbeat XREQ socket |
|
82 | 82 | """ |
|
83 | 83 | id=Int(0) |
|
84 | 84 | queue=Str() |
|
85 | 85 | control=Str() |
|
86 | 86 | registration=Str() |
|
87 | 87 | heartbeat=Str() |
|
88 | 88 | pending=Set() |
|
89 | 89 | |
|
90 | 90 | class HubFactory(RegistrationFactory): |
|
91 | 91 | """The Configurable for setting up a Hub.""" |
|
92 | 92 | |
|
93 | 93 | # name of a scheduler scheme |
|
94 | 94 | scheme = Str('leastload', config=True) |
|
95 | 95 | |
|
96 | 96 | # port-pairs for monitoredqueues: |
|
97 | 97 | hb = Instance(list, config=True) |
|
98 | 98 | def _hb_default(self): |
|
99 | 99 | return select_random_ports(2) |
|
100 | 100 | |
|
101 | 101 | mux = Instance(list, config=True) |
|
102 | 102 | def _mux_default(self): |
|
103 | 103 | return select_random_ports(2) |
|
104 | 104 | |
|
105 | 105 | task = Instance(list, config=True) |
|
106 | 106 | def _task_default(self): |
|
107 | 107 | return select_random_ports(2) |
|
108 | 108 | |
|
109 | 109 | control = Instance(list, config=True) |
|
110 | 110 | def _control_default(self): |
|
111 | 111 | return select_random_ports(2) |
|
112 | 112 | |
|
113 | 113 | iopub = Instance(list, config=True) |
|
114 | 114 | def _iopub_default(self): |
|
115 | 115 | return select_random_ports(2) |
|
116 | 116 | |
|
117 | 117 | # single ports: |
|
118 | 118 | mon_port = Instance(int, config=True) |
|
119 | 119 | def _mon_port_default(self): |
|
120 | 120 | return select_random_ports(1)[0] |
|
121 | 121 | |
|
122 | 122 | notifier_port = Instance(int, config=True) |
|
123 | 123 | def _notifier_port_default(self): |
|
124 | 124 | return select_random_ports(1)[0] |
|
125 | 125 | |
|
126 | 126 | ping = Int(1000, config=True) # ping frequency |
|
127 | 127 | |
|
128 | 128 | engine_ip = CStr('127.0.0.1', config=True) |
|
129 | 129 | engine_transport = CStr('tcp', config=True) |
|
130 | 130 | |
|
131 | 131 | client_ip = CStr('127.0.0.1', config=True) |
|
132 | 132 | client_transport = CStr('tcp', config=True) |
|
133 | 133 | |
|
134 | 134 | monitor_ip = CStr('127.0.0.1', config=True) |
|
135 | 135 | monitor_transport = CStr('tcp', config=True) |
|
136 | 136 | |
|
137 | 137 | monitor_url = CStr('') |
|
138 | 138 | |
|
139 | 139 | db_class = CStr('IPython.zmq.parallel.dictdb.DictDB', config=True) |
|
140 | 140 | |
|
141 | 141 | # not configurable |
|
142 | 142 | db = Instance('IPython.zmq.parallel.dictdb.BaseDB') |
|
143 | 143 | heartmonitor = Instance('IPython.zmq.parallel.heartmonitor.HeartMonitor') |
|
144 | 144 | subconstructors = List() |
|
145 | 145 | _constructed = Bool(False) |
|
146 | 146 | |
|
147 | 147 | def _ip_changed(self, name, old, new): |
|
148 | 148 | self.engine_ip = new |
|
149 | 149 | self.client_ip = new |
|
150 | 150 | self.monitor_ip = new |
|
151 | 151 | self._update_monitor_url() |
|
152 | 152 | |
|
153 | 153 | def _update_monitor_url(self): |
|
154 | 154 | self.monitor_url = "%s://%s:%i"%(self.monitor_transport, self.monitor_ip, self.mon_port) |
|
155 | 155 | |
|
156 | 156 | def _transport_changed(self, name, old, new): |
|
157 | 157 | self.engine_transport = new |
|
158 | 158 | self.client_transport = new |
|
159 | 159 | self.monitor_transport = new |
|
160 | 160 | self._update_monitor_url() |
|
161 | 161 | |
|
162 | 162 | def __init__(self, **kwargs): |
|
163 | 163 | super(HubFactory, self).__init__(**kwargs) |
|
164 | 164 | self._update_monitor_url() |
|
165 | 165 | # self.on_trait_change(self._sync_ips, 'ip') |
|
166 | 166 | # self.on_trait_change(self._sync_transports, 'transport') |
|
167 | 167 | self.subconstructors.append(self.construct_hub) |
|
168 | 168 | |
|
169 | 169 | |
|
170 | 170 | def construct(self): |
|
171 | 171 | assert not self._constructed, "already constructed!" |
|
172 | 172 | |
|
173 | 173 | for subc in self.subconstructors: |
|
174 | 174 | subc() |
|
175 | 175 | |
|
176 | 176 | self._constructed = True |
|
177 | 177 | |
|
178 | 178 | |
|
179 | 179 | def start(self): |
|
180 | 180 | assert self._constructed, "must be constructed by self.construct() first!" |
|
181 | 181 | self.heartmonitor.start() |
|
182 | 182 | self.log.info("Heartmonitor started") |
|
183 | 183 | |
|
184 | 184 | def construct_hub(self): |
|
185 | 185 | """construct""" |
|
186 | 186 | client_iface = "%s://%s:"%(self.client_transport, self.client_ip) + "%i" |
|
187 | 187 | engine_iface = "%s://%s:"%(self.engine_transport, self.engine_ip) + "%i" |
|
188 | 188 | |
|
189 | 189 | ctx = self.context |
|
190 | 190 | loop = self.loop |
|
191 | 191 | |
|
192 | 192 | # Registrar socket |
|
193 | 193 | q = ZMQStream(ctx.socket(zmq.XREP), loop) |
|
194 | 194 | q.bind(client_iface % self.regport) |
|
195 | 195 | self.log.info("Hub listening on %s for registration."%(client_iface%self.regport)) |
|
196 | 196 | if self.client_ip != self.engine_ip: |
|
197 | 197 | q.bind(engine_iface % self.regport) |
|
198 | 198 | self.log.info("Hub listening on %s for registration."%(engine_iface%self.regport)) |
|
199 | 199 | |
|
200 | 200 | ### Engine connections ### |
|
201 | 201 | |
|
202 | 202 | # heartbeat |
|
203 | 203 | hpub = ctx.socket(zmq.PUB) |
|
204 | 204 | hpub.bind(engine_iface % self.hb[0]) |
|
205 | 205 | hrep = ctx.socket(zmq.XREP) |
|
206 | 206 | hrep.bind(engine_iface % self.hb[1]) |
|
207 | 207 | self.heartmonitor = HeartMonitor(loop=loop, pingstream=ZMQStream(hpub,loop), pongstream=ZMQStream(hrep,loop), |
|
208 | 208 | period=self.ping, logname=self.log.name) |
|
209 | 209 | |
|
210 | 210 | ### Client connections ### |
|
211 | 211 | # Notifier socket |
|
212 | 212 | n = ZMQStream(ctx.socket(zmq.PUB), loop) |
|
213 | 213 | n.bind(client_iface%self.notifier_port) |
|
214 | 214 | |
|
215 | 215 | ### build and launch the queues ### |
|
216 | 216 | |
|
217 | 217 | # monitor socket |
|
218 | 218 | sub = ctx.socket(zmq.SUB) |
|
219 | 219 | sub.setsockopt(zmq.SUBSCRIBE, "") |
|
220 | 220 | sub.bind(self.monitor_url) |
|
221 | 221 | sub.bind('inproc://monitor') |
|
222 | 222 | sub = ZMQStream(sub, loop) |
|
223 | 223 | |
|
224 | 224 | # connect the db |
|
225 | 225 | self.log.info('Hub using DB backend: %r'%(self.db_class.split()[-1])) |
|
226 | 226 | # cdir = self.config.Global.cluster_dir |
|
227 | 227 | self.db = import_item(self.db_class)(session=self.session.session, config=self.config) |
|
228 | 228 | time.sleep(.25) |
|
229 | 229 | |
|
230 | 230 | # build connection dicts |
|
231 | 231 | self.engine_info = { |
|
232 | 232 | 'control' : engine_iface%self.control[1], |
|
233 | 233 | 'mux': engine_iface%self.mux[1], |
|
234 | 234 | 'heartbeat': (engine_iface%self.hb[0], engine_iface%self.hb[1]), |
|
235 | 235 | 'task' : engine_iface%self.task[1], |
|
236 | 236 | 'iopub' : engine_iface%self.iopub[1], |
|
237 | 237 | # 'monitor' : engine_iface%self.mon_port, |
|
238 | 238 | } |
|
239 | 239 | |
|
240 | 240 | self.client_info = { |
|
241 | 241 | 'control' : client_iface%self.control[0], |
|
242 | 242 | 'mux': client_iface%self.mux[0], |
|
243 | 243 | 'task' : (self.scheme, client_iface%self.task[0]), |
|
244 | 244 | 'iopub' : client_iface%self.iopub[0], |
|
245 | 245 | 'notification': client_iface%self.notifier_port |
|
246 | 246 | } |
|
247 | 247 | self.log.debug("Hub engine addrs: %s"%self.engine_info) |
|
248 | 248 | self.log.debug("Hub client addrs: %s"%self.client_info) |
|
249 | 249 | self.hub = Hub(loop=loop, session=self.session, monitor=sub, heartmonitor=self.heartmonitor, |
|
250 | 250 | query=q, notifier=n, db=self.db, |
|
251 | 251 | engine_info=self.engine_info, client_info=self.client_info, |
|
252 | 252 | logname=self.log.name) |
|
253 | 253 | |
|
254 | 254 | |
|
255 | 255 | class Hub(LoggingFactory): |
|
256 | 256 | """The IPython Controller Hub with 0MQ connections |
|
257 | 257 | |
|
258 | 258 | Parameters |
|
259 | 259 | ========== |
|
260 | 260 | loop: zmq IOLoop instance |
|
261 | 261 | session: StreamSession object |
|
262 | 262 | <removed> context: zmq context for creating new connections (?) |
|
263 | 263 | queue: ZMQStream for monitoring the command queue (SUB) |
|
264 | 264 | query: ZMQStream for engine registration and client queries requests (XREP) |
|
265 | 265 | heartbeat: HeartMonitor object checking the pulse of the engines |
|
266 | 266 | notifier: ZMQStream for broadcasting engine registration changes (PUB) |
|
267 | 267 | db: connection to db for out of memory logging of commands |
|
268 | 268 | NotImplemented |
|
269 | 269 | engine_info: dict of zmq connection information for engines to connect |
|
270 | 270 | to the queues. |
|
271 | 271 | client_info: dict of zmq connection information for engines to connect |
|
272 | 272 | to the queues. |
|
273 | 273 | """ |
|
274 | 274 | # internal data structures: |
|
275 | 275 | ids=Set() # engine IDs |
|
276 | 276 | keytable=Dict() |
|
277 | 277 | by_ident=Dict() |
|
278 | 278 | engines=Dict() |
|
279 | 279 | clients=Dict() |
|
280 | 280 | hearts=Dict() |
|
281 | 281 | pending=Set() |
|
282 | 282 | queues=Dict() # pending msg_ids keyed by engine_id |
|
283 | 283 | tasks=Dict() # pending msg_ids submitted as tasks, keyed by client_id |
|
284 | 284 | completed=Dict() # completed msg_ids keyed by engine_id |
|
285 | 285 | all_completed=Set() # completed msg_ids keyed by engine_id |
|
286 | 286 | # mia=None |
|
287 | 287 | incoming_registrations=Dict() |
|
288 | 288 | registration_timeout=Int() |
|
289 | 289 | _idcounter=Int(0) |
|
290 | 290 | |
|
291 | 291 | # objects from constructor: |
|
292 | 292 | loop=Instance(ioloop.IOLoop) |
|
293 | 293 | query=Instance(ZMQStream) |
|
294 | 294 | monitor=Instance(ZMQStream) |
|
295 | 295 | heartmonitor=Instance(HeartMonitor) |
|
296 | 296 | notifier=Instance(ZMQStream) |
|
297 | 297 | db=Instance(object) |
|
298 | 298 | client_info=Dict() |
|
299 | 299 | engine_info=Dict() |
|
300 | 300 | |
|
301 | 301 | |
|
302 | 302 | def __init__(self, **kwargs): |
|
303 | 303 | """ |
|
304 | 304 | # universal: |
|
305 | 305 | loop: IOLoop for creating future connections |
|
306 | 306 | session: streamsession for sending serialized data |
|
307 | 307 | # engine: |
|
308 | 308 | queue: ZMQStream for monitoring queue messages |
|
309 | 309 | query: ZMQStream for engine+client registration and client requests |
|
310 | 310 | heartbeat: HeartMonitor object for tracking engines |
|
311 | 311 | # extra: |
|
312 | 312 | db: ZMQStream for db connection (NotImplemented) |
|
313 | 313 | engine_info: zmq address/protocol dict for engine connections |
|
314 | 314 | client_info: zmq address/protocol dict for client connections |
|
315 | 315 | """ |
|
316 | 316 | |
|
317 | 317 | super(Hub, self).__init__(**kwargs) |
|
318 | 318 | self.registration_timeout = max(5000, 2*self.heartmonitor.period) |
|
319 | 319 | |
|
320 | 320 | # validate connection dicts: |
|
321 | 321 | for k,v in self.client_info.iteritems(): |
|
322 | 322 | if k == 'task': |
|
323 | 323 | validate_url_container(v[1]) |
|
324 | 324 | else: |
|
325 | 325 | validate_url_container(v) |
|
326 | 326 | # validate_url_container(self.client_info) |
|
327 | 327 | validate_url_container(self.engine_info) |
|
328 | 328 | |
|
329 | 329 | # register our callbacks |
|
330 | 330 | self.query.on_recv(self.dispatch_query) |
|
331 | 331 | self.monitor.on_recv(self.dispatch_monitor_traffic) |
|
332 | 332 | |
|
333 | 333 | self.heartmonitor.add_heart_failure_handler(self.handle_heart_failure) |
|
334 | 334 | self.heartmonitor.add_new_heart_handler(self.handle_new_heart) |
|
335 | 335 | |
|
336 | 336 | self.monitor_handlers = { 'in' : self.save_queue_request, |
|
337 | 337 | 'out': self.save_queue_result, |
|
338 | 338 | 'intask': self.save_task_request, |
|
339 | 339 | 'outtask': self.save_task_result, |
|
340 | 340 | 'tracktask': self.save_task_destination, |
|
341 | 341 | 'incontrol': _passer, |
|
342 | 342 | 'outcontrol': _passer, |
|
343 | 343 | 'iopub': self.save_iopub_message, |
|
344 | 344 | } |
|
345 | 345 | |
|
346 | 346 | self.query_handlers = {'queue_request': self.queue_status, |
|
347 | 347 | 'result_request': self.get_results, |
|
348 | 348 | 'purge_request': self.purge_results, |
|
349 | 349 | 'load_request': self.check_load, |
|
350 | 350 | 'resubmit_request': self.resubmit_task, |
|
351 | 351 | 'shutdown_request': self.shutdown_request, |
|
352 | 352 | 'registration_request' : self.register_engine, |
|
353 | 353 | 'unregistration_request' : self.unregister_engine, |
|
354 | 354 | 'connection_request': self.connection_request, |
|
355 | 355 | } |
|
356 | 356 | |
|
357 | 357 | self.log.info("hub::created hub") |
|
358 | 358 | |
|
359 | 359 | @property |
|
360 | 360 | def _next_id(self): |
|
361 | 361 | """gemerate a new ID. |
|
362 | 362 | |
|
363 | 363 | No longer reuse old ids, just count from 0.""" |
|
364 | 364 | newid = self._idcounter |
|
365 | 365 | self._idcounter += 1 |
|
366 | 366 | return newid |
|
367 | 367 | # newid = 0 |
|
368 | 368 | # incoming = [id[0] for id in self.incoming_registrations.itervalues()] |
|
369 | 369 | # # print newid, self.ids, self.incoming_registrations |
|
370 | 370 | # while newid in self.ids or newid in incoming: |
|
371 | 371 | # newid += 1 |
|
372 | 372 | # return newid |
|
373 | 373 | |
|
374 | 374 | #----------------------------------------------------------------------------- |
|
375 | 375 | # message validation |
|
376 | 376 | #----------------------------------------------------------------------------- |
|
377 | 377 | |
|
378 | 378 | def _validate_targets(self, targets): |
|
379 | 379 | """turn any valid targets argument into a list of integer ids""" |
|
380 | 380 | if targets is None: |
|
381 | 381 | # default to all |
|
382 | 382 | targets = self.ids |
|
383 | 383 | |
|
384 | 384 | if isinstance(targets, (int,str,unicode)): |
|
385 | 385 | # only one target specified |
|
386 | 386 | targets = [targets] |
|
387 | 387 | _targets = [] |
|
388 | 388 | for t in targets: |
|
389 | 389 | # map raw identities to ids |
|
390 | 390 | if isinstance(t, (str,unicode)): |
|
391 | 391 | t = self.by_ident.get(t, t) |
|
392 | 392 | _targets.append(t) |
|
393 | 393 | targets = _targets |
|
394 | 394 | bad_targets = [ t for t in targets if t not in self.ids ] |
|
395 | 395 | if bad_targets: |
|
396 | 396 | raise IndexError("No Such Engine: %r"%bad_targets) |
|
397 | 397 | if not targets: |
|
398 | 398 | raise IndexError("No Engines Registered") |
|
399 | 399 | return targets |
|
400 | 400 | |
|
401 | 401 | #----------------------------------------------------------------------------- |
|
402 | 402 | # dispatch methods (1 per stream) |
|
403 | 403 | #----------------------------------------------------------------------------- |
|
404 | 404 | |
|
405 | 405 | # def dispatch_registration_request(self, msg): |
|
406 | 406 | # """""" |
|
407 | 407 | # self.log.debug("registration::dispatch_register_request(%s)"%msg) |
|
408 | 408 | # idents,msg = self.session.feed_identities(msg) |
|
409 | 409 | # if not idents: |
|
410 | 410 | # self.log.error("Bad Query Message: %s"%msg, exc_info=True) |
|
411 | 411 | # return |
|
412 | 412 | # try: |
|
413 | 413 | # msg = self.session.unpack_message(msg,content=True) |
|
414 | 414 | # except: |
|
415 | 415 | # self.log.error("registration::got bad registration message: %s"%msg, exc_info=True) |
|
416 | 416 | # return |
|
417 | 417 | # |
|
418 | 418 | # msg_type = msg['msg_type'] |
|
419 | 419 | # content = msg['content'] |
|
420 | 420 | # |
|
421 | 421 | # handler = self.query_handlers.get(msg_type, None) |
|
422 | 422 | # if handler is None: |
|
423 | 423 | # self.log.error("registration::got bad registration message: %s"%msg) |
|
424 | 424 | # else: |
|
425 | 425 | # handler(idents, msg) |
|
426 | 426 | |
|
427 | 427 | def dispatch_monitor_traffic(self, msg): |
|
428 | 428 | """all ME and Task queue messages come through here, as well as |
|
429 | 429 | IOPub traffic.""" |
|
430 | 430 | self.log.debug("monitor traffic: %s"%msg[:2]) |
|
431 | 431 | switch = msg[0] |
|
432 | 432 | idents, msg = self.session.feed_identities(msg[1:]) |
|
433 | 433 | if not idents: |
|
434 | 434 | self.log.error("Bad Monitor Message: %s"%msg) |
|
435 | 435 | return |
|
436 | 436 | handler = self.monitor_handlers.get(switch, None) |
|
437 | 437 | if handler is not None: |
|
438 | 438 | handler(idents, msg) |
|
439 | 439 | else: |
|
440 | 440 | self.log.error("Invalid monitor topic: %s"%switch) |
|
441 | 441 | |
|
442 | 442 | |
|
443 | 443 | def dispatch_query(self, msg): |
|
444 | 444 | """Route registration requests and queries from clients.""" |
|
445 | 445 | idents, msg = self.session.feed_identities(msg) |
|
446 | 446 | if not idents: |
|
447 | 447 | self.log.error("Bad Query Message: %s"%msg) |
|
448 | 448 | return |
|
449 | 449 | client_id = idents[0] |
|
450 | 450 | try: |
|
451 | 451 | msg = self.session.unpack_message(msg, content=True) |
|
452 | 452 | except: |
|
453 | 453 | content = error.wrap_exception() |
|
454 | 454 | self.log.error("Bad Query Message: %s"%msg, exc_info=True) |
|
455 | 455 | self.session.send(self.query, "hub_error", ident=client_id, |
|
456 | 456 | content=content) |
|
457 | 457 | return |
|
458 | 458 | |
|
459 | 459 | # print client_id, header, parent, content |
|
460 | 460 | #switch on message type: |
|
461 | 461 | msg_type = msg['msg_type'] |
|
462 | 462 | self.log.info("client::client %s requested %s"%(client_id, msg_type)) |
|
463 | 463 | handler = self.query_handlers.get(msg_type, None) |
|
464 | 464 | try: |
|
465 | 465 | assert handler is not None, "Bad Message Type: %s"%msg_type |
|
466 | 466 | except: |
|
467 | 467 | content = error.wrap_exception() |
|
468 | 468 | self.log.error("Bad Message Type: %s"%msg_type, exc_info=True) |
|
469 | 469 | self.session.send(self.query, "hub_error", ident=client_id, |
|
470 | 470 | content=content) |
|
471 | 471 | return |
|
472 | 472 | else: |
|
473 | 473 | handler(idents, msg) |
|
474 | 474 | |
|
475 | 475 | def dispatch_db(self, msg): |
|
476 | 476 | """""" |
|
477 | 477 | raise NotImplementedError |
|
478 | 478 | |
|
479 | 479 | #--------------------------------------------------------------------------- |
|
480 | 480 | # handler methods (1 per event) |
|
481 | 481 | #--------------------------------------------------------------------------- |
|
482 | 482 | |
|
483 | 483 | #----------------------- Heartbeat -------------------------------------- |
|
484 | 484 | |
|
485 | 485 | def handle_new_heart(self, heart): |
|
486 | 486 | """handler to attach to heartbeater. |
|
487 | 487 | Called when a new heart starts to beat. |
|
488 | 488 | Triggers completion of registration.""" |
|
489 | 489 | self.log.debug("heartbeat::handle_new_heart(%r)"%heart) |
|
490 | 490 | if heart not in self.incoming_registrations: |
|
491 | 491 | self.log.info("heartbeat::ignoring new heart: %r"%heart) |
|
492 | 492 | else: |
|
493 | 493 | self.finish_registration(heart) |
|
494 | 494 | |
|
495 | 495 | |
|
496 | 496 | def handle_heart_failure(self, heart): |
|
497 | 497 | """handler to attach to heartbeater. |
|
498 | 498 | called when a previously registered heart fails to respond to beat request. |
|
499 | 499 | triggers unregistration""" |
|
500 | 500 | self.log.debug("heartbeat::handle_heart_failure(%r)"%heart) |
|
501 | 501 | eid = self.hearts.get(heart, None) |
|
502 | 502 | queue = self.engines[eid].queue |
|
503 | 503 | if eid is None: |
|
504 | 504 | self.log.info("heartbeat::ignoring heart failure %r"%heart) |
|
505 | 505 | else: |
|
506 | 506 | self.unregister_engine(heart, dict(content=dict(id=eid, queue=queue))) |
|
507 | 507 | |
|
508 | 508 | #----------------------- MUX Queue Traffic ------------------------------ |
|
509 | 509 | |
|
510 | 510 | def save_queue_request(self, idents, msg): |
|
511 | 511 | if len(idents) < 2: |
|
512 | 512 | self.log.error("invalid identity prefix: %s"%idents) |
|
513 | 513 | return |
|
514 | 514 | queue_id, client_id = idents[:2] |
|
515 | 515 | try: |
|
516 | 516 | msg = self.session.unpack_message(msg, content=False) |
|
517 | 517 | except: |
|
518 | 518 | self.log.error("queue::client %r sent invalid message to %r: %s"%(client_id, queue_id, msg), exc_info=True) |
|
519 | 519 | return |
|
520 | 520 | |
|
521 | 521 | eid = self.by_ident.get(queue_id, None) |
|
522 | 522 | if eid is None: |
|
523 | 523 | self.log.error("queue::target %r not registered"%queue_id) |
|
524 | 524 | self.log.debug("queue:: valid are: %s"%(self.by_ident.keys())) |
|
525 | 525 | return |
|
526 | 526 | |
|
527 | 527 | header = msg['header'] |
|
528 | 528 | msg_id = header['msg_id'] |
|
529 | 529 | record = init_record(msg) |
|
530 | 530 | record['engine_uuid'] = queue_id |
|
531 | 531 | record['client_uuid'] = client_id |
|
532 | 532 | record['queue'] = 'mux' |
|
533 | 533 | |
|
534 | 534 | self.pending.add(msg_id) |
|
535 | 535 | self.queues[eid].append(msg_id) |
|
536 | 536 | self.db.add_record(msg_id, record) |
|
537 | 537 | |
|
538 | 538 | def save_queue_result(self, idents, msg): |
|
539 | 539 | if len(idents) < 2: |
|
540 | 540 | self.log.error("invalid identity prefix: %s"%idents) |
|
541 | 541 | return |
|
542 | 542 | |
|
543 | 543 | client_id, queue_id = idents[:2] |
|
544 | 544 | try: |
|
545 | 545 | msg = self.session.unpack_message(msg, content=False) |
|
546 | 546 | except: |
|
547 | 547 | self.log.error("queue::engine %r sent invalid message to %r: %s"%( |
|
548 | 548 | queue_id,client_id, msg), exc_info=True) |
|
549 | 549 | return |
|
550 | 550 | |
|
551 | 551 | eid = self.by_ident.get(queue_id, None) |
|
552 | 552 | if eid is None: |
|
553 | 553 | self.log.error("queue::unknown engine %r is sending a reply: "%queue_id) |
|
554 | 554 | self.log.debug("queue:: %s"%msg[2:]) |
|
555 | 555 | return |
|
556 | 556 | |
|
557 | 557 | parent = msg['parent_header'] |
|
558 | 558 | if not parent: |
|
559 | 559 | return |
|
560 | 560 | msg_id = parent['msg_id'] |
|
561 | 561 | if msg_id in self.pending: |
|
562 | 562 | self.pending.remove(msg_id) |
|
563 | 563 | self.all_completed.add(msg_id) |
|
564 | 564 | self.queues[eid].remove(msg_id) |
|
565 | 565 | self.completed[eid].append(msg_id) |
|
566 | 566 | elif msg_id not in self.all_completed: |
|
567 | 567 | # it could be a result from a dead engine that died before delivering the |
|
568 | 568 | # result |
|
569 | 569 | self.log.warn("queue:: unknown msg finished %s"%msg_id) |
|
570 | 570 | return |
|
571 | 571 | # update record anyway, because the unregistration could have been premature |
|
572 | 572 | rheader = msg['header'] |
|
573 | 573 | completed = datetime.strptime(rheader['date'], ISO8601) |
|
574 | 574 | started = rheader.get('started', None) |
|
575 | 575 | if started is not None: |
|
576 | 576 | started = datetime.strptime(started, ISO8601) |
|
577 | 577 | result = { |
|
578 | 578 | 'result_header' : rheader, |
|
579 | 579 | 'result_content': msg['content'], |
|
580 | 580 | 'started' : started, |
|
581 | 581 | 'completed' : completed |
|
582 | 582 | } |
|
583 | 583 | |
|
584 | 584 | result['result_buffers'] = msg['buffers'] |
|
585 | 585 | self.db.update_record(msg_id, result) |
|
586 | 586 | |
|
587 | 587 | |
|
588 | 588 | #--------------------- Task Queue Traffic ------------------------------ |
|
589 | 589 | |
|
590 | 590 | def save_task_request(self, idents, msg): |
|
591 | 591 | """Save the submission of a task.""" |
|
592 | 592 | client_id = idents[0] |
|
593 | 593 | |
|
594 | 594 | try: |
|
595 | 595 | msg = self.session.unpack_message(msg, content=False) |
|
596 | 596 | except: |
|
597 | 597 | self.log.error("task::client %r sent invalid task message: %s"%( |
|
598 | 598 | client_id, msg), exc_info=True) |
|
599 | 599 | return |
|
600 | 600 | record = init_record(msg) |
|
601 | 601 | |
|
602 | 602 | record['client_uuid'] = client_id |
|
603 | 603 | record['queue'] = 'task' |
|
604 | 604 | header = msg['header'] |
|
605 | 605 | msg_id = header['msg_id'] |
|
606 | 606 | self.pending.add(msg_id) |
|
607 | 607 | self.db.add_record(msg_id, record) |
|
608 | 608 | |
|
609 | 609 | def save_task_result(self, idents, msg): |
|
610 | 610 | """save the result of a completed task.""" |
|
611 | 611 | client_id = idents[0] |
|
612 | 612 | try: |
|
613 | 613 | msg = self.session.unpack_message(msg, content=False) |
|
614 | 614 | except: |
|
615 | 615 | self.log.error("task::invalid task result message send to %r: %s"%( |
|
616 | 616 | client_id, msg), exc_info=True) |
|
617 | 617 | raise |
|
618 | 618 | return |
|
619 | 619 | |
|
620 | 620 | parent = msg['parent_header'] |
|
621 | 621 | if not parent: |
|
622 | 622 | # print msg |
|
623 | 623 | self.log.warn("Task %r had no parent!"%msg) |
|
624 | 624 | return |
|
625 | 625 | msg_id = parent['msg_id'] |
|
626 | 626 | |
|
627 | 627 | header = msg['header'] |
|
628 | 628 | engine_uuid = header.get('engine', None) |
|
629 | 629 | eid = self.by_ident.get(engine_uuid, None) |
|
630 | 630 | |
|
631 | 631 | if msg_id in self.pending: |
|
632 | 632 | self.pending.remove(msg_id) |
|
633 | 633 | self.all_completed.add(msg_id) |
|
634 | 634 | if eid is not None: |
|
635 | 635 | self.completed[eid].append(msg_id) |
|
636 | 636 | if msg_id in self.tasks[eid]: |
|
637 | 637 | self.tasks[eid].remove(msg_id) |
|
638 | 638 | completed = datetime.strptime(header['date'], ISO8601) |
|
639 | 639 | started = header.get('started', None) |
|
640 | 640 | if started is not None: |
|
641 | 641 | started = datetime.strptime(started, ISO8601) |
|
642 | 642 | result = { |
|
643 | 643 | 'result_header' : header, |
|
644 | 644 | 'result_content': msg['content'], |
|
645 | 645 | 'started' : started, |
|
646 | 646 | 'completed' : completed, |
|
647 | 647 | 'engine_uuid': engine_uuid |
|
648 | 648 | } |
|
649 | 649 | |
|
650 | 650 | result['result_buffers'] = msg['buffers'] |
|
651 | 651 | self.db.update_record(msg_id, result) |
|
652 | 652 | |
|
653 | 653 | else: |
|
654 | 654 | self.log.debug("task::unknown task %s finished"%msg_id) |
|
655 | 655 | |
|
656 | 656 | def save_task_destination(self, idents, msg): |
|
657 | 657 | try: |
|
658 | 658 | msg = self.session.unpack_message(msg, content=True) |
|
659 | 659 | except: |
|
660 | 660 | self.log.error("task::invalid task tracking message", exc_info=True) |
|
661 | 661 | return |
|
662 | 662 | content = msg['content'] |
|
663 | 663 | # print (content) |
|
664 | 664 | msg_id = content['msg_id'] |
|
665 | 665 | engine_uuid = content['engine_id'] |
|
666 | 666 | eid = self.by_ident[engine_uuid] |
|
667 | 667 | |
|
668 | 668 | self.log.info("task::task %s arrived on %s"%(msg_id, eid)) |
|
669 | 669 | # if msg_id in self.mia: |
|
670 | 670 | # self.mia.remove(msg_id) |
|
671 | 671 | # else: |
|
672 | 672 | # self.log.debug("task::task %s not listed as MIA?!"%(msg_id)) |
|
673 | 673 | |
|
674 | 674 | self.tasks[eid].append(msg_id) |
|
675 | 675 | # self.pending[msg_id][1].update(received=datetime.now(),engine=(eid,engine_uuid)) |
|
676 | 676 | self.db.update_record(msg_id, dict(engine_uuid=engine_uuid)) |
|
677 | 677 | |
|
678 | 678 | def mia_task_request(self, idents, msg): |
|
679 | 679 | raise NotImplementedError |
|
680 | 680 | client_id = idents[0] |
|
681 | 681 | # content = dict(mia=self.mia,status='ok') |
|
682 | 682 | # self.session.send('mia_reply', content=content, idents=client_id) |
|
683 | 683 | |
|
684 | 684 | |
|
685 | 685 | #--------------------- IOPub Traffic ------------------------------ |
|
686 | 686 | |
|
687 | 687 | def save_iopub_message(self, topics, msg): |
|
688 | 688 | """save an iopub message into the db""" |
|
689 | 689 | # print (topics) |
|
690 | 690 | try: |
|
691 | 691 | msg = self.session.unpack_message(msg, content=True) |
|
692 | 692 | except: |
|
693 | 693 | self.log.error("iopub::invalid IOPub message", exc_info=True) |
|
694 | 694 | return |
|
695 | 695 | |
|
696 | 696 | parent = msg['parent_header'] |
|
697 | 697 | if not parent: |
|
698 | 698 | self.log.error("iopub::invalid IOPub message: %s"%msg) |
|
699 | 699 | return |
|
700 | 700 | msg_id = parent['msg_id'] |
|
701 | 701 | msg_type = msg['msg_type'] |
|
702 | 702 | content = msg['content'] |
|
703 | 703 | |
|
704 | 704 | # ensure msg_id is in db |
|
705 | 705 | try: |
|
706 | 706 | rec = self.db.get_record(msg_id) |
|
707 | 707 | except: |
|
708 | 708 | self.log.error("iopub::IOPub message has invalid parent", exc_info=True) |
|
709 | 709 | return |
|
710 | 710 | # stream |
|
711 | 711 | d = {} |
|
712 | 712 | if msg_type == 'stream': |
|
713 | 713 | name = content['name'] |
|
714 | 714 | s = rec[name] or '' |
|
715 | 715 | d[name] = s + content['data'] |
|
716 | 716 | |
|
717 | 717 | elif msg_type == 'pyerr': |
|
718 | 718 | d['pyerr'] = content |
|
719 | 719 | else: |
|
720 | 720 | d[msg_type] = content['data'] |
|
721 | 721 | |
|
722 | 722 | self.db.update_record(msg_id, d) |
|
723 | 723 | |
|
724 | 724 | |
|
725 | 725 | |
|
726 | 726 | #------------------------------------------------------------------------- |
|
727 | 727 | # Registration requests |
|
728 | 728 | #------------------------------------------------------------------------- |
|
729 | 729 | |
|
730 | 730 | def connection_request(self, client_id, msg): |
|
731 | 731 | """Reply with connection addresses for clients.""" |
|
732 | 732 | self.log.info("client::client %s connected"%client_id) |
|
733 | 733 | content = dict(status='ok') |
|
734 | 734 | content.update(self.client_info) |
|
735 | 735 | jsonable = {} |
|
736 | 736 | for k,v in self.keytable.iteritems(): |
|
737 | 737 | jsonable[str(k)] = v |
|
738 | 738 | content['engines'] = jsonable |
|
739 | 739 | self.session.send(self.query, 'connection_reply', content, parent=msg, ident=client_id) |
|
740 | 740 | |
|
741 | 741 | def register_engine(self, reg, msg): |
|
742 | 742 | """Register a new engine.""" |
|
743 | 743 | content = msg['content'] |
|
744 | 744 | try: |
|
745 | 745 | queue = content['queue'] |
|
746 | 746 | except KeyError: |
|
747 | 747 | self.log.error("registration::queue not specified", exc_info=True) |
|
748 | 748 | return |
|
749 | 749 | heart = content.get('heartbeat', None) |
|
750 | 750 | """register a new engine, and create the socket(s) necessary""" |
|
751 | 751 | eid = self._next_id |
|
752 | 752 | # print (eid, queue, reg, heart) |
|
753 | 753 | |
|
754 | 754 | self.log.debug("registration::register_engine(%i, %r, %r, %r)"%(eid, queue, reg, heart)) |
|
755 | 755 | |
|
756 | 756 | content = dict(id=eid,status='ok') |
|
757 | 757 | content.update(self.engine_info) |
|
758 | 758 | # check if requesting available IDs: |
|
759 | 759 | if queue in self.by_ident: |
|
760 | 760 | try: |
|
761 | 761 | raise KeyError("queue_id %r in use"%queue) |
|
762 | 762 | except: |
|
763 | 763 | content = error.wrap_exception() |
|
764 | 764 | self.log.error("queue_id %r in use"%queue, exc_info=True) |
|
765 | 765 | elif heart in self.hearts: # need to check unique hearts? |
|
766 | 766 | try: |
|
767 | 767 | raise KeyError("heart_id %r in use"%heart) |
|
768 | 768 | except: |
|
769 | 769 | self.log.error("heart_id %r in use"%heart, exc_info=True) |
|
770 | 770 | content = error.wrap_exception() |
|
771 | 771 | else: |
|
772 | 772 | for h, pack in self.incoming_registrations.iteritems(): |
|
773 | 773 | if heart == h: |
|
774 | 774 | try: |
|
775 | 775 | raise KeyError("heart_id %r in use"%heart) |
|
776 | 776 | except: |
|
777 | 777 | self.log.error("heart_id %r in use"%heart, exc_info=True) |
|
778 | 778 | content = error.wrap_exception() |
|
779 | 779 | break |
|
780 | 780 | elif queue == pack[1]: |
|
781 | 781 | try: |
|
782 | 782 | raise KeyError("queue_id %r in use"%queue) |
|
783 | 783 | except: |
|
784 | 784 | self.log.error("queue_id %r in use"%queue, exc_info=True) |
|
785 | 785 | content = error.wrap_exception() |
|
786 | 786 | break |
|
787 | 787 | |
|
788 | 788 | msg = self.session.send(self.query, "registration_reply", |
|
789 | 789 | content=content, |
|
790 | 790 | ident=reg) |
|
791 | 791 | |
|
792 | 792 | if content['status'] == 'ok': |
|
793 | 793 | if heart in self.heartmonitor.hearts: |
|
794 | 794 | # already beating |
|
795 | 795 | self.incoming_registrations[heart] = (eid,queue,reg[0],None) |
|
796 | 796 | self.finish_registration(heart) |
|
797 | 797 | else: |
|
798 | 798 | purge = lambda : self._purge_stalled_registration(heart) |
|
799 | 799 | dc = ioloop.DelayedCallback(purge, self.registration_timeout, self.loop) |
|
800 | 800 | dc.start() |
|
801 | 801 | self.incoming_registrations[heart] = (eid,queue,reg[0],dc) |
|
802 | 802 | else: |
|
803 | 803 | self.log.error("registration::registration %i failed: %s"%(eid, content['evalue'])) |
|
804 | 804 | return eid |
|
805 | 805 | |
|
806 | 806 | def unregister_engine(self, ident, msg): |
|
807 | 807 | """Unregister an engine that explicitly requested to leave.""" |
|
808 | 808 | try: |
|
809 | 809 | eid = msg['content']['id'] |
|
810 | 810 | except: |
|
811 | 811 | self.log.error("registration::bad engine id for unregistration: %s"%ident, exc_info=True) |
|
812 | 812 | return |
|
813 | 813 | self.log.info("registration::unregister_engine(%s)"%eid) |
|
814 | 814 | # print (eid) |
|
815 | 815 | content=dict(id=eid, queue=self.engines[eid].queue) |
|
816 | 816 | self.ids.remove(eid) |
|
817 | 817 | uuid = self.keytable.pop(eid) |
|
818 | 818 | ec = self.engines.pop(eid) |
|
819 | 819 | self.hearts.pop(ec.heartbeat) |
|
820 | 820 | self.by_ident.pop(ec.queue) |
|
821 | 821 | self.completed.pop(eid) |
|
822 | 822 | self._handle_stranded_msgs(eid, uuid) |
|
823 | 823 | ############## TODO: HANDLE IT ################ |
|
824 | 824 | |
|
825 | 825 | if self.notifier: |
|
826 | 826 | self.session.send(self.notifier, "unregistration_notification", content=content) |
|
827 | 827 | |
|
828 | 828 | def _handle_stranded_msgs(self, eid, uuid): |
|
829 | 829 | """Handle messages known to be on an engine when the engine unregisters. |
|
830 | 830 | |
|
831 | 831 | It is possible that this will fire prematurely - that is, an engine will |
|
832 | 832 | go down after completing a result, and the client will be notified |
|
833 | 833 | that the result failed and later receive the actual result. |
|
834 | 834 | """ |
|
835 | 835 | |
|
836 | 836 | outstanding = self.queues.pop(eid) |
|
837 | 837 | |
|
838 | 838 | for msg_id in outstanding: |
|
839 | 839 | self.pending.remove(msg_id) |
|
840 | 840 | self.all_completed.add(msg_id) |
|
841 | 841 | try: |
|
842 | 842 | raise error.EngineError("Engine %r died while running task %r"%(eid, msg_id)) |
|
843 | 843 | except: |
|
844 | 844 | content = error.wrap_exception() |
|
845 | 845 | # build a fake header: |
|
846 | 846 | header = {} |
|
847 | 847 | header['engine'] = uuid |
|
848 | 848 | header['date'] = datetime.now().strftime(ISO8601) |
|
849 | 849 | rec = dict(result_content=content, result_header=header, result_buffers=[]) |
|
850 | 850 | rec['completed'] = header['date'] |
|
851 | 851 | rec['engine_uuid'] = uuid |
|
852 | 852 | self.db.update_record(msg_id, rec) |
|
853 | 853 | |
|
854 | 854 | def finish_registration(self, heart): |
|
855 | 855 | """Second half of engine registration, called after our HeartMonitor |
|
856 | 856 | has received a beat from the Engine's Heart.""" |
|
857 | 857 | try: |
|
858 | 858 | (eid,queue,reg,purge) = self.incoming_registrations.pop(heart) |
|
859 | 859 | except KeyError: |
|
860 | 860 | self.log.error("registration::tried to finish nonexistant registration", exc_info=True) |
|
861 | 861 | return |
|
862 | 862 | self.log.info("registration::finished registering engine %i:%r"%(eid,queue)) |
|
863 | 863 | if purge is not None: |
|
864 | 864 | purge.stop() |
|
865 | 865 | control = queue |
|
866 | 866 | self.ids.add(eid) |
|
867 | 867 | self.keytable[eid] = queue |
|
868 | 868 | self.engines[eid] = EngineConnector(id=eid, queue=queue, registration=reg, |
|
869 | 869 | control=control, heartbeat=heart) |
|
870 | 870 | self.by_ident[queue] = eid |
|
871 | 871 | self.queues[eid] = list() |
|
872 | 872 | self.tasks[eid] = list() |
|
873 | 873 | self.completed[eid] = list() |
|
874 | 874 | self.hearts[heart] = eid |
|
875 | 875 | content = dict(id=eid, queue=self.engines[eid].queue) |
|
876 | 876 | if self.notifier: |
|
877 | 877 | self.session.send(self.notifier, "registration_notification", content=content) |
|
878 | 878 | self.log.info("engine::Engine Connected: %i"%eid) |
|
879 | 879 | |
|
880 | 880 | def _purge_stalled_registration(self, heart): |
|
881 | 881 | if heart in self.incoming_registrations: |
|
882 | 882 | eid = self.incoming_registrations.pop(heart)[0] |
|
883 | 883 | self.log.info("registration::purging stalled registration: %i"%eid) |
|
884 | 884 | else: |
|
885 | 885 | pass |
|
886 | 886 | |
|
887 | 887 | #------------------------------------------------------------------------- |
|
888 | 888 | # Client Requests |
|
889 | 889 | #------------------------------------------------------------------------- |
|
890 | 890 | |
|
891 | 891 | def shutdown_request(self, client_id, msg): |
|
892 | 892 | """handle shutdown request.""" |
|
893 | # s = self.context.socket(zmq.XREQ) | |
|
894 | # s.connect(self.client_connections['mux']) | |
|
895 | # time.sleep(0.1) | |
|
896 | # for eid,ec in self.engines.iteritems(): | |
|
897 | # self.session.send(s, 'shutdown_request', content=dict(restart=False), ident=ec.queue) | |
|
898 | # time.sleep(1) | |
|
899 | 893 | self.session.send(self.query, 'shutdown_reply', content={'status': 'ok'}, ident=client_id) |
|
894 | # also notify other clients of shutdown | |
|
895 | self.session.send(self.notifier, 'shutdown_notice', content={'status': 'ok'}) | |
|
900 | 896 | dc = ioloop.DelayedCallback(lambda : self._shutdown(), 1000, self.loop) |
|
901 | 897 | dc.start() |
|
902 | 898 | |
|
903 | 899 | def _shutdown(self): |
|
904 | 900 | self.log.info("hub::hub shutting down.") |
|
905 | 901 | time.sleep(0.1) |
|
906 | 902 | sys.exit(0) |
|
907 | 903 | |
|
908 | 904 | |
|
909 | 905 | def check_load(self, client_id, msg): |
|
910 | 906 | content = msg['content'] |
|
911 | 907 | try: |
|
912 | 908 | targets = content['targets'] |
|
913 | 909 | targets = self._validate_targets(targets) |
|
914 | 910 | except: |
|
915 | 911 | content = error.wrap_exception() |
|
916 | 912 | self.session.send(self.query, "hub_error", |
|
917 | 913 | content=content, ident=client_id) |
|
918 | 914 | return |
|
919 | 915 | |
|
920 | 916 | content = dict(status='ok') |
|
921 | 917 | # loads = {} |
|
922 | 918 | for t in targets: |
|
923 | 919 | content[bytes(t)] = len(self.queues[t])+len(self.tasks[t]) |
|
924 | 920 | self.session.send(self.query, "load_reply", content=content, ident=client_id) |
|
925 | 921 | |
|
926 | 922 | |
|
927 | 923 | def queue_status(self, client_id, msg): |
|
928 | 924 | """Return the Queue status of one or more targets. |
|
929 | 925 | if verbose: return the msg_ids |
|
930 | 926 | else: return len of each type. |
|
931 | 927 | keys: queue (pending MUX jobs) |
|
932 | 928 | tasks (pending Task jobs) |
|
933 | 929 | completed (finished jobs from both queues)""" |
|
934 | 930 | content = msg['content'] |
|
935 | 931 | targets = content['targets'] |
|
936 | 932 | try: |
|
937 | 933 | targets = self._validate_targets(targets) |
|
938 | 934 | except: |
|
939 | 935 | content = error.wrap_exception() |
|
940 | 936 | self.session.send(self.query, "hub_error", |
|
941 | 937 | content=content, ident=client_id) |
|
942 | 938 | return |
|
943 | 939 | verbose = content.get('verbose', False) |
|
944 | 940 | content = dict(status='ok') |
|
945 | 941 | for t in targets: |
|
946 | 942 | queue = self.queues[t] |
|
947 | 943 | completed = self.completed[t] |
|
948 | 944 | tasks = self.tasks[t] |
|
949 | 945 | if not verbose: |
|
950 | 946 | queue = len(queue) |
|
951 | 947 | completed = len(completed) |
|
952 | 948 | tasks = len(tasks) |
|
953 | 949 | content[bytes(t)] = {'queue': queue, 'completed': completed , 'tasks': tasks} |
|
954 | 950 | # pending |
|
955 | 951 | self.session.send(self.query, "queue_reply", content=content, ident=client_id) |
|
956 | 952 | |
|
957 | 953 | def purge_results(self, client_id, msg): |
|
958 | 954 | """Purge results from memory. This method is more valuable before we move |
|
959 | 955 | to a DB based message storage mechanism.""" |
|
960 | 956 | content = msg['content'] |
|
961 | 957 | msg_ids = content.get('msg_ids', []) |
|
962 | 958 | reply = dict(status='ok') |
|
963 | 959 | if msg_ids == 'all': |
|
964 | 960 | self.db.drop_matching_records(dict(completed={'$ne':None})) |
|
965 | 961 | else: |
|
966 | 962 | for msg_id in msg_ids: |
|
967 | 963 | if msg_id in self.all_completed: |
|
968 | 964 | self.db.drop_record(msg_id) |
|
969 | 965 | else: |
|
970 | 966 | if msg_id in self.pending: |
|
971 | 967 | try: |
|
972 | 968 | raise IndexError("msg pending: %r"%msg_id) |
|
973 | 969 | except: |
|
974 | 970 | reply = error.wrap_exception() |
|
975 | 971 | else: |
|
976 | 972 | try: |
|
977 | 973 | raise IndexError("No such msg: %r"%msg_id) |
|
978 | 974 | except: |
|
979 | 975 | reply = error.wrap_exception() |
|
980 | 976 | break |
|
981 | 977 | eids = content.get('engine_ids', []) |
|
982 | 978 | for eid in eids: |
|
983 | 979 | if eid not in self.engines: |
|
984 | 980 | try: |
|
985 | 981 | raise IndexError("No such engine: %i"%eid) |
|
986 | 982 | except: |
|
987 | 983 | reply = error.wrap_exception() |
|
988 | 984 | break |
|
989 | 985 | msg_ids = self.completed.pop(eid) |
|
990 | 986 | uid = self.engines[eid].queue |
|
991 | 987 | self.db.drop_matching_records(dict(engine_uuid=uid, completed={'$ne':None})) |
|
992 | 988 | |
|
993 | 989 | self.session.send(self.query, 'purge_reply', content=reply, ident=client_id) |
|
994 | 990 | |
|
995 | 991 | def resubmit_task(self, client_id, msg, buffers): |
|
996 | 992 | """Resubmit a task.""" |
|
997 | 993 | raise NotImplementedError |
|
998 | 994 | |
|
999 | 995 | def get_results(self, client_id, msg): |
|
1000 | 996 | """Get the result of 1 or more messages.""" |
|
1001 | 997 | content = msg['content'] |
|
1002 | 998 | msg_ids = sorted(set(content['msg_ids'])) |
|
1003 | 999 | statusonly = content.get('status_only', False) |
|
1004 | 1000 | pending = [] |
|
1005 | 1001 | completed = [] |
|
1006 | 1002 | content = dict(status='ok') |
|
1007 | 1003 | content['pending'] = pending |
|
1008 | 1004 | content['completed'] = completed |
|
1009 | 1005 | buffers = [] |
|
1010 | 1006 | if not statusonly: |
|
1011 | 1007 | content['results'] = {} |
|
1012 | 1008 | records = self.db.find_records(dict(msg_id={'$in':msg_ids})) |
|
1013 | 1009 | for msg_id in msg_ids: |
|
1014 | 1010 | if msg_id in self.pending: |
|
1015 | 1011 | pending.append(msg_id) |
|
1016 | 1012 | elif msg_id in self.all_completed: |
|
1017 | 1013 | completed.append(msg_id) |
|
1018 | 1014 | if not statusonly: |
|
1019 | 1015 | rec = records[msg_id] |
|
1020 | 1016 | io_dict = {} |
|
1021 | 1017 | for key in 'pyin pyout pyerr stdout stderr'.split(): |
|
1022 | 1018 | io_dict[key] = rec[key] |
|
1023 | 1019 | content[msg_id] = { 'result_content': rec['result_content'], |
|
1024 | 1020 | 'header': rec['header'], |
|
1025 | 1021 | 'result_header' : rec['result_header'], |
|
1026 | 1022 | 'io' : io_dict, |
|
1027 | 1023 | } |
|
1028 | 1024 | if rec['result_buffers']: |
|
1029 | 1025 | buffers.extend(map(str, rec['result_buffers'])) |
|
1030 | 1026 | else: |
|
1031 | 1027 | try: |
|
1032 | 1028 | raise KeyError('No such message: '+msg_id) |
|
1033 | 1029 | except: |
|
1034 | 1030 | content = error.wrap_exception() |
|
1035 | 1031 | break |
|
1036 | 1032 | self.session.send(self.query, "result_reply", content=content, |
|
1037 | 1033 | parent=msg, ident=client_id, |
|
1038 | 1034 | buffers=buffers) |
|
1039 | 1035 |
@@ -1,203 +1,200 b'' | |||
|
1 |
"""Remote Functions and decorators for |
|
|
1 | """Remote Functions and decorators for Views.""" | |
|
2 | 2 | #----------------------------------------------------------------------------- |
|
3 | 3 | # Copyright (C) 2010 The IPython Development Team |
|
4 | 4 | # |
|
5 | 5 | # Distributed under the terms of the BSD License. The full license is in |
|
6 | 6 | # the file COPYING, distributed as part of this software. |
|
7 | 7 | #----------------------------------------------------------------------------- |
|
8 | 8 | |
|
9 | 9 | #----------------------------------------------------------------------------- |
|
10 | 10 | # Imports |
|
11 | 11 | #----------------------------------------------------------------------------- |
|
12 | 12 | |
|
13 | 13 | import warnings |
|
14 | 14 | |
|
15 | 15 | from IPython.testing import decorators as testdec |
|
16 | 16 | |
|
17 | 17 | from . import map as Map |
|
18 | 18 | from .asyncresult import AsyncMapResult |
|
19 | 19 | |
|
20 | 20 | #----------------------------------------------------------------------------- |
|
21 | 21 | # Decorators |
|
22 | 22 | #----------------------------------------------------------------------------- |
|
23 | 23 | |
|
24 | 24 | @testdec.skip_doctest |
|
25 | def remote(client, bound=False, block=None, targets=None, balanced=None): | |
|
25 | def remote(view, block=None, **flags): | |
|
26 | 26 | """Turn a function into a remote function. |
|
27 | 27 | |
|
28 | 28 | This method can be used for map: |
|
29 | 29 | |
|
30 |
In [1]: @remote( |
|
|
30 | In [1]: @remote(view,block=True) | |
|
31 | 31 | ...: def func(a): |
|
32 | 32 | ...: pass |
|
33 | 33 | """ |
|
34 | 34 | |
|
35 | 35 | def remote_function(f): |
|
36 |
return RemoteFunction( |
|
|
36 | return RemoteFunction(view, f, block=block, **flags) | |
|
37 | 37 | return remote_function |
|
38 | 38 | |
|
39 | 39 | @testdec.skip_doctest |
|
40 |
def parallel( |
|
|
40 | def parallel(view, dist='b', block=None, **flags): | |
|
41 | 41 | """Turn a function into a parallel remote function. |
|
42 | 42 | |
|
43 | 43 | This method can be used for map: |
|
44 | 44 | |
|
45 |
In [1]: @parallel( |
|
|
45 | In [1]: @parallel(view, block=True) | |
|
46 | 46 | ...: def func(a): |
|
47 | 47 | ...: pass |
|
48 | 48 | """ |
|
49 | 49 | |
|
50 | 50 | def parallel_function(f): |
|
51 |
return ParallelFunction( |
|
|
51 | return ParallelFunction(view, f, dist=dist, block=block, **flags) | |
|
52 | 52 | return parallel_function |
|
53 | 53 | |
|
54 | 54 | #-------------------------------------------------------------------------- |
|
55 | 55 | # Classes |
|
56 | 56 | #-------------------------------------------------------------------------- |
|
57 | 57 | |
|
58 | 58 | class RemoteFunction(object): |
|
59 | 59 | """Turn an existing function into a remote function. |
|
60 | 60 | |
|
61 | 61 | Parameters |
|
62 | 62 | ---------- |
|
63 | 63 | |
|
64 |
|
|
|
65 |
The |
|
|
64 | view : View instance | |
|
65 | The view to be used for execution | |
|
66 | 66 | f : callable |
|
67 | 67 | The function to be wrapped into a remote function |
|
68 | bound : bool [default: False] | |
|
69 | Whether the affect the remote namespace when called | |
|
70 | 68 | block : bool [default: None] |
|
71 | 69 | Whether to wait for results or not. The default behavior is |
|
72 |
to use the current `block` attribute of ` |
|
|
73 | targets : valid target list [default: all] | |
|
74 | The targets on which to execute. | |
|
75 | balanced : bool | |
|
76 | Whether to load-balance with the Task scheduler or not | |
|
70 | to use the current `block` attribute of `view` | |
|
71 | ||
|
72 | **flags : remaining kwargs are passed to View.temp_flags | |
|
77 | 73 | """ |
|
78 | 74 | |
|
79 |
|
|
|
75 | view = None # the remote connection | |
|
80 | 76 | func = None # the wrapped function |
|
81 | 77 | block = None # whether to block |
|
82 | bound = None # whether to affect the namespace | |
|
83 | targets = None # where to execute | |
|
84 | balanced = None # whether to load-balance | |
|
78 | flags = None # dict of extra kwargs for temp_flags | |
|
85 | 79 | |
|
86 |
def __init__(self, |
|
|
87 |
self. |
|
|
80 | def __init__(self, view, f, block=None, **flags): | |
|
81 | self.view = view | |
|
88 | 82 | self.func = f |
|
89 | 83 | self.block=block |
|
90 | self.bound=bound | |
|
91 | self.targets=targets | |
|
92 | if balanced is None: | |
|
93 | if targets is None: | |
|
94 | balanced = True | |
|
95 | else: | |
|
96 | balanced = False | |
|
97 | self.balanced = balanced | |
|
84 | self.flags=flags | |
|
98 | 85 | |
|
99 | 86 | def __call__(self, *args, **kwargs): |
|
100 | return self.client.apply(self.func, args=args, kwargs=kwargs, | |
|
101 | block=self.block, targets=self.targets, bound=self.bound, balanced=self.balanced) | |
|
87 | block = self.view.block if self.block is None else self.block | |
|
88 | with self.view.temp_flags(block=block, **self.flags): | |
|
89 | return self.view.apply(self.func, *args, **kwargs) | |
|
102 | 90 | |
|
103 | 91 | |
|
104 | 92 | class ParallelFunction(RemoteFunction): |
|
105 | 93 | """Class for mapping a function to sequences. |
|
106 | 94 | |
|
107 | 95 | This will distribute the sequences according the a mapper, and call |
|
108 | 96 | the function on each sub-sequence. If called via map, then the function |
|
109 | 97 | will be called once on each element, rather that each sub-sequence. |
|
110 | 98 | |
|
111 | 99 | Parameters |
|
112 | 100 | ---------- |
|
113 | 101 | |
|
114 |
|
|
|
115 |
The |
|
|
102 | view : View instance | |
|
103 | The view to be used for execution | |
|
116 | 104 | f : callable |
|
117 | 105 | The function to be wrapped into a remote function |
|
118 |
|
|
|
119 | Whether the affect the remote namespace when called | |
|
106 | dist : str [default: 'b'] | |
|
107 | The key for which mapObject to use to distribute sequences | |
|
108 | options are: | |
|
109 | * 'b' : use contiguous chunks in order | |
|
110 | * 'r' : use round-robin striping | |
|
120 | 111 | block : bool [default: None] |
|
121 | 112 | Whether to wait for results or not. The default behavior is |
|
122 |
to use the current `block` attribute of ` |
|
|
123 | targets : valid target list [default: all] | |
|
124 | The targets on which to execute. | |
|
125 | balanced : bool | |
|
126 | Whether to load-balance with the Task scheduler or not | |
|
127 | chunk_size : int or None | |
|
113 | to use the current `block` attribute of `view` | |
|
114 | chunksize : int or None | |
|
128 | 115 | The size of chunk to use when breaking up sequences in a load-balanced manner |
|
116 | **flags : remaining kwargs are passed to View.temp_flags | |
|
129 | 117 | """ |
|
130 | def __init__(self, client, f, dist='b', bound=False, block=None, targets='all', balanced=None, chunk_size=None): | |
|
131 | super(ParallelFunction, self).__init__(client,f,bound,block,targets,balanced) | |
|
132 | self.chunk_size = chunk_size | |
|
118 | ||
|
119 | chunksize=None | |
|
120 | mapObject=None | |
|
121 | ||
|
122 | def __init__(self, view, f, dist='b', block=None, chunksize=None, **flags): | |
|
123 | super(ParallelFunction, self).__init__(view, f, block=block, **flags) | |
|
124 | self.chunksize = chunksize | |
|
133 | 125 | |
|
134 | 126 | mapClass = Map.dists[dist] |
|
135 | 127 | self.mapObject = mapClass() |
|
136 | 128 | |
|
137 | 129 | def __call__(self, *sequences): |
|
130 | # check that the length of sequences match | |
|
138 | 131 | len_0 = len(sequences[0]) |
|
139 | 132 | for s in sequences: |
|
140 | 133 | if len(s)!=len_0: |
|
141 | 134 | msg = 'all sequences must have equal length, but %i!=%i'%(len_0,len(s)) |
|
142 | 135 | raise ValueError(msg) |
|
143 | ||
|
144 |
if |
|
|
145 |
if self.chunk |
|
|
146 |
nparts = len_0/self.chunk |
|
|
136 | balanced = 'Balanced' in self.view.__class__.__name__ | |
|
137 | if balanced: | |
|
138 | if self.chunksize: | |
|
139 | nparts = len_0/self.chunksize + int(len_0%self.chunksize > 0) | |
|
147 | 140 | else: |
|
148 | 141 | nparts = len_0 |
|
149 |
targets = [ |
|
|
142 | targets = [None]*nparts | |
|
150 | 143 | else: |
|
151 |
if self.chunk |
|
|
152 |
warnings.warn("`chunk |
|
|
144 | if self.chunksize: | |
|
145 | warnings.warn("`chunksize` is ignored unless load balancing", UserWarning) | |
|
153 | 146 | # multiplexed: |
|
154 |
targets = self. |
|
|
147 | targets = self.view.targets | |
|
155 | 148 | nparts = len(targets) |
|
156 | 149 | |
|
157 | 150 | msg_ids = [] |
|
158 | 151 | # my_f = lambda *a: map(self.func, *a) |
|
152 | client = self.view.client | |
|
159 | 153 | for index, t in enumerate(targets): |
|
160 | 154 | args = [] |
|
161 | 155 | for seq in sequences: |
|
162 | 156 | part = self.mapObject.getPartition(seq, index, nparts) |
|
163 | 157 | if len(part) == 0: |
|
164 | 158 | continue |
|
165 | 159 | else: |
|
166 | 160 | args.append(part) |
|
167 | 161 | if not args: |
|
168 | 162 | continue |
|
169 | 163 | |
|
170 | 164 | # print (args) |
|
171 | 165 | if hasattr(self, '_map'): |
|
172 | 166 | f = map |
|
173 | 167 | args = [self.func]+args |
|
174 | 168 | else: |
|
175 | 169 | f=self.func |
|
176 | ar = self.client.apply(f, args=args, block=False, bound=self.bound, | |
|
177 | targets=t, balanced=self.balanced) | |
|
170 | ||
|
171 | view = self.view if balanced else client[t] | |
|
172 | with view.temp_flags(block=False, **self.flags): | |
|
173 | ar = view.apply(f, *args) | |
|
178 | 174 | |
|
179 | 175 | msg_ids.append(ar.msg_ids[0]) |
|
180 | 176 | |
|
181 | r = AsyncMapResult(self.client, msg_ids, self.mapObject, fname=self.func.__name__) | |
|
177 | r = AsyncMapResult(self.view.client, msg_ids, self.mapObject, fname=self.func.__name__) | |
|
178 | ||
|
182 | 179 | if self.block: |
|
183 | 180 | try: |
|
184 | 181 | return r.get() |
|
185 | 182 | except KeyboardInterrupt: |
|
186 | 183 | return r |
|
187 | 184 | else: |
|
188 | 185 | return r |
|
189 | 186 | |
|
190 | 187 | def map(self, *sequences): |
|
191 | 188 | """call a function on each element of a sequence remotely. |
|
192 | 189 | This should behave very much like the builtin map, but return an AsyncMapResult |
|
193 | 190 | if self.block is False. |
|
194 | 191 | """ |
|
195 | 192 | # set _map as a flag for use inside self.__call__ |
|
196 | 193 | self._map = True |
|
197 | 194 | try: |
|
198 | 195 | ret = self.__call__(*sequences) |
|
199 | 196 | finally: |
|
200 | 197 | del self._map |
|
201 | 198 | return ret |
|
202 | 199 | |
|
203 | 200 | __all__ = ['remote', 'parallel', 'RemoteFunction', 'ParallelFunction'] No newline at end of file |
@@ -1,590 +1,592 b'' | |||
|
1 | 1 | """The Python scheduler for rich scheduling. |
|
2 | 2 | |
|
3 | 3 | The Pure ZMQ scheduler does not allow routing schemes other than LRU, |
|
4 | 4 | nor does it check msg_id DAG dependencies. For those, a slightly slower |
|
5 | 5 | Python Scheduler exists. |
|
6 | 6 | """ |
|
7 | 7 | #----------------------------------------------------------------------------- |
|
8 | 8 | # Copyright (C) 2010-2011 The IPython Development Team |
|
9 | 9 | # |
|
10 | 10 | # Distributed under the terms of the BSD License. The full license is in |
|
11 | 11 | # the file COPYING, distributed as part of this software. |
|
12 | 12 | #----------------------------------------------------------------------------- |
|
13 | 13 | |
|
14 | 14 | #---------------------------------------------------------------------- |
|
15 | 15 | # Imports |
|
16 | 16 | #---------------------------------------------------------------------- |
|
17 | 17 | |
|
18 | 18 | from __future__ import print_function |
|
19 | 19 | |
|
20 | 20 | import logging |
|
21 | 21 | import sys |
|
22 | 22 | |
|
23 | 23 | from datetime import datetime, timedelta |
|
24 | 24 | from random import randint, random |
|
25 | 25 | from types import FunctionType |
|
26 | 26 | |
|
27 | 27 | try: |
|
28 | 28 | import numpy |
|
29 | 29 | except ImportError: |
|
30 | 30 | numpy = None |
|
31 | 31 | |
|
32 | 32 | import zmq |
|
33 | 33 | from zmq.eventloop import ioloop, zmqstream |
|
34 | 34 | |
|
35 | 35 | # local imports |
|
36 | 36 | from IPython.external.decorator import decorator |
|
37 | 37 | from IPython.utils.traitlets import Instance, Dict, List, Set |
|
38 | 38 | |
|
39 | 39 | from . import error |
|
40 | 40 | from .dependency import Dependency |
|
41 | 41 | from .entry_point import connect_logger, local_logger |
|
42 | 42 | from .factory import SessionFactory |
|
43 | 43 | |
|
44 | 44 | |
|
45 | 45 | @decorator |
|
46 | 46 | def logged(f,self,*args,**kwargs): |
|
47 | 47 | # print ("#--------------------") |
|
48 | 48 | self.log.debug("scheduler::%s(*%s,**%s)"%(f.func_name, args, kwargs)) |
|
49 | 49 | # print ("#--") |
|
50 | 50 | return f(self,*args, **kwargs) |
|
51 | 51 | |
|
52 | 52 | #---------------------------------------------------------------------- |
|
53 | 53 | # Chooser functions |
|
54 | 54 | #---------------------------------------------------------------------- |
|
55 | 55 | |
|
56 | 56 | def plainrandom(loads): |
|
57 | 57 | """Plain random pick.""" |
|
58 | 58 | n = len(loads) |
|
59 | 59 | return randint(0,n-1) |
|
60 | 60 | |
|
61 | 61 | def lru(loads): |
|
62 | 62 | """Always pick the front of the line. |
|
63 | 63 | |
|
64 | 64 | The content of `loads` is ignored. |
|
65 | 65 | |
|
66 | 66 | Assumes LRU ordering of loads, with oldest first. |
|
67 | 67 | """ |
|
68 | 68 | return 0 |
|
69 | 69 | |
|
70 | 70 | def twobin(loads): |
|
71 | 71 | """Pick two at random, use the LRU of the two. |
|
72 | 72 | |
|
73 | 73 | The content of loads is ignored. |
|
74 | 74 | |
|
75 | 75 | Assumes LRU ordering of loads, with oldest first. |
|
76 | 76 | """ |
|
77 | 77 | n = len(loads) |
|
78 | 78 | a = randint(0,n-1) |
|
79 | 79 | b = randint(0,n-1) |
|
80 | 80 | return min(a,b) |
|
81 | 81 | |
|
82 | 82 | def weighted(loads): |
|
83 | 83 | """Pick two at random using inverse load as weight. |
|
84 | 84 | |
|
85 | 85 | Return the less loaded of the two. |
|
86 | 86 | """ |
|
87 | 87 | # weight 0 a million times more than 1: |
|
88 | 88 | weights = 1./(1e-6+numpy.array(loads)) |
|
89 | 89 | sums = weights.cumsum() |
|
90 | 90 | t = sums[-1] |
|
91 | 91 | x = random()*t |
|
92 | 92 | y = random()*t |
|
93 | 93 | idx = 0 |
|
94 | 94 | idy = 0 |
|
95 | 95 | while sums[idx] < x: |
|
96 | 96 | idx += 1 |
|
97 | 97 | while sums[idy] < y: |
|
98 | 98 | idy += 1 |
|
99 | 99 | if weights[idy] > weights[idx]: |
|
100 | 100 | return idy |
|
101 | 101 | else: |
|
102 | 102 | return idx |
|
103 | 103 | |
|
104 | 104 | def leastload(loads): |
|
105 | 105 | """Always choose the lowest load. |
|
106 | 106 | |
|
107 | 107 | If the lowest load occurs more than once, the first |
|
108 | 108 | occurance will be used. If loads has LRU ordering, this means |
|
109 | 109 | the LRU of those with the lowest load is chosen. |
|
110 | 110 | """ |
|
111 | 111 | return loads.index(min(loads)) |
|
112 | 112 | |
|
113 | 113 | #--------------------------------------------------------------------- |
|
114 | 114 | # Classes |
|
115 | 115 | #--------------------------------------------------------------------- |
|
116 | 116 | # store empty default dependency: |
|
117 | 117 | MET = Dependency([]) |
|
118 | 118 | |
|
119 | 119 | class TaskScheduler(SessionFactory): |
|
120 | 120 | """Python TaskScheduler object. |
|
121 | 121 | |
|
122 | 122 | This is the simplest object that supports msg_id based |
|
123 | 123 | DAG dependencies. *Only* task msg_ids are checked, not |
|
124 | 124 | msg_ids of jobs submitted via the MUX queue. |
|
125 | 125 | |
|
126 | 126 | """ |
|
127 | 127 | |
|
128 | 128 | # input arguments: |
|
129 | 129 | scheme = Instance(FunctionType, default=leastload) # function for determining the destination |
|
130 | 130 | client_stream = Instance(zmqstream.ZMQStream) # client-facing stream |
|
131 | 131 | engine_stream = Instance(zmqstream.ZMQStream) # engine-facing stream |
|
132 | 132 | notifier_stream = Instance(zmqstream.ZMQStream) # hub-facing sub stream |
|
133 | 133 | mon_stream = Instance(zmqstream.ZMQStream) # hub-facing pub stream |
|
134 | 134 | |
|
135 | 135 | # internals: |
|
136 | 136 | graph = Dict() # dict by msg_id of [ msg_ids that depend on key ] |
|
137 | 137 | depending = Dict() # dict by msg_id of (msg_id, raw_msg, after, follow) |
|
138 | 138 | pending = Dict() # dict by engine_uuid of submitted tasks |
|
139 | 139 | completed = Dict() # dict by engine_uuid of completed tasks |
|
140 | 140 | failed = Dict() # dict by engine_uuid of failed tasks |
|
141 | 141 | destinations = Dict() # dict by msg_id of engine_uuids where jobs ran (reverse of completed+failed) |
|
142 | 142 | clients = Dict() # dict by msg_id for who submitted the task |
|
143 | 143 | targets = List() # list of target IDENTs |
|
144 | 144 | loads = List() # list of engine loads |
|
145 | 145 | all_completed = Set() # set of all completed tasks |
|
146 | 146 | all_failed = Set() # set of all failed tasks |
|
147 | 147 | all_done = Set() # set of all finished tasks=union(completed,failed) |
|
148 | 148 | all_ids = Set() # set of all submitted task IDs |
|
149 | 149 | blacklist = Dict() # dict by msg_id of locations where a job has encountered UnmetDependency |
|
150 | 150 | auditor = Instance('zmq.eventloop.ioloop.PeriodicCallback') |
|
151 | 151 | |
|
152 | 152 | |
|
153 | 153 | def start(self): |
|
154 | 154 | self.engine_stream.on_recv(self.dispatch_result, copy=False) |
|
155 | 155 | self._notification_handlers = dict( |
|
156 | 156 | registration_notification = self._register_engine, |
|
157 | 157 | unregistration_notification = self._unregister_engine |
|
158 | 158 | ) |
|
159 | 159 | self.notifier_stream.on_recv(self.dispatch_notification) |
|
160 | 160 | self.auditor = ioloop.PeriodicCallback(self.audit_timeouts, 2e3, self.loop) # 1 Hz |
|
161 | 161 | self.auditor.start() |
|
162 | 162 | self.log.info("Scheduler started...%r"%self) |
|
163 | 163 | |
|
164 | 164 | def resume_receiving(self): |
|
165 | 165 | """Resume accepting jobs.""" |
|
166 | 166 | self.client_stream.on_recv(self.dispatch_submission, copy=False) |
|
167 | 167 | |
|
168 | 168 | def stop_receiving(self): |
|
169 | 169 | """Stop accepting jobs while there are no engines. |
|
170 | 170 | Leave them in the ZMQ queue.""" |
|
171 | 171 | self.client_stream.on_recv(None) |
|
172 | 172 | |
|
173 | 173 | #----------------------------------------------------------------------- |
|
174 | 174 | # [Un]Registration Handling |
|
175 | 175 | #----------------------------------------------------------------------- |
|
176 | 176 | |
|
177 | 177 | def dispatch_notification(self, msg): |
|
178 | 178 | """dispatch register/unregister events.""" |
|
179 | 179 | idents,msg = self.session.feed_identities(msg) |
|
180 | 180 | msg = self.session.unpack_message(msg) |
|
181 | 181 | msg_type = msg['msg_type'] |
|
182 | 182 | handler = self._notification_handlers.get(msg_type, None) |
|
183 | 183 | if handler is None: |
|
184 | 184 | raise Exception("Unhandled message type: %s"%msg_type) |
|
185 | 185 | else: |
|
186 | 186 | try: |
|
187 | 187 | handler(str(msg['content']['queue'])) |
|
188 | 188 | except KeyError: |
|
189 | 189 | self.log.error("task::Invalid notification msg: %s"%msg) |
|
190 | 190 | |
|
191 | 191 | @logged |
|
192 | 192 | def _register_engine(self, uid): |
|
193 | 193 | """New engine with ident `uid` became available.""" |
|
194 | 194 | # head of the line: |
|
195 | 195 | self.targets.insert(0,uid) |
|
196 | 196 | self.loads.insert(0,0) |
|
197 | 197 | # initialize sets |
|
198 | 198 | self.completed[uid] = set() |
|
199 | 199 | self.failed[uid] = set() |
|
200 | 200 | self.pending[uid] = {} |
|
201 | 201 | if len(self.targets) == 1: |
|
202 | 202 | self.resume_receiving() |
|
203 | 203 | |
|
204 | 204 | def _unregister_engine(self, uid): |
|
205 | 205 | """Existing engine with ident `uid` became unavailable.""" |
|
206 | 206 | if len(self.targets) == 1: |
|
207 | 207 | # this was our only engine |
|
208 | 208 | self.stop_receiving() |
|
209 | 209 | |
|
210 | 210 | # handle any potentially finished tasks: |
|
211 | 211 | self.engine_stream.flush() |
|
212 | 212 | |
|
213 | 213 | self.completed.pop(uid) |
|
214 | 214 | self.failed.pop(uid) |
|
215 | 215 | # don't pop destinations, because it might be used later |
|
216 | 216 | # map(self.destinations.pop, self.completed.pop(uid)) |
|
217 | 217 | # map(self.destinations.pop, self.failed.pop(uid)) |
|
218 | 218 | |
|
219 | 219 | idx = self.targets.index(uid) |
|
220 | 220 | self.targets.pop(idx) |
|
221 | 221 | self.loads.pop(idx) |
|
222 | 222 | |
|
223 | 223 | # wait 5 seconds before cleaning up pending jobs, since the results might |
|
224 | 224 | # still be incoming |
|
225 | 225 | if self.pending[uid]: |
|
226 | 226 | dc = ioloop.DelayedCallback(lambda : self.handle_stranded_tasks(uid), 5000, self.loop) |
|
227 | 227 | dc.start() |
|
228 | 228 | |
|
229 | 229 | @logged |
|
230 | 230 | def handle_stranded_tasks(self, engine): |
|
231 | 231 | """Deal with jobs resident in an engine that died.""" |
|
232 | 232 | lost = self.pending.pop(engine) |
|
233 | 233 | |
|
234 | 234 | for msg_id, (raw_msg, targets, MET, follow, timeout) in lost.iteritems(): |
|
235 | 235 | self.all_failed.add(msg_id) |
|
236 | 236 | self.all_done.add(msg_id) |
|
237 | 237 | idents,msg = self.session.feed_identities(raw_msg, copy=False) |
|
238 | 238 | msg = self.session.unpack_message(msg, copy=False, content=False) |
|
239 | 239 | parent = msg['header'] |
|
240 | 240 | idents = [idents[0],engine]+idents[1:] |
|
241 | print (idents) | |
|
241 | # print (idents) | |
|
242 | 242 | try: |
|
243 | 243 | raise error.EngineError("Engine %r died while running task %r"%(engine, msg_id)) |
|
244 | 244 | except: |
|
245 | 245 | content = error.wrap_exception() |
|
246 | 246 | msg = self.session.send(self.client_stream, 'apply_reply', content, |
|
247 | 247 | parent=parent, ident=idents) |
|
248 | 248 | self.session.send(self.mon_stream, msg, ident=['outtask']+idents) |
|
249 | 249 | self.update_graph(msg_id) |
|
250 | 250 | |
|
251 | 251 | |
|
252 | 252 | #----------------------------------------------------------------------- |
|
253 | 253 | # Job Submission |
|
254 | 254 | #----------------------------------------------------------------------- |
|
255 | 255 | @logged |
|
256 | 256 | def dispatch_submission(self, raw_msg): |
|
257 | 257 | """Dispatch job submission to appropriate handlers.""" |
|
258 | 258 | # ensure targets up to date: |
|
259 | 259 | self.notifier_stream.flush() |
|
260 | 260 | try: |
|
261 | 261 | idents, msg = self.session.feed_identities(raw_msg, copy=False) |
|
262 | 262 | msg = self.session.unpack_message(msg, content=False, copy=False) |
|
263 | 263 | except: |
|
264 | 264 | self.log.error("task::Invaid task: %s"%raw_msg, exc_info=True) |
|
265 | 265 | return |
|
266 | 266 | |
|
267 | 267 | # send to monitor |
|
268 | 268 | self.mon_stream.send_multipart(['intask']+raw_msg, copy=False) |
|
269 | 269 | |
|
270 | 270 | header = msg['header'] |
|
271 | 271 | msg_id = header['msg_id'] |
|
272 | 272 | self.all_ids.add(msg_id) |
|
273 | 273 | |
|
274 | 274 | # targets |
|
275 | 275 | targets = set(header.get('targets', [])) |
|
276 | 276 | |
|
277 | 277 | # time dependencies |
|
278 | 278 | after = Dependency(header.get('after', [])) |
|
279 | 279 | if after.all: |
|
280 | if after.success: | |
|
280 | 281 | after.difference_update(self.all_completed) |
|
281 |
if |
|
|
282 | if after.failure: | |
|
282 | 283 | after.difference_update(self.all_failed) |
|
283 | 284 | if after.check(self.all_completed, self.all_failed): |
|
284 | 285 | # recast as empty set, if `after` already met, |
|
285 | 286 | # to prevent unnecessary set comparisons |
|
286 | 287 | after = MET |
|
287 | 288 | |
|
288 | 289 | # location dependencies |
|
289 | 290 | follow = Dependency(header.get('follow', [])) |
|
290 | 291 | |
|
291 | 292 | # turn timeouts into datetime objects: |
|
292 | 293 | timeout = header.get('timeout', None) |
|
293 | 294 | if timeout: |
|
294 | 295 | timeout = datetime.now() + timedelta(0,timeout,0) |
|
295 | 296 | |
|
296 | 297 | args = [raw_msg, targets, after, follow, timeout] |
|
297 | 298 | |
|
298 | 299 | # validate and reduce dependencies: |
|
299 | 300 | for dep in after,follow: |
|
300 | 301 | # check valid: |
|
301 | 302 | if msg_id in dep or dep.difference(self.all_ids): |
|
302 | 303 | self.depending[msg_id] = args |
|
303 | 304 | return self.fail_unreachable(msg_id, error.InvalidDependency) |
|
304 | 305 | # check if unreachable: |
|
305 | if dep.unreachable(self.all_failed): | |
|
306 | if dep.unreachable(self.all_completed, self.all_failed): | |
|
306 | 307 | self.depending[msg_id] = args |
|
307 | 308 | return self.fail_unreachable(msg_id) |
|
308 | 309 | |
|
309 | 310 | if after.check(self.all_completed, self.all_failed): |
|
310 | 311 | # time deps already met, try to run |
|
311 | 312 | if not self.maybe_run(msg_id, *args): |
|
312 | 313 | # can't run yet |
|
313 | 314 | self.save_unmet(msg_id, *args) |
|
314 | 315 | else: |
|
315 | 316 | self.save_unmet(msg_id, *args) |
|
316 | 317 | |
|
317 | 318 | # @logged |
|
318 | 319 | def audit_timeouts(self): |
|
319 | 320 | """Audit all waiting tasks for expired timeouts.""" |
|
320 | 321 | now = datetime.now() |
|
321 | 322 | for msg_id in self.depending.keys(): |
|
322 | 323 | # must recheck, in case one failure cascaded to another: |
|
323 | 324 | if msg_id in self.depending: |
|
324 | 325 | raw,after,targets,follow,timeout = self.depending[msg_id] |
|
325 | 326 | if timeout and timeout < now: |
|
326 | 327 | self.fail_unreachable(msg_id, timeout=True) |
|
327 | 328 | |
|
328 | 329 | @logged |
|
329 | 330 | def fail_unreachable(self, msg_id, why=error.ImpossibleDependency): |
|
330 | 331 | """a task has become unreachable, send a reply with an ImpossibleDependency |
|
331 | 332 | error.""" |
|
332 | 333 | if msg_id not in self.depending: |
|
333 | 334 | self.log.error("msg %r already failed!"%msg_id) |
|
334 | 335 | return |
|
335 | 336 | raw_msg,targets,after,follow,timeout = self.depending.pop(msg_id) |
|
336 | 337 | for mid in follow.union(after): |
|
337 | 338 | if mid in self.graph: |
|
338 | 339 | self.graph[mid].remove(msg_id) |
|
339 | 340 | |
|
340 | 341 | # FIXME: unpacking a message I've already unpacked, but didn't save: |
|
341 | 342 | idents,msg = self.session.feed_identities(raw_msg, copy=False) |
|
342 | 343 | msg = self.session.unpack_message(msg, copy=False, content=False) |
|
343 | 344 | header = msg['header'] |
|
344 | 345 | |
|
345 | 346 | try: |
|
346 | 347 | raise why() |
|
347 | 348 | except: |
|
348 | 349 | content = error.wrap_exception() |
|
349 | 350 | |
|
350 | 351 | self.all_done.add(msg_id) |
|
351 | 352 | self.all_failed.add(msg_id) |
|
352 | 353 | |
|
353 | 354 | msg = self.session.send(self.client_stream, 'apply_reply', content, |
|
354 | 355 | parent=header, ident=idents) |
|
355 | 356 | self.session.send(self.mon_stream, msg, ident=['outtask']+idents) |
|
356 | 357 | |
|
357 | 358 | self.update_graph(msg_id, success=False) |
|
358 | 359 | |
|
359 | 360 | @logged |
|
360 | 361 | def maybe_run(self, msg_id, raw_msg, targets, after, follow, timeout): |
|
361 | 362 | """check location dependencies, and run if they are met.""" |
|
362 | 363 | blacklist = self.blacklist.setdefault(msg_id, set()) |
|
363 | 364 | if follow or targets or blacklist: |
|
364 | 365 | # we need a can_run filter |
|
365 | 366 | def can_run(idx): |
|
366 | 367 | target = self.targets[idx] |
|
367 | 368 | # check targets |
|
368 | 369 | if targets and target not in targets: |
|
369 | 370 | return False |
|
370 | 371 | # check blacklist |
|
371 | 372 | if target in blacklist: |
|
372 | 373 | return False |
|
373 | 374 | # check follow |
|
374 | 375 | return follow.check(self.completed[target], self.failed[target]) |
|
375 | 376 | |
|
376 | 377 | indices = filter(can_run, range(len(self.targets))) |
|
377 | 378 | if not indices: |
|
378 | 379 | # couldn't run |
|
379 | 380 | if follow.all: |
|
380 | 381 | # check follow for impossibility |
|
381 | 382 | dests = set() |
|
382 | relevant = self.all_completed if follow.success_only else self.all_done | |
|
383 | relevant = set() | |
|
384 | if follow.success: | |
|
385 | relevant = self.all_completed | |
|
386 | if follow.failure: | |
|
387 | relevant = relevant.union(self.all_failed) | |
|
383 | 388 | for m in follow.intersection(relevant): |
|
384 | 389 | dests.add(self.destinations[m]) |
|
385 | 390 | if len(dests) > 1: |
|
386 | 391 | self.fail_unreachable(msg_id) |
|
387 | 392 | return False |
|
388 | 393 | if targets: |
|
389 | 394 | # check blacklist+targets for impossibility |
|
390 | 395 | targets.difference_update(blacklist) |
|
391 | 396 | if not targets or not targets.intersection(self.targets): |
|
392 | 397 | self.fail_unreachable(msg_id) |
|
393 | 398 | return False |
|
394 | 399 | return False |
|
395 | 400 | else: |
|
396 | 401 | indices = None |
|
397 | 402 | |
|
398 | 403 | self.submit_task(msg_id, raw_msg, targets, follow, timeout, indices) |
|
399 | 404 | return True |
|
400 | 405 | |
|
401 | 406 | @logged |
|
402 | 407 | def save_unmet(self, msg_id, raw_msg, targets, after, follow, timeout): |
|
403 | 408 | """Save a message for later submission when its dependencies are met.""" |
|
404 | 409 | self.depending[msg_id] = [raw_msg,targets,after,follow,timeout] |
|
405 | 410 | # track the ids in follow or after, but not those already finished |
|
406 | 411 | for dep_id in after.union(follow).difference(self.all_done): |
|
407 | 412 | if dep_id not in self.graph: |
|
408 | 413 | self.graph[dep_id] = set() |
|
409 | 414 | self.graph[dep_id].add(msg_id) |
|
410 | 415 | |
|
411 | 416 | @logged |
|
412 | 417 | def submit_task(self, msg_id, raw_msg, targets, follow, timeout, indices=None): |
|
413 | 418 | """Submit a task to any of a subset of our targets.""" |
|
414 | 419 | if indices: |
|
415 | 420 | loads = [self.loads[i] for i in indices] |
|
416 | 421 | else: |
|
417 | 422 | loads = self.loads |
|
418 | 423 | idx = self.scheme(loads) |
|
419 | 424 | if indices: |
|
420 | 425 | idx = indices[idx] |
|
421 | 426 | target = self.targets[idx] |
|
422 | 427 | # print (target, map(str, msg[:3])) |
|
423 | 428 | self.engine_stream.send(target, flags=zmq.SNDMORE, copy=False) |
|
424 | 429 | self.engine_stream.send_multipart(raw_msg, copy=False) |
|
425 | 430 | self.add_job(idx) |
|
426 | 431 | self.pending[target][msg_id] = (raw_msg, targets, MET, follow, timeout) |
|
427 | 432 | content = dict(msg_id=msg_id, engine_id=target) |
|
428 | 433 | self.session.send(self.mon_stream, 'task_destination', content=content, |
|
429 | 434 | ident=['tracktask',self.session.session]) |
|
430 | 435 | |
|
431 | 436 | #----------------------------------------------------------------------- |
|
432 | 437 | # Result Handling |
|
433 | 438 | #----------------------------------------------------------------------- |
|
434 | 439 | @logged |
|
435 | 440 | def dispatch_result(self, raw_msg): |
|
436 | 441 | """dispatch method for result replies""" |
|
437 | 442 | try: |
|
438 | 443 | idents,msg = self.session.feed_identities(raw_msg, copy=False) |
|
439 | 444 | msg = self.session.unpack_message(msg, content=False, copy=False) |
|
440 | 445 | except: |
|
441 | 446 | self.log.error("task::Invaid result: %s"%raw_msg, exc_info=True) |
|
442 | 447 | return |
|
443 | 448 | |
|
444 | 449 | header = msg['header'] |
|
445 | 450 | if header.get('dependencies_met', True): |
|
446 | 451 | success = (header['status'] == 'ok') |
|
447 | 452 | self.handle_result(idents, msg['parent_header'], raw_msg, success) |
|
448 | 453 | # send to Hub monitor |
|
449 | 454 | self.mon_stream.send_multipart(['outtask']+raw_msg, copy=False) |
|
450 | 455 | else: |
|
451 | 456 | self.handle_unmet_dependency(idents, msg['parent_header']) |
|
452 | 457 | |
|
453 | 458 | @logged |
|
454 | 459 | def handle_result(self, idents, parent, raw_msg, success=True): |
|
455 | 460 | """handle a real task result, either success or failure""" |
|
456 | 461 | # first, relay result to client |
|
457 | 462 | engine = idents[0] |
|
458 | 463 | client = idents[1] |
|
459 | 464 | # swap_ids for XREP-XREP mirror |
|
460 | 465 | raw_msg[:2] = [client,engine] |
|
461 | 466 | # print (map(str, raw_msg[:4])) |
|
462 | 467 | self.client_stream.send_multipart(raw_msg, copy=False) |
|
463 | 468 | # now, update our data structures |
|
464 | 469 | msg_id = parent['msg_id'] |
|
465 | 470 | self.blacklist.pop(msg_id, None) |
|
466 | 471 | self.pending[engine].pop(msg_id) |
|
467 | 472 | if success: |
|
468 | 473 | self.completed[engine].add(msg_id) |
|
469 | 474 | self.all_completed.add(msg_id) |
|
470 | 475 | else: |
|
471 | 476 | self.failed[engine].add(msg_id) |
|
472 | 477 | self.all_failed.add(msg_id) |
|
473 | 478 | self.all_done.add(msg_id) |
|
474 | 479 | self.destinations[msg_id] = engine |
|
475 | 480 | |
|
476 | 481 | self.update_graph(msg_id, success) |
|
477 | 482 | |
|
478 | 483 | @logged |
|
479 | 484 | def handle_unmet_dependency(self, idents, parent): |
|
480 | 485 | """handle an unmet dependency""" |
|
481 | 486 | engine = idents[0] |
|
482 | 487 | msg_id = parent['msg_id'] |
|
483 | 488 | |
|
484 | 489 | if msg_id not in self.blacklist: |
|
485 | 490 | self.blacklist[msg_id] = set() |
|
486 | 491 | self.blacklist[msg_id].add(engine) |
|
487 | 492 | |
|
488 | 493 | args = self.pending[engine].pop(msg_id) |
|
489 | 494 | raw,targets,after,follow,timeout = args |
|
490 | 495 | |
|
491 | 496 | if self.blacklist[msg_id] == targets: |
|
492 | 497 | self.depending[msg_id] = args |
|
493 | 498 | return self.fail_unreachable(msg_id) |
|
494 | 499 | |
|
495 | 500 | elif not self.maybe_run(msg_id, *args): |
|
496 | 501 | # resubmit failed, put it back in our dependency tree |
|
497 | 502 | self.save_unmet(msg_id, *args) |
|
498 | 503 | |
|
499 | 504 | |
|
500 | 505 | @logged |
|
501 | 506 | def update_graph(self, dep_id, success=True): |
|
502 | 507 | """dep_id just finished. Update our dependency |
|
503 | 508 | graph and submit any jobs that just became runable.""" |
|
504 | 509 | # print ("\n\n***********") |
|
505 | 510 | # pprint (dep_id) |
|
506 | 511 | # pprint (self.graph) |
|
507 | 512 | # pprint (self.depending) |
|
508 | 513 | # pprint (self.all_completed) |
|
509 | 514 | # pprint (self.all_failed) |
|
510 | 515 | # print ("\n\n***********\n\n") |
|
511 | 516 | if dep_id not in self.graph: |
|
512 | 517 | return |
|
513 | 518 | jobs = self.graph.pop(dep_id) |
|
514 | 519 | |
|
515 | 520 | for msg_id in jobs: |
|
516 | 521 | raw_msg, targets, after, follow, timeout = self.depending[msg_id] |
|
517 | # if dep_id in after: | |
|
518 | # if after.all and (success or not after.success_only): | |
|
519 | # after.remove(dep_id) | |
|
520 | 522 | |
|
521 | if after.unreachable(self.all_failed) or follow.unreachable(self.all_failed): | |
|
523 | if after.unreachable(self.all_completed, self.all_failed) or follow.unreachable(self.all_completed, self.all_failed): | |
|
522 | 524 | self.fail_unreachable(msg_id) |
|
523 | 525 | |
|
524 | 526 | elif after.check(self.all_completed, self.all_failed): # time deps met, maybe run |
|
525 | 527 | if self.maybe_run(msg_id, raw_msg, targets, MET, follow, timeout): |
|
526 | 528 | |
|
527 | 529 | self.depending.pop(msg_id) |
|
528 | 530 | for mid in follow.union(after): |
|
529 | 531 | if mid in self.graph: |
|
530 | 532 | self.graph[mid].remove(msg_id) |
|
531 | 533 | |
|
532 | 534 | #---------------------------------------------------------------------- |
|
533 | 535 | # methods to be overridden by subclasses |
|
534 | 536 | #---------------------------------------------------------------------- |
|
535 | 537 | |
|
536 | 538 | def add_job(self, idx): |
|
537 | 539 | """Called after self.targets[idx] just got the job with header. |
|
538 | 540 | Override with subclasses. The default ordering is simple LRU. |
|
539 | 541 | The default loads are the number of outstanding jobs.""" |
|
540 | 542 | self.loads[idx] += 1 |
|
541 | 543 | for lis in (self.targets, self.loads): |
|
542 | 544 | lis.append(lis.pop(idx)) |
|
543 | 545 | |
|
544 | 546 | |
|
545 | 547 | def finish_job(self, idx): |
|
546 | 548 | """Called after self.targets[idx] just finished a job. |
|
547 | 549 | Override with subclasses.""" |
|
548 | 550 | self.loads[idx] -= 1 |
|
549 | 551 | |
|
550 | 552 | |
|
551 | 553 | |
|
552 | 554 | def launch_scheduler(in_addr, out_addr, mon_addr, not_addr, config=None,logname='ZMQ', |
|
553 | 555 | log_addr=None, loglevel=logging.DEBUG, scheme='lru', |
|
554 | 556 | identity=b'task'): |
|
555 | 557 | from zmq.eventloop import ioloop |
|
556 | 558 | from zmq.eventloop.zmqstream import ZMQStream |
|
557 | 559 | |
|
558 | 560 | ctx = zmq.Context() |
|
559 | 561 | loop = ioloop.IOLoop() |
|
560 | 562 | print (in_addr, out_addr, mon_addr, not_addr) |
|
561 | 563 | ins = ZMQStream(ctx.socket(zmq.XREP),loop) |
|
562 | 564 | ins.setsockopt(zmq.IDENTITY, identity) |
|
563 | 565 | ins.bind(in_addr) |
|
564 | 566 | |
|
565 | 567 | outs = ZMQStream(ctx.socket(zmq.XREP),loop) |
|
566 | 568 | outs.setsockopt(zmq.IDENTITY, identity) |
|
567 | 569 | outs.bind(out_addr) |
|
568 | 570 | mons = ZMQStream(ctx.socket(zmq.PUB),loop) |
|
569 | 571 | mons.connect(mon_addr) |
|
570 | 572 | nots = ZMQStream(ctx.socket(zmq.SUB),loop) |
|
571 | 573 | nots.setsockopt(zmq.SUBSCRIBE, '') |
|
572 | 574 | nots.connect(not_addr) |
|
573 | 575 | |
|
574 | 576 | scheme = globals().get(scheme, None) |
|
575 | 577 | # setup logging |
|
576 | 578 | if log_addr: |
|
577 | 579 | connect_logger(logname, ctx, log_addr, root="scheduler", loglevel=loglevel) |
|
578 | 580 | else: |
|
579 | 581 | local_logger(logname, loglevel) |
|
580 | 582 | |
|
581 | 583 | scheduler = TaskScheduler(client_stream=ins, engine_stream=outs, |
|
582 | 584 | mon_stream=mons, notifier_stream=nots, |
|
583 | 585 | scheme=scheme, loop=loop, logname=logname, |
|
584 | 586 | config=config) |
|
585 | 587 | scheduler.start() |
|
586 | 588 | try: |
|
587 | 589 | loop.start() |
|
588 | 590 | except KeyboardInterrupt: |
|
589 | 591 | print ("interrupted, exiting...", file=sys.__stderr__) |
|
590 | 592 |
@@ -1,493 +1,489 b'' | |||
|
1 | 1 | #!/usr/bin/env python |
|
2 | 2 | """ |
|
3 | 3 | Kernel adapted from kernel.py to use ZMQ Streams |
|
4 | 4 | """ |
|
5 | 5 | #----------------------------------------------------------------------------- |
|
6 | 6 | # Copyright (C) 2010-2011 The IPython Development Team |
|
7 | 7 | # |
|
8 | 8 | # Distributed under the terms of the BSD License. The full license is in |
|
9 | 9 | # the file COPYING, distributed as part of this software. |
|
10 | 10 | #----------------------------------------------------------------------------- |
|
11 | 11 | |
|
12 | 12 | #----------------------------------------------------------------------------- |
|
13 | 13 | # Imports |
|
14 | 14 | #----------------------------------------------------------------------------- |
|
15 | 15 | |
|
16 | 16 | # Standard library imports. |
|
17 | 17 | from __future__ import print_function |
|
18 | 18 | |
|
19 | 19 | import sys |
|
20 | 20 | import time |
|
21 | 21 | |
|
22 | 22 | from code import CommandCompiler |
|
23 | 23 | from datetime import datetime |
|
24 | 24 | from pprint import pprint |
|
25 | 25 | from signal import SIGTERM, SIGKILL |
|
26 | 26 | |
|
27 | 27 | # System library imports. |
|
28 | 28 | import zmq |
|
29 | 29 | from zmq.eventloop import ioloop, zmqstream |
|
30 | 30 | |
|
31 | 31 | # Local imports. |
|
32 | 32 | from IPython.core import ultratb |
|
33 | 33 | from IPython.utils.traitlets import Instance, List, Int, Dict, Set, Str |
|
34 | 34 | from IPython.zmq.completer import KernelCompleter |
|
35 | 35 | from IPython.zmq.iostream import OutStream |
|
36 | 36 | from IPython.zmq.displayhook import DisplayHook |
|
37 | 37 | |
|
38 | 38 | from . import heartmonitor |
|
39 | 39 | from .client import Client |
|
40 | 40 | from .error import wrap_exception |
|
41 | 41 | from .factory import SessionFactory |
|
42 | 42 | from .streamsession import StreamSession |
|
43 | 43 | from .util import serialize_object, unpack_apply_message, ISO8601, Namespace |
|
44 | 44 | |
|
45 | 45 | def printer(*args): |
|
46 | 46 | pprint(args, stream=sys.__stdout__) |
|
47 | 47 | |
|
48 | 48 | |
|
49 | 49 | class _Passer: |
|
50 | 50 | """Empty class that implements `send()` that does nothing.""" |
|
51 | 51 | def send(self, *args, **kwargs): |
|
52 | 52 | pass |
|
53 | 53 | send_multipart = send |
|
54 | 54 | |
|
55 | 55 | |
|
56 | 56 | #----------------------------------------------------------------------------- |
|
57 | 57 | # Main kernel class |
|
58 | 58 | #----------------------------------------------------------------------------- |
|
59 | 59 | |
|
60 | 60 | class Kernel(SessionFactory): |
|
61 | 61 | |
|
62 | 62 | #--------------------------------------------------------------------------- |
|
63 | 63 | # Kernel interface |
|
64 | 64 | #--------------------------------------------------------------------------- |
|
65 | 65 | |
|
66 | 66 | # kwargs: |
|
67 | 67 | int_id = Int(-1, config=True) |
|
68 | 68 | user_ns = Dict(config=True) |
|
69 | 69 | exec_lines = List(config=True) |
|
70 | 70 | |
|
71 | 71 | control_stream = Instance(zmqstream.ZMQStream) |
|
72 | 72 | task_stream = Instance(zmqstream.ZMQStream) |
|
73 | 73 | iopub_stream = Instance(zmqstream.ZMQStream) |
|
74 | 74 | client = Instance('IPython.zmq.parallel.client.Client') |
|
75 | 75 | |
|
76 | 76 | # internals |
|
77 | 77 | shell_streams = List() |
|
78 | 78 | compiler = Instance(CommandCompiler, (), {}) |
|
79 | 79 | completer = Instance(KernelCompleter) |
|
80 | 80 | |
|
81 | 81 | aborted = Set() |
|
82 | 82 | shell_handlers = Dict() |
|
83 | 83 | control_handlers = Dict() |
|
84 | 84 | |
|
85 | 85 | def _set_prefix(self): |
|
86 | 86 | self.prefix = "engine.%s"%self.int_id |
|
87 | 87 | |
|
88 | 88 | def _connect_completer(self): |
|
89 | 89 | self.completer = KernelCompleter(self.user_ns) |
|
90 | 90 | |
|
91 | 91 | def __init__(self, **kwargs): |
|
92 | 92 | super(Kernel, self).__init__(**kwargs) |
|
93 | 93 | self._set_prefix() |
|
94 | 94 | self._connect_completer() |
|
95 | 95 | |
|
96 | 96 | self.on_trait_change(self._set_prefix, 'id') |
|
97 | 97 | self.on_trait_change(self._connect_completer, 'user_ns') |
|
98 | 98 | |
|
99 | 99 | # Build dict of handlers for message types |
|
100 | 100 | for msg_type in ['execute_request', 'complete_request', 'apply_request', |
|
101 | 101 | 'clear_request']: |
|
102 | 102 | self.shell_handlers[msg_type] = getattr(self, msg_type) |
|
103 | 103 | |
|
104 | 104 | for msg_type in ['shutdown_request', 'abort_request']+self.shell_handlers.keys(): |
|
105 | 105 | self.control_handlers[msg_type] = getattr(self, msg_type) |
|
106 | 106 | |
|
107 | 107 | self._initial_exec_lines() |
|
108 | 108 | |
|
109 | 109 | def _wrap_exception(self, method=None): |
|
110 | 110 | e_info = dict(engine_uuid=self.ident, engine_id=self.int_id, method=method) |
|
111 | 111 | content=wrap_exception(e_info) |
|
112 | 112 | return content |
|
113 | 113 | |
|
114 | 114 | def _initial_exec_lines(self): |
|
115 | 115 | s = _Passer() |
|
116 | 116 | content = dict(silent=True, user_variable=[],user_expressions=[]) |
|
117 | 117 | for line in self.exec_lines: |
|
118 | 118 | self.log.debug("executing initialization: %s"%line) |
|
119 | 119 | content.update({'code':line}) |
|
120 | 120 | msg = self.session.msg('execute_request', content) |
|
121 | 121 | self.execute_request(s, [], msg) |
|
122 | 122 | |
|
123 | 123 | |
|
124 | 124 | #-------------------- control handlers ----------------------------- |
|
125 | 125 | def abort_queues(self): |
|
126 | 126 | for stream in self.shell_streams: |
|
127 | 127 | if stream: |
|
128 | 128 | self.abort_queue(stream) |
|
129 | 129 | |
|
130 | 130 | def abort_queue(self, stream): |
|
131 | 131 | while True: |
|
132 | 132 | try: |
|
133 | 133 | msg = self.session.recv(stream, zmq.NOBLOCK,content=True) |
|
134 | 134 | except zmq.ZMQError as e: |
|
135 | 135 | if e.errno == zmq.EAGAIN: |
|
136 | 136 | break |
|
137 | 137 | else: |
|
138 | 138 | return |
|
139 | 139 | else: |
|
140 | 140 | if msg is None: |
|
141 | 141 | return |
|
142 | 142 | else: |
|
143 | 143 | idents,msg = msg |
|
144 | 144 | |
|
145 | 145 | # assert self.reply_socketly_socket.rcvmore(), "Unexpected missing message part." |
|
146 | 146 | # msg = self.reply_socket.recv_json() |
|
147 | 147 | self.log.info("Aborting:") |
|
148 | 148 | self.log.info(str(msg)) |
|
149 | 149 | msg_type = msg['msg_type'] |
|
150 | 150 | reply_type = msg_type.split('_')[0] + '_reply' |
|
151 | 151 | # reply_msg = self.session.msg(reply_type, {'status' : 'aborted'}, msg) |
|
152 | 152 | # self.reply_socket.send(ident,zmq.SNDMORE) |
|
153 | 153 | # self.reply_socket.send_json(reply_msg) |
|
154 | 154 | reply_msg = self.session.send(stream, reply_type, |
|
155 | 155 | content={'status' : 'aborted'}, parent=msg, ident=idents)[0] |
|
156 | 156 | self.log.debug(str(reply_msg)) |
|
157 | 157 | # We need to wait a bit for requests to come in. This can probably |
|
158 | 158 | # be set shorter for true asynchronous clients. |
|
159 | 159 | time.sleep(0.05) |
|
160 | 160 | |
|
161 | 161 | def abort_request(self, stream, ident, parent): |
|
162 | 162 | """abort a specifig msg by id""" |
|
163 | 163 | msg_ids = parent['content'].get('msg_ids', None) |
|
164 | 164 | if isinstance(msg_ids, basestring): |
|
165 | 165 | msg_ids = [msg_ids] |
|
166 | 166 | if not msg_ids: |
|
167 | 167 | self.abort_queues() |
|
168 | 168 | for mid in msg_ids: |
|
169 | 169 | self.aborted.add(str(mid)) |
|
170 | 170 | |
|
171 | 171 | content = dict(status='ok') |
|
172 | 172 | reply_msg = self.session.send(stream, 'abort_reply', content=content, |
|
173 |
parent=parent, ident=ident) |
|
|
173 | parent=parent, ident=ident) | |
|
174 | 174 | self.log.debug(str(reply_msg)) |
|
175 | 175 | |
|
176 | 176 | def shutdown_request(self, stream, ident, parent): |
|
177 | 177 | """kill ourself. This should really be handled in an external process""" |
|
178 | 178 | try: |
|
179 | 179 | self.abort_queues() |
|
180 | 180 | except: |
|
181 | 181 | content = self._wrap_exception('shutdown') |
|
182 | 182 | else: |
|
183 | 183 | content = dict(parent['content']) |
|
184 | 184 | content['status'] = 'ok' |
|
185 | 185 | msg = self.session.send(stream, 'shutdown_reply', |
|
186 | 186 | content=content, parent=parent, ident=ident) |
|
187 | # msg = self.session.send(self.pub_socket, 'shutdown_reply', | |
|
188 | # content, parent, ident) | |
|
189 | # print >> sys.__stdout__, msg | |
|
190 | # time.sleep(0.2) | |
|
187 | self.log.debug(str(msg)) | |
|
191 | 188 | dc = ioloop.DelayedCallback(lambda : sys.exit(0), 1000, self.loop) |
|
192 | 189 | dc.start() |
|
193 | 190 | |
|
194 | 191 | def dispatch_control(self, msg): |
|
195 | 192 | idents,msg = self.session.feed_identities(msg, copy=False) |
|
196 | 193 | try: |
|
197 | 194 | msg = self.session.unpack_message(msg, content=True, copy=False) |
|
198 | 195 | except: |
|
199 | 196 | self.log.error("Invalid Message", exc_info=True) |
|
200 | 197 | return |
|
201 | 198 | |
|
202 | 199 | header = msg['header'] |
|
203 | 200 | msg_id = header['msg_id'] |
|
204 | 201 | |
|
205 | 202 | handler = self.control_handlers.get(msg['msg_type'], None) |
|
206 | 203 | if handler is None: |
|
207 | 204 | self.log.error("UNKNOWN CONTROL MESSAGE TYPE: %r"%msg['msg_type']) |
|
208 | 205 | else: |
|
209 | 206 | handler(self.control_stream, idents, msg) |
|
210 | 207 | |
|
211 | 208 | |
|
212 | 209 | #-------------------- queue helpers ------------------------------ |
|
213 | 210 | |
|
214 | 211 | def check_dependencies(self, dependencies): |
|
215 | 212 | if not dependencies: |
|
216 | 213 | return True |
|
217 | 214 | if len(dependencies) == 2 and dependencies[0] in 'any all'.split(): |
|
218 | 215 | anyorall = dependencies[0] |
|
219 | 216 | dependencies = dependencies[1] |
|
220 | 217 | else: |
|
221 | 218 | anyorall = 'all' |
|
222 | 219 | results = self.client.get_results(dependencies,status_only=True) |
|
223 | 220 | if results['status'] != 'ok': |
|
224 | 221 | return False |
|
225 | 222 | |
|
226 | 223 | if anyorall == 'any': |
|
227 | 224 | if not results['completed']: |
|
228 | 225 | return False |
|
229 | 226 | else: |
|
230 | 227 | if results['pending']: |
|
231 | 228 | return False |
|
232 | 229 | |
|
233 | 230 | return True |
|
234 | 231 | |
|
235 | 232 | def check_aborted(self, msg_id): |
|
236 | 233 | return msg_id in self.aborted |
|
237 | 234 | |
|
238 | 235 | #-------------------- queue handlers ----------------------------- |
|
239 | 236 | |
|
240 | 237 | def clear_request(self, stream, idents, parent): |
|
241 | 238 | """Clear our namespace.""" |
|
242 | 239 | self.user_ns = {} |
|
243 | 240 | msg = self.session.send(stream, 'clear_reply', ident=idents, parent=parent, |
|
244 | 241 | content = dict(status='ok')) |
|
245 | 242 | self._initial_exec_lines() |
|
246 | 243 | |
|
247 | 244 | def execute_request(self, stream, ident, parent): |
|
248 | 245 | self.log.debug('execute request %s'%parent) |
|
249 | 246 | try: |
|
250 | 247 | code = parent[u'content'][u'code'] |
|
251 | 248 | except: |
|
252 | 249 | self.log.error("Got bad msg: %s"%parent, exc_info=True) |
|
253 | 250 | return |
|
254 | 251 | self.session.send(self.iopub_stream, u'pyin', {u'code':code},parent=parent, |
|
255 | 252 | ident='%s.pyin'%self.prefix) |
|
256 | 253 | started = datetime.now().strftime(ISO8601) |
|
257 | 254 | try: |
|
258 | 255 | comp_code = self.compiler(code, '<zmq-kernel>') |
|
259 | 256 | # allow for not overriding displayhook |
|
260 | 257 | if hasattr(sys.displayhook, 'set_parent'): |
|
261 | 258 | sys.displayhook.set_parent(parent) |
|
262 | 259 | sys.stdout.set_parent(parent) |
|
263 | 260 | sys.stderr.set_parent(parent) |
|
264 | 261 | exec comp_code in self.user_ns, self.user_ns |
|
265 | 262 | except: |
|
266 | 263 | exc_content = self._wrap_exception('execute') |
|
267 | 264 | # exc_msg = self.session.msg(u'pyerr', exc_content, parent) |
|
268 | 265 | self.session.send(self.iopub_stream, u'pyerr', exc_content, parent=parent, |
|
269 | 266 | ident='%s.pyerr'%self.prefix) |
|
270 | 267 | reply_content = exc_content |
|
271 | 268 | else: |
|
272 | 269 | reply_content = {'status' : 'ok'} |
|
273 | 270 | |
|
274 | 271 | reply_msg = self.session.send(stream, u'execute_reply', reply_content, parent=parent, |
|
275 | 272 | ident=ident, subheader = dict(started=started)) |
|
276 | 273 | self.log.debug(str(reply_msg)) |
|
277 | 274 | if reply_msg['content']['status'] == u'error': |
|
278 | 275 | self.abort_queues() |
|
279 | 276 | |
|
280 | 277 | def complete_request(self, stream, ident, parent): |
|
281 | 278 | matches = {'matches' : self.complete(parent), |
|
282 | 279 | 'status' : 'ok'} |
|
283 | 280 | completion_msg = self.session.send(stream, 'complete_reply', |
|
284 | 281 | matches, parent, ident) |
|
285 | 282 | # print >> sys.__stdout__, completion_msg |
|
286 | 283 | |
|
287 | 284 | def complete(self, msg): |
|
288 | 285 | return self.completer.complete(msg.content.line, msg.content.text) |
|
289 | 286 | |
|
290 | 287 | def apply_request(self, stream, ident, parent): |
|
291 | 288 | # flush previous reply, so this request won't block it |
|
292 | 289 | stream.flush(zmq.POLLOUT) |
|
293 | 290 | |
|
294 | 291 | try: |
|
295 | 292 | content = parent[u'content'] |
|
296 | 293 | bufs = parent[u'buffers'] |
|
297 | 294 | msg_id = parent['header']['msg_id'] |
|
298 |
bound = |
|
|
295 | # bound = parent['header'].get('bound', False) | |
|
299 | 296 | except: |
|
300 | 297 | self.log.error("Got bad msg: %s"%parent, exc_info=True) |
|
301 | 298 | return |
|
302 | 299 | # pyin_msg = self.session.msg(u'pyin',{u'code':code}, parent=parent) |
|
303 | 300 | # self.iopub_stream.send(pyin_msg) |
|
304 | 301 | # self.session.send(self.iopub_stream, u'pyin', {u'code':code},parent=parent) |
|
305 | 302 | sub = {'dependencies_met' : True, 'engine' : self.ident, |
|
306 | 303 | 'started': datetime.now().strftime(ISO8601)} |
|
307 | 304 | try: |
|
308 | 305 | # allow for not overriding displayhook |
|
309 | 306 | if hasattr(sys.displayhook, 'set_parent'): |
|
310 | 307 | sys.displayhook.set_parent(parent) |
|
311 | 308 | sys.stdout.set_parent(parent) |
|
312 | 309 | sys.stderr.set_parent(parent) |
|
313 | 310 | # exec "f(*args,**kwargs)" in self.user_ns, self.user_ns |
|
314 | 311 | working = self.user_ns |
|
315 | 312 | # suffix = |
|
316 | 313 | prefix = "_"+str(msg_id).replace("-","")+"_" |
|
317 |
|
|
|
318 | # | |
|
319 | # else: | |
|
320 | # working = dict() | |
|
321 | # suffix = prefix = "_" # prevent keyword collisions with lambda | |
|
314 | ||
|
322 | 315 | f,args,kwargs = unpack_apply_message(bufs, working, copy=False) |
|
323 | if bound: | |
|
324 | bound_ns = Namespace(working) | |
|
325 | args = [bound_ns]+list(args) | |
|
326 | # if f.fun | |
|
316 | # if bound: | |
|
317 | # bound_ns = Namespace(working) | |
|
318 | # args = [bound_ns]+list(args) | |
|
319 | ||
|
327 | 320 | fname = getattr(f, '__name__', 'f') |
|
328 | 321 | |
|
329 | 322 | fname = prefix+"f" |
|
330 | 323 | argname = prefix+"args" |
|
331 | 324 | kwargname = prefix+"kwargs" |
|
332 | 325 | resultname = prefix+"result" |
|
333 | 326 | |
|
334 | 327 | ns = { fname : f, argname : args, kwargname : kwargs , resultname : None } |
|
335 | 328 | # print ns |
|
336 | 329 | working.update(ns) |
|
337 | 330 | code = "%s=%s(*%s,**%s)"%(resultname, fname, argname, kwargname) |
|
338 | 331 | try: |
|
339 | 332 | exec code in working,working |
|
340 | 333 | result = working.get(resultname) |
|
341 | 334 | finally: |
|
342 | 335 | for key in ns.iterkeys(): |
|
343 | 336 | working.pop(key) |
|
344 | if bound: | |
|
345 | working.update(bound_ns) | |
|
337 | # if bound: | |
|
338 | # working.update(bound_ns) | |
|
346 | 339 | |
|
347 | 340 | packed_result,buf = serialize_object(result) |
|
348 | 341 | result_buf = [packed_result]+buf |
|
349 | 342 | except: |
|
350 | 343 | exc_content = self._wrap_exception('apply') |
|
351 | 344 | # exc_msg = self.session.msg(u'pyerr', exc_content, parent) |
|
352 | 345 | self.session.send(self.iopub_stream, u'pyerr', exc_content, parent=parent, |
|
353 | 346 | ident='%s.pyerr'%self.prefix) |
|
354 | 347 | reply_content = exc_content |
|
355 | 348 | result_buf = [] |
|
356 | 349 | |
|
357 | 350 | if exc_content['ename'] == 'UnmetDependency': |
|
358 | 351 | sub['dependencies_met'] = False |
|
359 | 352 | else: |
|
360 | 353 | reply_content = {'status' : 'ok'} |
|
361 | 354 | |
|
362 | 355 | # put 'ok'/'error' status in header, for scheduler introspection: |
|
363 | 356 | sub['status'] = reply_content['status'] |
|
364 | 357 | |
|
365 | 358 | reply_msg = self.session.send(stream, u'apply_reply', reply_content, |
|
366 | 359 | parent=parent, ident=ident,buffers=result_buf, subheader=sub) |
|
367 | 360 | |
|
368 | # if reply_msg['content']['status'] == u'error': | |
|
369 | # self.abort_queues() | |
|
361 | # flush i/o | |
|
362 | # should this be before reply_msg is sent, like in the single-kernel code, | |
|
363 | # or should nothing get in the way of real results? | |
|
364 | sys.stdout.flush() | |
|
365 | sys.stderr.flush() | |
|
370 | 366 | |
|
371 | 367 | def dispatch_queue(self, stream, msg): |
|
372 | 368 | self.control_stream.flush() |
|
373 | 369 | idents,msg = self.session.feed_identities(msg, copy=False) |
|
374 | 370 | try: |
|
375 | 371 | msg = self.session.unpack_message(msg, content=True, copy=False) |
|
376 | 372 | except: |
|
377 | 373 | self.log.error("Invalid Message", exc_info=True) |
|
378 | 374 | return |
|
379 | 375 | |
|
380 | 376 | |
|
381 | 377 | header = msg['header'] |
|
382 | 378 | msg_id = header['msg_id'] |
|
383 | 379 | if self.check_aborted(msg_id): |
|
384 | 380 | self.aborted.remove(msg_id) |
|
385 | 381 | # is it safe to assume a msg_id will not be resubmitted? |
|
386 | 382 | reply_type = msg['msg_type'].split('_')[0] + '_reply' |
|
387 | 383 | reply_msg = self.session.send(stream, reply_type, |
|
388 | 384 | content={'status' : 'aborted'}, parent=msg, ident=idents) |
|
389 | 385 | return |
|
390 | 386 | handler = self.shell_handlers.get(msg['msg_type'], None) |
|
391 | 387 | if handler is None: |
|
392 | 388 | self.log.error("UNKNOWN MESSAGE TYPE: %r"%msg['msg_type']) |
|
393 | 389 | else: |
|
394 | 390 | handler(stream, idents, msg) |
|
395 | 391 | |
|
396 | 392 | def start(self): |
|
397 | 393 | #### stream mode: |
|
398 | 394 | if self.control_stream: |
|
399 | 395 | self.control_stream.on_recv(self.dispatch_control, copy=False) |
|
400 | 396 | self.control_stream.on_err(printer) |
|
401 | 397 | |
|
402 | 398 | def make_dispatcher(stream): |
|
403 | 399 | def dispatcher(msg): |
|
404 | 400 | return self.dispatch_queue(stream, msg) |
|
405 | 401 | return dispatcher |
|
406 | 402 | |
|
407 | 403 | for s in self.shell_streams: |
|
408 | 404 | s.on_recv(make_dispatcher(s), copy=False) |
|
409 | 405 | s.on_err(printer) |
|
410 | 406 | |
|
411 | 407 | if self.iopub_stream: |
|
412 | 408 | self.iopub_stream.on_err(printer) |
|
413 | 409 | |
|
414 | 410 | #### while True mode: |
|
415 | 411 | # while True: |
|
416 | 412 | # idle = True |
|
417 | 413 | # try: |
|
418 | 414 | # msg = self.shell_stream.socket.recv_multipart( |
|
419 | 415 | # zmq.NOBLOCK, copy=False) |
|
420 | 416 | # except zmq.ZMQError, e: |
|
421 | 417 | # if e.errno != zmq.EAGAIN: |
|
422 | 418 | # raise e |
|
423 | 419 | # else: |
|
424 | 420 | # idle=False |
|
425 | 421 | # self.dispatch_queue(self.shell_stream, msg) |
|
426 | 422 | # |
|
427 | 423 | # if not self.task_stream.empty(): |
|
428 | 424 | # idle=False |
|
429 | 425 | # msg = self.task_stream.recv_multipart() |
|
430 | 426 | # self.dispatch_queue(self.task_stream, msg) |
|
431 | 427 | # if idle: |
|
432 | 428 | # # don't busywait |
|
433 | 429 | # time.sleep(1e-3) |
|
434 | 430 | |
|
435 | 431 | def make_kernel(int_id, identity, control_addr, shell_addrs, iopub_addr, hb_addrs, |
|
436 | 432 | client_addr=None, loop=None, context=None, key=None, |
|
437 | 433 | out_stream_factory=OutStream, display_hook_factory=DisplayHook): |
|
438 | 434 | """NO LONGER IN USE""" |
|
439 | 435 | # create loop, context, and session: |
|
440 | 436 | if loop is None: |
|
441 | 437 | loop = ioloop.IOLoop.instance() |
|
442 | 438 | if context is None: |
|
443 | 439 | context = zmq.Context() |
|
444 | 440 | c = context |
|
445 | 441 | session = StreamSession(key=key) |
|
446 | 442 | # print (session.key) |
|
447 | 443 | # print (control_addr, shell_addrs, iopub_addr, hb_addrs) |
|
448 | 444 | |
|
449 | 445 | # create Control Stream |
|
450 | 446 | control_stream = zmqstream.ZMQStream(c.socket(zmq.PAIR), loop) |
|
451 | 447 | control_stream.setsockopt(zmq.IDENTITY, identity) |
|
452 | 448 | control_stream.connect(control_addr) |
|
453 | 449 | |
|
454 | 450 | # create Shell Streams (MUX, Task, etc.): |
|
455 | 451 | shell_streams = [] |
|
456 | 452 | for addr in shell_addrs: |
|
457 | 453 | stream = zmqstream.ZMQStream(c.socket(zmq.PAIR), loop) |
|
458 | 454 | stream.setsockopt(zmq.IDENTITY, identity) |
|
459 | 455 | stream.connect(addr) |
|
460 | 456 | shell_streams.append(stream) |
|
461 | 457 | |
|
462 | 458 | # create iopub stream: |
|
463 | 459 | iopub_stream = zmqstream.ZMQStream(c.socket(zmq.PUB), loop) |
|
464 | 460 | iopub_stream.setsockopt(zmq.IDENTITY, identity) |
|
465 | 461 | iopub_stream.connect(iopub_addr) |
|
466 | 462 | |
|
467 | 463 | # Redirect input streams and set a display hook. |
|
468 | 464 | if out_stream_factory: |
|
469 | 465 | sys.stdout = out_stream_factory(session, iopub_stream, u'stdout') |
|
470 | 466 | sys.stdout.topic = 'engine.%i.stdout'%int_id |
|
471 | 467 | sys.stderr = out_stream_factory(session, iopub_stream, u'stderr') |
|
472 | 468 | sys.stderr.topic = 'engine.%i.stderr'%int_id |
|
473 | 469 | if display_hook_factory: |
|
474 | 470 | sys.displayhook = display_hook_factory(session, iopub_stream) |
|
475 | 471 | sys.displayhook.topic = 'engine.%i.pyout'%int_id |
|
476 | 472 | |
|
477 | 473 | |
|
478 | 474 | # launch heartbeat |
|
479 | 475 | heart = heartmonitor.Heart(*map(str, hb_addrs), heart_id=identity) |
|
480 | 476 | heart.start() |
|
481 | 477 | |
|
482 | 478 | # create (optional) Client |
|
483 | 479 | if client_addr: |
|
484 | 480 | client = Client(client_addr, username=identity) |
|
485 | 481 | else: |
|
486 | 482 | client = None |
|
487 | 483 | |
|
488 | 484 | kernel = Kernel(id=int_id, session=session, control_stream=control_stream, |
|
489 | 485 | shell_streams=shell_streams, iopub_stream=iopub_stream, |
|
490 | 486 | client=client, loop=loop) |
|
491 | 487 | kernel.start() |
|
492 | 488 | return loop, c, kernel |
|
493 | 489 |
@@ -1,48 +1,69 b'' | |||
|
1 | 1 | """toplevel setup/teardown for parallel tests.""" |
|
2 | 2 | |
|
3 | #------------------------------------------------------------------------------- | |
|
4 | # Copyright (C) 2011 The IPython Development Team | |
|
5 | # | |
|
6 | # Distributed under the terms of the BSD License. The full license is in | |
|
7 | # the file COPYING, distributed as part of this software. | |
|
8 | #------------------------------------------------------------------------------- | |
|
9 | ||
|
10 | #------------------------------------------------------------------------------- | |
|
11 | # Imports | |
|
12 | #------------------------------------------------------------------------------- | |
|
13 | ||
|
3 | 14 | import tempfile |
|
4 | 15 | import time |
|
5 | 16 | from subprocess import Popen, PIPE, STDOUT |
|
6 | 17 | |
|
7 | 18 | from IPython.zmq.parallel import client |
|
8 | 19 | |
|
9 | 20 | processes = [] |
|
10 | 21 | blackhole = tempfile.TemporaryFile() |
|
11 | 22 | |
|
12 | 23 | # nose setup/teardown |
|
13 | 24 | |
|
14 | 25 | def setup(): |
|
15 | 26 | cp = Popen('ipcontrollerz --profile iptest -r --log-level 10 --log-to-file'.split(), stdout=blackhole, stderr=STDOUT) |
|
16 | 27 | processes.append(cp) |
|
17 | 28 | time.sleep(.5) |
|
18 | add_engine() | |
|
29 | add_engines(1) | |
|
19 | 30 | c = client.Client(profile='iptest') |
|
20 | 31 | while not c.ids: |
|
21 | 32 | time.sleep(.1) |
|
22 | 33 | c.spin() |
|
34 | c.close() | |
|
23 | 35 | |
|
24 | def add_engine(profile='iptest'): | |
|
36 | def add_engines(n=1, profile='iptest'): | |
|
37 | rc = client.Client(profile=profile) | |
|
38 | base = len(rc) | |
|
39 | eps = [] | |
|
40 | for i in range(n): | |
|
25 | 41 | ep = Popen(['ipenginez']+ ['--profile', profile, '--log-level', '10', '--log-to-file'], stdout=blackhole, stderr=STDOUT) |
|
26 | 42 | # ep.start() |
|
27 | 43 | processes.append(ep) |
|
28 | return ep | |
|
44 | eps.append(ep) | |
|
45 | while len(rc) < base+n: | |
|
46 | time.sleep(.1) | |
|
47 | rc.spin() | |
|
48 | rc.close() | |
|
49 | return eps | |
|
29 | 50 | |
|
30 | 51 | def teardown(): |
|
31 | 52 | time.sleep(1) |
|
32 | 53 | while processes: |
|
33 | 54 | p = processes.pop() |
|
34 | 55 | if p.poll() is None: |
|
35 | 56 | try: |
|
36 | 57 | p.terminate() |
|
37 | 58 | except Exception, e: |
|
38 | 59 | print e |
|
39 | 60 | pass |
|
40 | 61 | if p.poll() is None: |
|
41 | 62 | time.sleep(.25) |
|
42 | 63 | if p.poll() is None: |
|
43 | 64 | try: |
|
44 | 65 | print 'killing' |
|
45 | 66 | p.kill() |
|
46 | 67 | except: |
|
47 | 68 | print "couldn't shutdown process: ", p |
|
48 | 69 |
@@ -1,106 +1,119 b'' | |||
|
1 | """base class for parallel client tests""" | |
|
2 | ||
|
3 | #------------------------------------------------------------------------------- | |
|
4 | # Copyright (C) 2011 The IPython Development Team | |
|
5 | # | |
|
6 | # Distributed under the terms of the BSD License. The full license is in | |
|
7 | # the file COPYING, distributed as part of this software. | |
|
8 | #------------------------------------------------------------------------------- | |
|
9 | ||
|
1 | 10 | import sys |
|
2 | 11 | import tempfile |
|
3 | 12 | import time |
|
4 | 13 | from signal import SIGINT |
|
5 | 14 | from multiprocessing import Process |
|
6 | 15 | |
|
7 | 16 | from nose import SkipTest |
|
8 | 17 | |
|
18 | import zmq | |
|
9 | 19 | from zmq.tests import BaseZMQTestCase |
|
10 | 20 | |
|
11 | 21 | from IPython.external.decorator import decorator |
|
12 | 22 | |
|
13 | 23 | from IPython.zmq.parallel import error |
|
14 | 24 | from IPython.zmq.parallel.client import Client |
|
15 | 25 | from IPython.zmq.parallel.ipcluster import launch_process |
|
16 | 26 | from IPython.zmq.parallel.entry_point import select_random_ports |
|
17 | from IPython.zmq.parallel.tests import processes,add_engine | |
|
27 | from IPython.zmq.parallel.tests import processes,add_engines | |
|
18 | 28 | |
|
19 | 29 | # simple tasks for use in apply tests |
|
20 | 30 | |
|
21 | 31 | def segfault(): |
|
22 | 32 | """this will segfault""" |
|
23 | 33 | import ctypes |
|
24 | 34 | ctypes.memset(-1,0,1) |
|
25 | 35 | |
|
26 | 36 | def wait(n): |
|
27 | 37 | """sleep for a time""" |
|
28 | 38 | import time |
|
29 | 39 | time.sleep(n) |
|
30 | 40 | return n |
|
31 | 41 | |
|
32 | 42 | def raiser(eclass): |
|
33 | 43 | """raise an exception""" |
|
34 | 44 | raise eclass() |
|
35 | 45 | |
|
36 | 46 | # test decorator for skipping tests when libraries are unavailable |
|
37 | 47 | def skip_without(*names): |
|
38 | 48 | """skip a test if some names are not importable""" |
|
39 | 49 | @decorator |
|
40 | 50 | def skip_without_names(f, *args, **kwargs): |
|
41 | 51 | """decorator to skip tests in the absence of numpy.""" |
|
42 | 52 | for name in names: |
|
43 | 53 | try: |
|
44 | 54 | __import__(name) |
|
45 | 55 | except ImportError: |
|
46 | 56 | raise SkipTest |
|
47 | 57 | return f(*args, **kwargs) |
|
48 | 58 | return skip_without_names |
|
49 | 59 | |
|
50 | ||
|
51 | 60 | class ClusterTestCase(BaseZMQTestCase): |
|
52 | 61 | |
|
53 | 62 | def add_engines(self, n=1, block=True): |
|
54 | 63 | """add multiple engines to our cluster""" |
|
55 | for i in range(n): | |
|
56 | self.engines.append(add_engine()) | |
|
64 | self.engines.extend(add_engines(n)) | |
|
57 | 65 | if block: |
|
58 | 66 | self.wait_on_engines() |
|
59 | 67 | |
|
60 | 68 | def wait_on_engines(self, timeout=5): |
|
61 | 69 | """wait for our engines to connect.""" |
|
62 | 70 | n = len(self.engines)+self.base_engine_count |
|
63 | 71 | tic = time.time() |
|
64 | 72 | while time.time()-tic < timeout and len(self.client.ids) < n: |
|
65 | 73 | time.sleep(0.1) |
|
66 | 74 | |
|
67 | 75 | assert not len(self.client.ids) < n, "waiting for engines timed out" |
|
68 | 76 | |
|
69 | 77 | def connect_client(self): |
|
70 | 78 | """connect a client with my Context, and track its sockets for cleanup""" |
|
71 | 79 | c = Client(profile='iptest',context=self.context) |
|
72 | ||
|
73 | # for name in filter(lambda n:n.endswith('socket'), dir(c)): | |
|
74 | # self.sockets.append(getattr(c, name)) | |
|
80 | for name in filter(lambda n:n.endswith('socket'), dir(c)): | |
|
81 | s = getattr(c, name) | |
|
82 | s.setsockopt(zmq.LINGER, 0) | |
|
83 | self.sockets.append(s) | |
|
75 | 84 | return c |
|
76 | 85 | |
|
77 | 86 | def assertRaisesRemote(self, etype, f, *args, **kwargs): |
|
78 | 87 | try: |
|
79 | 88 | try: |
|
80 | 89 | f(*args, **kwargs) |
|
81 | 90 | except error.CompositeError as e: |
|
82 | 91 | e.raise_exception() |
|
83 | 92 | except error.RemoteError as e: |
|
84 | 93 | self.assertEquals(etype.__name__, e.ename, "Should have raised %r, but raised %r"%(e.ename, etype.__name__)) |
|
85 | 94 | else: |
|
86 | 95 | self.fail("should have raised a RemoteError") |
|
87 | 96 | |
|
88 | 97 | def setUp(self): |
|
89 | 98 | BaseZMQTestCase.setUp(self) |
|
90 | 99 | self.client = self.connect_client() |
|
91 | 100 | self.base_engine_count=len(self.client.ids) |
|
92 | 101 | self.engines=[] |
|
93 | 102 | |
|
94 | 103 | def tearDown(self): |
|
95 | ||
|
104 | # self.client.clear(block=True) | |
|
96 | 105 | # close fds: |
|
97 | 106 | for e in filter(lambda e: e.poll() is not None, processes): |
|
98 | 107 | processes.remove(e) |
|
99 | 108 | |
|
109 | # allow flushing of incoming messages to prevent crash on socket close | |
|
110 | self.client.wait(timeout=2) | |
|
111 | # time.sleep(2) | |
|
112 | self.client.spin() | |
|
100 | 113 | self.client.close() |
|
101 | 114 | BaseZMQTestCase.tearDown(self) |
|
102 |
# this will be |
|
|
103 | self.context.term() | |
|
115 | # this will be redundant when pyzmq merges PR #88 | |
|
116 | # self.context.term() | |
|
104 | 117 | # print tempfile.TemporaryFile().fileno(), |
|
105 | 118 | # sys.stdout.flush() |
|
106 | 119 | No newline at end of file |
@@ -1,262 +1,147 b'' | |||
|
1 | """Tests for parallel client.py""" | |
|
2 | ||
|
3 | #------------------------------------------------------------------------------- | |
|
4 | # Copyright (C) 2011 The IPython Development Team | |
|
5 | # | |
|
6 | # Distributed under the terms of the BSD License. The full license is in | |
|
7 | # the file COPYING, distributed as part of this software. | |
|
8 | #------------------------------------------------------------------------------- | |
|
9 | ||
|
10 | #------------------------------------------------------------------------------- | |
|
11 | # Imports | |
|
12 | #------------------------------------------------------------------------------- | |
|
13 | ||
|
1 | 14 | import time |
|
2 | 15 | from tempfile import mktemp |
|
3 | 16 | |
|
4 | 17 | import zmq |
|
5 | 18 | |
|
6 | 19 | from IPython.zmq.parallel import client as clientmod |
|
7 | 20 | from IPython.zmq.parallel import error |
|
8 | 21 | from IPython.zmq.parallel.asyncresult import AsyncResult, AsyncHubResult |
|
9 | 22 | from IPython.zmq.parallel.view import LoadBalancedView, DirectView |
|
10 | 23 | |
|
11 | from clienttest import ClusterTestCase, segfault, wait | |
|
24 | from clienttest import ClusterTestCase, segfault, wait, add_engines | |
|
25 | ||
|
26 | def setup(): | |
|
27 | add_engines(4) | |
|
12 | 28 | |
|
13 | 29 | class TestClient(ClusterTestCase): |
|
14 | 30 | |
|
15 | 31 | def test_ids(self): |
|
16 | 32 | n = len(self.client.ids) |
|
17 | 33 | self.add_engines(3) |
|
18 | 34 | self.assertEquals(len(self.client.ids), n+3) |
|
19 | 35 | |
|
20 | def test_segfault_task(self): | |
|
21 | """test graceful handling of engine death (balanced)""" | |
|
22 | self.add_engines(1) | |
|
23 | ar = self.client.apply(segfault, block=False) | |
|
24 | self.assertRaisesRemote(error.EngineError, ar.get) | |
|
25 | eid = ar.engine_id | |
|
26 | while eid in self.client.ids: | |
|
27 | time.sleep(.01) | |
|
28 | self.client.spin() | |
|
29 | ||
|
30 | def test_segfault_mux(self): | |
|
31 | """test graceful handling of engine death (direct)""" | |
|
32 | self.add_engines(1) | |
|
33 | eid = self.client.ids[-1] | |
|
34 | ar = self.client[eid].apply_async(segfault) | |
|
35 | self.assertRaisesRemote(error.EngineError, ar.get) | |
|
36 | eid = ar.engine_id | |
|
37 | while eid in self.client.ids: | |
|
38 | time.sleep(.01) | |
|
39 | self.client.spin() | |
|
40 | ||
|
41 | 36 | def test_view_indexing(self): |
|
42 | 37 | """test index access for views""" |
|
43 | 38 | self.add_engines(2) |
|
44 | 39 | targets = self.client._build_targets('all')[-1] |
|
45 | 40 | v = self.client[:] |
|
46 | 41 | self.assertEquals(v.targets, targets) |
|
47 | 42 | t = self.client.ids[2] |
|
48 | 43 | v = self.client[t] |
|
49 | 44 | self.assert_(isinstance(v, DirectView)) |
|
50 | 45 | self.assertEquals(v.targets, t) |
|
51 | 46 | t = self.client.ids[2:4] |
|
52 | 47 | v = self.client[t] |
|
53 | 48 | self.assert_(isinstance(v, DirectView)) |
|
54 | 49 | self.assertEquals(v.targets, t) |
|
55 | 50 | v = self.client[::2] |
|
56 | 51 | self.assert_(isinstance(v, DirectView)) |
|
57 | 52 | self.assertEquals(v.targets, targets[::2]) |
|
58 | 53 | v = self.client[1::3] |
|
59 | 54 | self.assert_(isinstance(v, DirectView)) |
|
60 | 55 | self.assertEquals(v.targets, targets[1::3]) |
|
61 | 56 | v = self.client[:-3] |
|
62 | 57 | self.assert_(isinstance(v, DirectView)) |
|
63 | 58 | self.assertEquals(v.targets, targets[:-3]) |
|
64 | 59 | v = self.client[-1] |
|
65 | 60 | self.assert_(isinstance(v, DirectView)) |
|
66 | 61 | self.assertEquals(v.targets, targets[-1]) |
|
67 | 62 | self.assertRaises(TypeError, lambda : self.client[None]) |
|
68 | 63 | |
|
69 | 64 | def test_view_cache(self): |
|
70 | 65 | """test that multiple view requests return the same object""" |
|
71 | 66 | v = self.client[:2] |
|
72 | 67 | v2 =self.client[:2] |
|
73 | 68 | self.assertTrue(v is v2) |
|
74 | v = self.client.view() | |
|
75 |
v2 = self.client.view( |
|
|
69 | v = self.client.load_balanced_view() | |
|
70 | v2 = self.client.load_balanced_view(targets=None) | |
|
76 | 71 | self.assertTrue(v is v2) |
|
77 | 72 | |
|
78 | 73 | def test_targets(self): |
|
79 | 74 | """test various valid targets arguments""" |
|
80 | 75 | build = self.client._build_targets |
|
81 | 76 | ids = self.client.ids |
|
82 | 77 | idents,targets = build(None) |
|
83 | 78 | self.assertEquals(ids, targets) |
|
84 | 79 | |
|
85 | 80 | def test_clear(self): |
|
86 | 81 | """test clear behavior""" |
|
87 | self.add_engines(2) | |
|
88 |
self.client |
|
|
89 | self.client.push(dict(a=5)) | |
|
90 | self.client.pull('a') | |
|
82 | # self.add_engines(2) | |
|
83 | v = self.client[:] | |
|
84 | v.block=True | |
|
85 | v.push(dict(a=5)) | |
|
86 | v.pull('a') | |
|
91 | 87 | id0 = self.client.ids[-1] |
|
92 | 88 | self.client.clear(targets=id0) |
|
93 |
self.client.pull('a' |
|
|
94 |
self.assertRaisesRemote(NameError, self.client. |
|
|
95 | self.client.clear() | |
|
89 | self.client[:-1].pull('a') | |
|
90 | self.assertRaisesRemote(NameError, self.client[id0].get, 'a') | |
|
91 | self.client.clear(block=True) | |
|
96 | 92 | for i in self.client.ids: |
|
97 | self.assertRaisesRemote(NameError, self.client.pull, 'a', targets=i) | |
|
98 | ||
|
99 | ||
|
100 | def test_push_pull(self): | |
|
101 | """test pushing and pulling""" | |
|
102 | data = dict(a=10, b=1.05, c=range(10), d={'e':(1,2),'f':'hi'}) | |
|
103 | t = self.client.ids[-1] | |
|
104 | self.add_engines(2) | |
|
105 | push = self.client.push | |
|
106 | pull = self.client.pull | |
|
107 | self.client.block=True | |
|
108 | nengines = len(self.client) | |
|
109 | push({'data':data}, targets=t) | |
|
110 | d = pull('data', targets=t) | |
|
111 | self.assertEquals(d, data) | |
|
112 | push({'data':data}) | |
|
113 | d = pull('data') | |
|
114 | self.assertEquals(d, nengines*[data]) | |
|
115 | ar = push({'data':data}, block=False) | |
|
116 | self.assertTrue(isinstance(ar, AsyncResult)) | |
|
117 | r = ar.get() | |
|
118 | ar = pull('data', block=False) | |
|
119 | self.assertTrue(isinstance(ar, AsyncResult)) | |
|
120 | r = ar.get() | |
|
121 | self.assertEquals(r, nengines*[data]) | |
|
122 | push(dict(a=10,b=20)) | |
|
123 | r = pull(('a','b')) | |
|
124 | self.assertEquals(r, nengines*[[10,20]]) | |
|
125 | ||
|
126 | def test_push_pull_function(self): | |
|
127 | "test pushing and pulling functions" | |
|
128 | def testf(x): | |
|
129 | return 2.0*x | |
|
130 | ||
|
131 | self.add_engines(4) | |
|
132 | t = self.client.ids[-1] | |
|
133 | self.client.block=True | |
|
134 | push = self.client.push | |
|
135 | pull = self.client.pull | |
|
136 | execute = self.client.execute | |
|
137 | push({'testf':testf}, targets=t) | |
|
138 | r = pull('testf', targets=t) | |
|
139 | self.assertEqual(r(1.0), testf(1.0)) | |
|
140 | execute('r = testf(10)', targets=t) | |
|
141 | r = pull('r', targets=t) | |
|
142 | self.assertEquals(r, testf(10)) | |
|
143 | ar = push({'testf':testf}, block=False) | |
|
144 | ar.get() | |
|
145 | ar = pull('testf', block=False) | |
|
146 | rlist = ar.get() | |
|
147 | for r in rlist: | |
|
148 | self.assertEqual(r(1.0), testf(1.0)) | |
|
149 | execute("def g(x): return x*x", targets=t) | |
|
150 | r = pull(('testf','g'),targets=t) | |
|
151 | self.assertEquals((r[0](10),r[1](10)), (testf(10), 100)) | |
|
152 | ||
|
153 | def test_push_function_globals(self): | |
|
154 | """test that pushed functions have access to globals""" | |
|
155 | def geta(): | |
|
156 | return a | |
|
157 | self.add_engines(1) | |
|
158 | v = self.client[-1] | |
|
159 | v.block=True | |
|
160 | v['f'] = geta | |
|
161 | self.assertRaisesRemote(NameError, v.execute, 'b=f()') | |
|
162 | v.execute('a=5') | |
|
163 | v.execute('b=f()') | |
|
164 | self.assertEquals(v['b'], 5) | |
|
165 | ||
|
166 | def test_push_function_defaults(self): | |
|
167 | """test that pushed functions preserve default args""" | |
|
168 | def echo(a=10): | |
|
169 | return a | |
|
170 | self.add_engines(1) | |
|
171 | v = self.client[-1] | |
|
172 | v.block=True | |
|
173 | v['f'] = echo | |
|
174 | v.execute('b=f()') | |
|
175 | self.assertEquals(v['b'], 10) | |
|
93 | # print i | |
|
94 | self.assertRaisesRemote(NameError, self.client[i].get, 'a') | |
|
176 | 95 | |
|
177 | 96 | def test_get_result(self): |
|
178 | 97 | """test getting results from the Hub.""" |
|
179 | 98 | c = clientmod.Client(profile='iptest') |
|
180 | self.add_engines(1) | |
|
99 | # self.add_engines(1) | |
|
181 | 100 | t = c.ids[-1] |
|
182 |
ar = c.apply(wait, |
|
|
101 | ar = c[t].apply_async(wait, 1) | |
|
183 | 102 | # give the monitor time to notice the message |
|
184 | 103 | time.sleep(.25) |
|
185 | 104 | ahr = self.client.get_result(ar.msg_ids) |
|
186 | 105 | self.assertTrue(isinstance(ahr, AsyncHubResult)) |
|
187 | 106 | self.assertEquals(ahr.get(), ar.get()) |
|
188 | 107 | ar2 = self.client.get_result(ar.msg_ids) |
|
189 | 108 | self.assertFalse(isinstance(ar2, AsyncHubResult)) |
|
109 | c.close() | |
|
190 | 110 | |
|
191 | 111 | def test_ids_list(self): |
|
192 | 112 | """test client.ids""" |
|
193 | self.add_engines(2) | |
|
113 | # self.add_engines(2) | |
|
194 | 114 | ids = self.client.ids |
|
195 | 115 | self.assertEquals(ids, self.client._ids) |
|
196 | 116 | self.assertFalse(ids is self.client._ids) |
|
197 | 117 | ids.remove(ids[-1]) |
|
198 | 118 | self.assertNotEquals(ids, self.client._ids) |
|
199 | 119 | |
|
200 |
def test_ |
|
|
201 | """test that run appends newline to files""" | |
|
202 | tmpfile = mktemp() | |
|
203 | with open(tmpfile, 'w') as f: | |
|
204 | f.write("""def g(): | |
|
205 | return 5 | |
|
206 | """) | |
|
207 |
|
|
|
208 | v.run(tmpfile, block=True) | |
|
209 | self.assertEquals(v.apply_sync(lambda : g()), 5) | |
|
210 | ||
|
211 | def test_apply_tracked(self): | |
|
212 | """test tracking for apply""" | |
|
213 | # self.add_engines(1) | |
|
214 | t = self.client.ids[-1] | |
|
215 | self.client.block=False | |
|
216 | def echo(n=1024*1024, **kwargs): | |
|
217 | return self.client.apply(lambda x: x, args=('x'*n,), targets=t, **kwargs) | |
|
218 | ar = echo(1) | |
|
219 | self.assertTrue(ar._tracker is None) | |
|
220 | self.assertTrue(ar.sent) | |
|
221 | ar = echo(track=True) | |
|
222 | self.assertTrue(isinstance(ar._tracker, zmq.MessageTracker)) | |
|
223 | self.assertEquals(ar.sent, ar._tracker.done) | |
|
224 | ar._tracker.wait() | |
|
225 | self.assertTrue(ar.sent) | |
|
226 | ||
|
227 | def test_push_tracked(self): | |
|
228 | t = self.client.ids[-1] | |
|
229 | ns = dict(x='x'*1024*1024) | |
|
230 | ar = self.client.push(ns, targets=t, block=False) | |
|
231 | self.assertTrue(ar._tracker is None) | |
|
232 | self.assertTrue(ar.sent) | |
|
233 | ||
|
234 | ar = self.client.push(ns, targets=t, block=False, track=True) | |
|
235 | self.assertTrue(isinstance(ar._tracker, zmq.MessageTracker)) | |
|
236 | self.assertEquals(ar.sent, ar._tracker.done) | |
|
237 | ar._tracker.wait() | |
|
238 | self.assertTrue(ar.sent) | |
|
239 | ar.get() | |
|
240 | ||
|
241 | def test_scatter_tracked(self): | |
|
242 | t = self.client.ids | |
|
243 | x='x'*1024*1024 | |
|
244 | ar = self.client.scatter('x', x, targets=t, block=False) | |
|
245 | self.assertTrue(ar._tracker is None) | |
|
246 | self.assertTrue(ar.sent) | |
|
247 | ||
|
248 | ar = self.client.scatter('x', x, targets=t, block=False, track=True) | |
|
249 | self.assertTrue(isinstance(ar._tracker, zmq.MessageTracker)) | |
|
250 | self.assertEquals(ar.sent, ar._tracker.done) | |
|
251 | ar._tracker.wait() | |
|
252 | self.assertTrue(ar.sent) | |
|
253 | ar.get() | |
|
254 | ||
|
255 | def test_remote_reference(self): | |
|
256 | v = self.client[-1] | |
|
257 | v['a'] = 123 | |
|
258 | ra = clientmod.Reference('a') | |
|
259 | b = v.apply_sync(lambda x: x, ra) | |
|
260 | self.assertEquals(b, 123) | |
|
120 | def test_queue_status(self): | |
|
121 | # self.addEngine(4) | |
|
122 | ids = self.client.ids | |
|
123 | id0 = ids[0] | |
|
124 | qs = self.client.queue_status(targets=id0) | |
|
125 | self.assertTrue(isinstance(qs, dict)) | |
|
126 | self.assertEquals(sorted(qs.keys()), ['completed', 'queue', 'tasks']) | |
|
127 | allqs = self.client.queue_status() | |
|
128 | self.assertTrue(isinstance(allqs, dict)) | |
|
129 | self.assertEquals(sorted(allqs.keys()), self.client.ids) | |
|
130 | for eid,qs in allqs.items(): | |
|
131 | self.assertTrue(isinstance(qs, dict)) | |
|
132 | self.assertEquals(sorted(qs.keys()), ['completed', 'queue', 'tasks']) | |
|
133 | ||
|
134 | def test_shutdown(self): | |
|
135 | # self.addEngine(4) | |
|
136 | ids = self.client.ids | |
|
137 | id0 = ids[0] | |
|
138 | self.client.shutdown(id0, block=True) | |
|
139 | while id0 in self.client.ids: | |
|
140 | time.sleep(0.1) | |
|
141 | self.client.spin() | |
|
261 | 142 | |
|
143 | self.assertRaises(IndexError, lambda : self.client[id0]) | |
|
262 | 144 | |
|
145 | def test_result_status(self): | |
|
146 | pass | |
|
147 | # to be written |
@@ -1,87 +1,108 b'' | |||
|
1 | 1 | """test serialization with newserialized""" |
|
2 | 2 | |
|
3 | #------------------------------------------------------------------------------- | |
|
4 | # Copyright (C) 2011 The IPython Development Team | |
|
5 | # | |
|
6 | # Distributed under the terms of the BSD License. The full license is in | |
|
7 | # the file COPYING, distributed as part of this software. | |
|
8 | #------------------------------------------------------------------------------- | |
|
9 | ||
|
10 | #------------------------------------------------------------------------------- | |
|
11 | # Imports | |
|
12 | #------------------------------------------------------------------------------- | |
|
13 | ||
|
3 | 14 | from unittest import TestCase |
|
4 | 15 | |
|
5 | 16 | from IPython.testing.parametric import parametric |
|
6 | 17 | from IPython.utils import newserialized as ns |
|
7 | 18 | from IPython.utils.pickleutil import can, uncan, CannedObject, CannedFunction |
|
8 | 19 | from IPython.zmq.parallel.tests.clienttest import skip_without |
|
9 | 20 | |
|
10 | 21 | |
|
11 | 22 | class CanningTestCase(TestCase): |
|
12 | 23 | def test_canning(self): |
|
13 | 24 | d = dict(a=5,b=6) |
|
14 | 25 | cd = can(d) |
|
15 | 26 | self.assertTrue(isinstance(cd, dict)) |
|
16 | 27 | |
|
17 | 28 | def test_canned_function(self): |
|
18 | 29 | f = lambda : 7 |
|
19 | 30 | cf = can(f) |
|
20 | 31 | self.assertTrue(isinstance(cf, CannedFunction)) |
|
21 | 32 | |
|
22 | 33 | @parametric |
|
23 | 34 | def test_can_roundtrip(cls): |
|
24 | 35 | objs = [ |
|
25 | 36 | dict(), |
|
26 | 37 | set(), |
|
27 | 38 | list(), |
|
28 | 39 | ['a',1,['a',1],u'e'], |
|
29 | 40 | ] |
|
30 | 41 | return map(cls.run_roundtrip, objs) |
|
31 | 42 | |
|
32 | 43 | @classmethod |
|
33 | 44 | def run_roundtrip(self, obj): |
|
34 | 45 | o = uncan(can(obj)) |
|
35 | 46 | assert o == obj, "failed assertion: %r == %r"%(o,obj) |
|
36 | 47 | |
|
37 | 48 | def test_serialized_interfaces(self): |
|
38 | 49 | |
|
39 | 50 | us = {'a':10, 'b':range(10)} |
|
40 | 51 | s = ns.serialize(us) |
|
41 | 52 | uus = ns.unserialize(s) |
|
42 | 53 | self.assertTrue(isinstance(s, ns.SerializeIt)) |
|
43 | 54 | self.assertEquals(uus, us) |
|
44 | 55 | |
|
45 | 56 | def test_pickle_serialized(self): |
|
46 | 57 | obj = {'a':1.45345, 'b':'asdfsdf', 'c':10000L} |
|
47 | 58 | original = ns.UnSerialized(obj) |
|
48 | 59 | originalSer = ns.SerializeIt(original) |
|
49 | 60 | firstData = originalSer.getData() |
|
50 | 61 | firstTD = originalSer.getTypeDescriptor() |
|
51 | 62 | firstMD = originalSer.getMetadata() |
|
52 | 63 | self.assertEquals(firstTD, 'pickle') |
|
53 | 64 | self.assertEquals(firstMD, {}) |
|
54 | 65 | unSerialized = ns.UnSerializeIt(originalSer) |
|
55 | 66 | secondObj = unSerialized.getObject() |
|
56 | 67 | for k, v in secondObj.iteritems(): |
|
57 | 68 | self.assertEquals(obj[k], v) |
|
58 | 69 | secondSer = ns.SerializeIt(ns.UnSerialized(secondObj)) |
|
59 | 70 | self.assertEquals(firstData, secondSer.getData()) |
|
60 | 71 | self.assertEquals(firstTD, secondSer.getTypeDescriptor() ) |
|
61 | 72 | self.assertEquals(firstMD, secondSer.getMetadata()) |
|
62 | 73 | |
|
63 | 74 | @skip_without('numpy') |
|
64 | 75 | def test_ndarray_serialized(self): |
|
65 | 76 | import numpy |
|
66 | 77 | a = numpy.linspace(0.0, 1.0, 1000) |
|
67 | 78 | unSer1 = ns.UnSerialized(a) |
|
68 | 79 | ser1 = ns.SerializeIt(unSer1) |
|
69 | 80 | td = ser1.getTypeDescriptor() |
|
70 | 81 | self.assertEquals(td, 'ndarray') |
|
71 | 82 | md = ser1.getMetadata() |
|
72 | 83 | self.assertEquals(md['shape'], a.shape) |
|
73 | 84 | self.assertEquals(md['dtype'], a.dtype.str) |
|
74 | 85 | buff = ser1.getData() |
|
75 | 86 | self.assertEquals(buff, numpy.getbuffer(a)) |
|
76 | 87 | s = ns.Serialized(buff, td, md) |
|
77 | 88 | final = ns.unserialize(s) |
|
78 | 89 | self.assertEquals(numpy.getbuffer(a), numpy.getbuffer(final)) |
|
79 | 90 | self.assertTrue((a==final).all()) |
|
80 | 91 | self.assertEquals(a.dtype.str, final.dtype.str) |
|
81 | 92 | self.assertEquals(a.shape, final.shape) |
|
82 | 93 | # test non-copying: |
|
83 | 94 | a[2] = 1e9 |
|
84 | 95 | self.assertTrue((a==final).all()) |
|
85 | 96 |
|
|
97 | def test_uncan_function_globals(self): | |
|
98 | """test that uncanning a module function restores it into its module""" | |
|
99 | from re import search | |
|
100 | cf = can(search) | |
|
101 | csearch = uncan(cf) | |
|
102 | self.assertEqual(csearch.__module__, search.__module__) | |
|
103 | self.assertNotEqual(csearch('asd', 'asdf'), None) | |
|
104 | csearch = uncan(cf, dict(a=5)) | |
|
105 | self.assertEqual(csearch.__module__, search.__module__) | |
|
106 | self.assertNotEqual(csearch('asd', 'asdf'), None) | |
|
86 | 107 | |
|
87 | 108 | No newline at end of file |
@@ -1,99 +1,111 b'' | |||
|
1 | """test building messages with streamsession""" | |
|
2 | ||
|
3 | #------------------------------------------------------------------------------- | |
|
4 | # Copyright (C) 2011 The IPython Development Team | |
|
5 | # | |
|
6 | # Distributed under the terms of the BSD License. The full license is in | |
|
7 | # the file COPYING, distributed as part of this software. | |
|
8 | #------------------------------------------------------------------------------- | |
|
9 | ||
|
10 | #------------------------------------------------------------------------------- | |
|
11 | # Imports | |
|
12 | #------------------------------------------------------------------------------- | |
|
1 | 13 | |
|
2 | 14 | import os |
|
3 | 15 | import uuid |
|
4 | 16 | import zmq |
|
5 | 17 | |
|
6 | 18 | from zmq.tests import BaseZMQTestCase |
|
7 | 19 | from zmq.eventloop.zmqstream import ZMQStream |
|
8 | 20 | # from IPython.zmq.tests import SessionTestCase |
|
9 | 21 | from IPython.zmq.parallel import streamsession as ss |
|
10 | 22 | |
|
11 | 23 | class SessionTestCase(BaseZMQTestCase): |
|
12 | 24 | |
|
13 | 25 | def setUp(self): |
|
14 | 26 | BaseZMQTestCase.setUp(self) |
|
15 | 27 | self.session = ss.StreamSession() |
|
16 | 28 | |
|
17 | 29 | class TestSession(SessionTestCase): |
|
18 | 30 | |
|
19 | 31 | def test_msg(self): |
|
20 | 32 | """message format""" |
|
21 | 33 | msg = self.session.msg('execute') |
|
22 | 34 | thekeys = set('header msg_id parent_header msg_type content'.split()) |
|
23 | 35 | s = set(msg.keys()) |
|
24 | 36 | self.assertEquals(s, thekeys) |
|
25 | 37 | self.assertTrue(isinstance(msg['content'],dict)) |
|
26 | 38 | self.assertTrue(isinstance(msg['header'],dict)) |
|
27 | 39 | self.assertTrue(isinstance(msg['parent_header'],dict)) |
|
28 | 40 | self.assertEquals(msg['msg_type'], 'execute') |
|
29 | 41 | |
|
30 | 42 | |
|
31 | 43 | |
|
32 | 44 | def test_args(self): |
|
33 | 45 | """initialization arguments for StreamSession""" |
|
34 | 46 | s = self.session |
|
35 | 47 | self.assertTrue(s.pack is ss.default_packer) |
|
36 | 48 | self.assertTrue(s.unpack is ss.default_unpacker) |
|
37 | 49 | self.assertEquals(s.username, os.environ.get('USER', 'username')) |
|
38 | 50 | |
|
39 | 51 | s = ss.StreamSession(username=None) |
|
40 | 52 | self.assertEquals(s.username, os.environ.get('USER', 'username')) |
|
41 | 53 | |
|
42 | 54 | self.assertRaises(TypeError, ss.StreamSession, packer='hi') |
|
43 | 55 | self.assertRaises(TypeError, ss.StreamSession, unpacker='hi') |
|
44 | 56 | u = str(uuid.uuid4()) |
|
45 | 57 | s = ss.StreamSession(username='carrot', session=u) |
|
46 | 58 | self.assertEquals(s.session, u) |
|
47 | 59 | self.assertEquals(s.username, 'carrot') |
|
48 | 60 | |
|
49 | 61 | def test_tracking(self): |
|
50 | 62 | """test tracking messages""" |
|
51 | 63 | a,b = self.create_bound_pair(zmq.PAIR, zmq.PAIR) |
|
52 | 64 | s = self.session |
|
53 | 65 | stream = ZMQStream(a) |
|
54 | 66 | msg = s.send(a, 'hello', track=False) |
|
55 | 67 | self.assertTrue(msg['tracker'] is None) |
|
56 | 68 | msg = s.send(a, 'hello', track=True) |
|
57 | 69 | self.assertTrue(isinstance(msg['tracker'], zmq.MessageTracker)) |
|
58 | 70 | M = zmq.Message(b'hi there', track=True) |
|
59 | 71 | msg = s.send(a, 'hello', buffers=[M], track=True) |
|
60 | 72 | t = msg['tracker'] |
|
61 | 73 | self.assertTrue(isinstance(t, zmq.MessageTracker)) |
|
62 | 74 | self.assertRaises(zmq.NotDone, t.wait, .1) |
|
63 | 75 | del M |
|
64 | 76 | t.wait(1) # this will raise |
|
65 | 77 | |
|
66 | 78 | |
|
67 | 79 | # def test_rekey(self): |
|
68 | 80 | # """rekeying dict around json str keys""" |
|
69 | 81 | # d = {'0': uuid.uuid4(), 0:uuid.uuid4()} |
|
70 | 82 | # self.assertRaises(KeyError, ss.rekey, d) |
|
71 | 83 | # |
|
72 | 84 | # d = {'0': uuid.uuid4(), 1:uuid.uuid4(), 'asdf':uuid.uuid4()} |
|
73 | 85 | # d2 = {0:d['0'],1:d[1],'asdf':d['asdf']} |
|
74 | 86 | # rd = ss.rekey(d) |
|
75 | 87 | # self.assertEquals(d2,rd) |
|
76 | 88 | # |
|
77 | 89 | # d = {'1.5':uuid.uuid4(),'1':uuid.uuid4()} |
|
78 | 90 | # d2 = {1.5:d['1.5'],1:d['1']} |
|
79 | 91 | # rd = ss.rekey(d) |
|
80 | 92 | # self.assertEquals(d2,rd) |
|
81 | 93 | # |
|
82 | 94 | # d = {'1.0':uuid.uuid4(),'1':uuid.uuid4()} |
|
83 | 95 | # self.assertRaises(KeyError, ss.rekey, d) |
|
84 | 96 | # |
|
85 | 97 | def test_unique_msg_ids(self): |
|
86 | 98 | """test that messages receive unique ids""" |
|
87 | 99 | ids = set() |
|
88 | 100 | for i in range(2**12): |
|
89 | 101 | h = self.session.msg_header('test') |
|
90 | 102 | msg_id = h['msg_id'] |
|
91 | 103 | self.assertTrue(msg_id not in ids) |
|
92 | 104 | ids.add(msg_id) |
|
93 | 105 | |
|
94 | 106 | def test_feed_identities(self): |
|
95 | 107 | """scrub the front for zmq IDENTITIES""" |
|
96 | 108 | theids = "engine client other".split() |
|
97 | 109 | content = dict(code='whoda',stuff=object()) |
|
98 | 110 | themsg = self.session.msg('execute',content=content) |
|
99 | 111 | pmsg = theids |
@@ -1,318 +1,354 b'' | |||
|
1 | 1 | """some generic utilities for dealing with classes, urls, and serialization""" |
|
2 | 2 | #----------------------------------------------------------------------------- |
|
3 | 3 | # Copyright (C) 2010-2011 The IPython Development Team |
|
4 | 4 | # |
|
5 | 5 | # Distributed under the terms of the BSD License. The full license is in |
|
6 | 6 | # the file COPYING, distributed as part of this software. |
|
7 | 7 | #----------------------------------------------------------------------------- |
|
8 | 8 | |
|
9 | 9 | #----------------------------------------------------------------------------- |
|
10 | 10 | # Imports |
|
11 | 11 | #----------------------------------------------------------------------------- |
|
12 | 12 | |
|
13 | 13 | import re |
|
14 | 14 | import socket |
|
15 | 15 | |
|
16 | 16 | try: |
|
17 | 17 | import cPickle |
|
18 | 18 | pickle = cPickle |
|
19 | 19 | except: |
|
20 | 20 | cPickle = None |
|
21 | 21 | import pickle |
|
22 | 22 | |
|
23 | 23 | |
|
24 | 24 | from IPython.utils.pickleutil import can, uncan, canSequence, uncanSequence |
|
25 | 25 | from IPython.utils.newserialized import serialize, unserialize |
|
26 | 26 | |
|
27 | 27 | ISO8601="%Y-%m-%dT%H:%M:%S.%f" |
|
28 | 28 | |
|
29 | 29 | #----------------------------------------------------------------------------- |
|
30 | 30 | # Classes |
|
31 | 31 | #----------------------------------------------------------------------------- |
|
32 | 32 | |
|
33 | 33 | class Namespace(dict): |
|
34 | 34 | """Subclass of dict for attribute access to keys.""" |
|
35 | 35 | |
|
36 | 36 | def __getattr__(self, key): |
|
37 | 37 | """getattr aliased to getitem""" |
|
38 | 38 | if key in self.iterkeys(): |
|
39 | 39 | return self[key] |
|
40 | 40 | else: |
|
41 | 41 | raise NameError(key) |
|
42 | 42 | |
|
43 | 43 | def __setattr__(self, key, value): |
|
44 | 44 | """setattr aliased to setitem, with strict""" |
|
45 | 45 | if hasattr(dict, key): |
|
46 | 46 | raise KeyError("Cannot override dict keys %r"%key) |
|
47 | 47 | self[key] = value |
|
48 | 48 | |
|
49 | 49 | |
|
50 | 50 | class ReverseDict(dict): |
|
51 | 51 | """simple double-keyed subset of dict methods.""" |
|
52 | 52 | |
|
53 | 53 | def __init__(self, *args, **kwargs): |
|
54 | 54 | dict.__init__(self, *args, **kwargs) |
|
55 | 55 | self._reverse = dict() |
|
56 | 56 | for key, value in self.iteritems(): |
|
57 | 57 | self._reverse[value] = key |
|
58 | 58 | |
|
59 | 59 | def __getitem__(self, key): |
|
60 | 60 | try: |
|
61 | 61 | return dict.__getitem__(self, key) |
|
62 | 62 | except KeyError: |
|
63 | 63 | return self._reverse[key] |
|
64 | 64 | |
|
65 | 65 | def __setitem__(self, key, value): |
|
66 | 66 | if key in self._reverse: |
|
67 | 67 | raise KeyError("Can't have key %r on both sides!"%key) |
|
68 | 68 | dict.__setitem__(self, key, value) |
|
69 | 69 | self._reverse[value] = key |
|
70 | 70 | |
|
71 | 71 | def pop(self, key): |
|
72 | 72 | value = dict.pop(self, key) |
|
73 | 73 | self._reverse.pop(value) |
|
74 | 74 | return value |
|
75 | 75 | |
|
76 | 76 | def get(self, key, default=None): |
|
77 | 77 | try: |
|
78 | 78 | return self[key] |
|
79 | 79 | except KeyError: |
|
80 | 80 | return default |
|
81 | 81 | |
|
82 | 82 | #----------------------------------------------------------------------------- |
|
83 | 83 | # Functions |
|
84 | 84 | #----------------------------------------------------------------------------- |
|
85 | 85 | |
|
86 | 86 | def validate_url(url): |
|
87 | 87 | """validate a url for zeromq""" |
|
88 | 88 | if not isinstance(url, basestring): |
|
89 | 89 | raise TypeError("url must be a string, not %r"%type(url)) |
|
90 | 90 | url = url.lower() |
|
91 | 91 | |
|
92 | 92 | proto_addr = url.split('://') |
|
93 | 93 | assert len(proto_addr) == 2, 'Invalid url: %r'%url |
|
94 | 94 | proto, addr = proto_addr |
|
95 | 95 | assert proto in ['tcp','pgm','epgm','ipc','inproc'], "Invalid protocol: %r"%proto |
|
96 | 96 | |
|
97 | 97 | # domain pattern adapted from http://www.regexlib.com/REDetails.aspx?regexp_id=391 |
|
98 | 98 | # author: Remi Sabourin |
|
99 | 99 | pat = re.compile(r'^([\w\d]([\w\d\-]{0,61}[\w\d])?\.)*[\w\d]([\w\d\-]{0,61}[\w\d])?$') |
|
100 | 100 | |
|
101 | 101 | if proto == 'tcp': |
|
102 | 102 | lis = addr.split(':') |
|
103 | 103 | assert len(lis) == 2, 'Invalid url: %r'%url |
|
104 | 104 | addr,s_port = lis |
|
105 | 105 | try: |
|
106 | 106 | port = int(s_port) |
|
107 | 107 | except ValueError: |
|
108 | 108 | raise AssertionError("Invalid port %r in url: %r"%(port, url)) |
|
109 | 109 | |
|
110 | 110 | assert addr == '*' or pat.match(addr) is not None, 'Invalid url: %r'%url |
|
111 | 111 | |
|
112 | 112 | else: |
|
113 | 113 | # only validate tcp urls currently |
|
114 | 114 | pass |
|
115 | 115 | |
|
116 | 116 | return True |
|
117 | 117 | |
|
118 | 118 | |
|
119 | 119 | def validate_url_container(container): |
|
120 | 120 | """validate a potentially nested collection of urls.""" |
|
121 | 121 | if isinstance(container, basestring): |
|
122 | 122 | url = container |
|
123 | 123 | return validate_url(url) |
|
124 | 124 | elif isinstance(container, dict): |
|
125 | 125 | container = container.itervalues() |
|
126 | 126 | |
|
127 | 127 | for element in container: |
|
128 | 128 | validate_url_container(element) |
|
129 | 129 | |
|
130 | 130 | |
|
131 | 131 | def split_url(url): |
|
132 | 132 | """split a zmq url (tcp://ip:port) into ('tcp','ip','port').""" |
|
133 | 133 | proto_addr = url.split('://') |
|
134 | 134 | assert len(proto_addr) == 2, 'Invalid url: %r'%url |
|
135 | 135 | proto, addr = proto_addr |
|
136 | 136 | lis = addr.split(':') |
|
137 | 137 | assert len(lis) == 2, 'Invalid url: %r'%url |
|
138 | 138 | addr,s_port = lis |
|
139 | 139 | return proto,addr,s_port |
|
140 | 140 | |
|
141 | 141 | def disambiguate_ip_address(ip, location=None): |
|
142 | 142 | """turn multi-ip interfaces '0.0.0.0' and '*' into connectable |
|
143 | 143 | ones, based on the location (default interpretation of location is localhost).""" |
|
144 | 144 | if ip in ('0.0.0.0', '*'): |
|
145 | 145 | external_ips = socket.gethostbyname_ex(socket.gethostname())[2] |
|
146 | 146 | if location is None or location in external_ips: |
|
147 | 147 | ip='127.0.0.1' |
|
148 | 148 | elif location: |
|
149 | 149 | return location |
|
150 | 150 | return ip |
|
151 | 151 | |
|
152 | 152 | def disambiguate_url(url, location=None): |
|
153 | 153 | """turn multi-ip interfaces '0.0.0.0' and '*' into connectable |
|
154 | 154 | ones, based on the location (default interpretation is localhost). |
|
155 | 155 | |
|
156 | 156 | This is for zeromq urls, such as tcp://*:10101.""" |
|
157 | 157 | try: |
|
158 | 158 | proto,ip,port = split_url(url) |
|
159 | 159 | except AssertionError: |
|
160 | 160 | # probably not tcp url; could be ipc, etc. |
|
161 | 161 | return url |
|
162 | 162 | |
|
163 | 163 | ip = disambiguate_ip_address(ip,location) |
|
164 | 164 | |
|
165 | 165 | return "%s://%s:%s"%(proto,ip,port) |
|
166 | 166 | |
|
167 | 167 | |
|
168 | 168 | def rekey(dikt): |
|
169 | 169 | """Rekey a dict that has been forced to use str keys where there should be |
|
170 | 170 | ints by json. This belongs in the jsonutil added by fperez.""" |
|
171 | 171 | for k in dikt.iterkeys(): |
|
172 | 172 | if isinstance(k, str): |
|
173 | 173 | ik=fk=None |
|
174 | 174 | try: |
|
175 | 175 | ik = int(k) |
|
176 | 176 | except ValueError: |
|
177 | 177 | try: |
|
178 | 178 | fk = float(k) |
|
179 | 179 | except ValueError: |
|
180 | 180 | continue |
|
181 | 181 | if ik is not None: |
|
182 | 182 | nk = ik |
|
183 | 183 | else: |
|
184 | 184 | nk = fk |
|
185 | 185 | if nk in dikt: |
|
186 | 186 | raise KeyError("already have key %r"%nk) |
|
187 | 187 | dikt[nk] = dikt.pop(k) |
|
188 | 188 | return dikt |
|
189 | 189 | |
|
190 | 190 | def serialize_object(obj, threshold=64e-6): |
|
191 | 191 | """Serialize an object into a list of sendable buffers. |
|
192 | 192 | |
|
193 | 193 | Parameters |
|
194 | 194 | ---------- |
|
195 | 195 | |
|
196 | 196 | obj : object |
|
197 | 197 | The object to be serialized |
|
198 | 198 | threshold : float |
|
199 | 199 | The threshold for not double-pickling the content. |
|
200 | 200 | |
|
201 | 201 | |
|
202 | 202 | Returns |
|
203 | 203 | ------- |
|
204 | 204 | ('pmd', [bufs]) : |
|
205 | 205 | where pmd is the pickled metadata wrapper, |
|
206 | 206 | bufs is a list of data buffers |
|
207 | 207 | """ |
|
208 | 208 | databuffers = [] |
|
209 | 209 | if isinstance(obj, (list, tuple)): |
|
210 | 210 | clist = canSequence(obj) |
|
211 | 211 | slist = map(serialize, clist) |
|
212 | 212 | for s in slist: |
|
213 | 213 | if s.typeDescriptor in ('buffer', 'ndarray') or s.getDataSize() > threshold: |
|
214 | 214 | databuffers.append(s.getData()) |
|
215 | 215 | s.data = None |
|
216 | 216 | return pickle.dumps(slist,-1), databuffers |
|
217 | 217 | elif isinstance(obj, dict): |
|
218 | 218 | sobj = {} |
|
219 | 219 | for k in sorted(obj.iterkeys()): |
|
220 | 220 | s = serialize(can(obj[k])) |
|
221 | 221 | if s.typeDescriptor in ('buffer', 'ndarray') or s.getDataSize() > threshold: |
|
222 | 222 | databuffers.append(s.getData()) |
|
223 | 223 | s.data = None |
|
224 | 224 | sobj[k] = s |
|
225 | 225 | return pickle.dumps(sobj,-1),databuffers |
|
226 | 226 | else: |
|
227 | 227 | s = serialize(can(obj)) |
|
228 | 228 | if s.typeDescriptor in ('buffer', 'ndarray') or s.getDataSize() > threshold: |
|
229 | 229 | databuffers.append(s.getData()) |
|
230 | 230 | s.data = None |
|
231 | 231 | return pickle.dumps(s,-1),databuffers |
|
232 | 232 | |
|
233 | 233 | |
|
234 | 234 | def unserialize_object(bufs): |
|
235 | 235 | """reconstruct an object serialized by serialize_object from data buffers.""" |
|
236 | 236 | bufs = list(bufs) |
|
237 | 237 | sobj = pickle.loads(bufs.pop(0)) |
|
238 | 238 | if isinstance(sobj, (list, tuple)): |
|
239 | 239 | for s in sobj: |
|
240 | 240 | if s.data is None: |
|
241 | 241 | s.data = bufs.pop(0) |
|
242 | 242 | return uncanSequence(map(unserialize, sobj)), bufs |
|
243 | 243 | elif isinstance(sobj, dict): |
|
244 | 244 | newobj = {} |
|
245 | 245 | for k in sorted(sobj.iterkeys()): |
|
246 | 246 | s = sobj[k] |
|
247 | 247 | if s.data is None: |
|
248 | 248 | s.data = bufs.pop(0) |
|
249 | 249 | newobj[k] = uncan(unserialize(s)) |
|
250 | 250 | return newobj, bufs |
|
251 | 251 | else: |
|
252 | 252 | if sobj.data is None: |
|
253 | 253 | sobj.data = bufs.pop(0) |
|
254 | 254 | return uncan(unserialize(sobj)), bufs |
|
255 | 255 | |
|
256 | 256 | def pack_apply_message(f, args, kwargs, threshold=64e-6): |
|
257 | 257 | """pack up a function, args, and kwargs to be sent over the wire |
|
258 | 258 | as a series of buffers. Any object whose data is larger than `threshold` |
|
259 | 259 | will not have their data copied (currently only numpy arrays support zero-copy)""" |
|
260 | 260 | msg = [pickle.dumps(can(f),-1)] |
|
261 | 261 | databuffers = [] # for large objects |
|
262 | 262 | sargs, bufs = serialize_object(args,threshold) |
|
263 | 263 | msg.append(sargs) |
|
264 | 264 | databuffers.extend(bufs) |
|
265 | 265 | skwargs, bufs = serialize_object(kwargs,threshold) |
|
266 | 266 | msg.append(skwargs) |
|
267 | 267 | databuffers.extend(bufs) |
|
268 | 268 | msg.extend(databuffers) |
|
269 | 269 | return msg |
|
270 | 270 | |
|
271 | 271 | def unpack_apply_message(bufs, g=None, copy=True): |
|
272 | 272 | """unpack f,args,kwargs from buffers packed by pack_apply_message() |
|
273 | 273 | Returns: original f,args,kwargs""" |
|
274 | 274 | bufs = list(bufs) # allow us to pop |
|
275 | 275 | assert len(bufs) >= 3, "not enough buffers!" |
|
276 | 276 | if not copy: |
|
277 | 277 | for i in range(3): |
|
278 | 278 | bufs[i] = bufs[i].bytes |
|
279 | 279 | cf = pickle.loads(bufs.pop(0)) |
|
280 | 280 | sargs = list(pickle.loads(bufs.pop(0))) |
|
281 | 281 | skwargs = dict(pickle.loads(bufs.pop(0))) |
|
282 | 282 | # print sargs, skwargs |
|
283 | 283 | f = uncan(cf, g) |
|
284 | 284 | for sa in sargs: |
|
285 | 285 | if sa.data is None: |
|
286 | 286 | m = bufs.pop(0) |
|
287 | 287 | if sa.getTypeDescriptor() in ('buffer', 'ndarray'): |
|
288 | 288 | if copy: |
|
289 | 289 | sa.data = buffer(m) |
|
290 | 290 | else: |
|
291 | 291 | sa.data = m.buffer |
|
292 | 292 | else: |
|
293 | 293 | if copy: |
|
294 | 294 | sa.data = m |
|
295 | 295 | else: |
|
296 | 296 | sa.data = m.bytes |
|
297 | 297 | |
|
298 | 298 | args = uncanSequence(map(unserialize, sargs), g) |
|
299 | 299 | kwargs = {} |
|
300 | 300 | for k in sorted(skwargs.iterkeys()): |
|
301 | 301 | sa = skwargs[k] |
|
302 | 302 | if sa.data is None: |
|
303 | 303 | m = bufs.pop(0) |
|
304 | 304 | if sa.getTypeDescriptor() in ('buffer', 'ndarray'): |
|
305 | 305 | if copy: |
|
306 | 306 | sa.data = buffer(m) |
|
307 | 307 | else: |
|
308 | 308 | sa.data = m.buffer |
|
309 | 309 | else: |
|
310 | 310 | if copy: |
|
311 | 311 | sa.data = m |
|
312 | 312 | else: |
|
313 | 313 | sa.data = m.bytes |
|
314 | 314 | |
|
315 | 315 | kwargs[k] = uncan(unserialize(sa), g) |
|
316 | 316 | |
|
317 | 317 | return f,args,kwargs |
|
318 | 318 | |
|
319 | #-------------------------------------------------------------------------- | |
|
320 | # helpers for implementing old MEC API via view.apply | |
|
321 | #-------------------------------------------------------------------------- | |
|
322 | ||
|
323 | def interactive(f): | |
|
324 | """decorator for making functions appear as interactively defined. | |
|
325 | This results in the function being linked to the user_ns as globals() | |
|
326 | instead of the module globals(). | |
|
327 | """ | |
|
328 | f.__module__ = '__main__' | |
|
329 | return f | |
|
330 | ||
|
331 | @interactive | |
|
332 | def _push(ns): | |
|
333 | """helper method for implementing `client.push` via `client.apply`""" | |
|
334 | globals().update(ns) | |
|
335 | ||
|
336 | @interactive | |
|
337 | def _pull(keys): | |
|
338 | """helper method for implementing `client.pull` via `client.apply`""" | |
|
339 | user_ns = globals() | |
|
340 | if isinstance(keys, (list,tuple, set)): | |
|
341 | for key in keys: | |
|
342 | if not user_ns.has_key(key): | |
|
343 | raise NameError("name '%s' is not defined"%key) | |
|
344 | return map(user_ns.get, keys) | |
|
345 | else: | |
|
346 | if not user_ns.has_key(keys): | |
|
347 | raise NameError("name '%s' is not defined"%keys) | |
|
348 | return user_ns.get(keys) | |
|
349 | ||
|
350 | @interactive | |
|
351 | def _execute(code): | |
|
352 | """helper method for implementing `client.execute` via `client.apply`""" | |
|
353 | exec code in globals() | |
|
354 |
This diff has been collapsed as it changes many lines, (596 lines changed) Show them Hide them | |||
@@ -1,670 +1,920 b'' | |||
|
1 | 1 | """Views of remote engines.""" |
|
2 | 2 | #----------------------------------------------------------------------------- |
|
3 | 3 | # Copyright (C) 2010 The IPython Development Team |
|
4 | 4 | # |
|
5 | 5 | # Distributed under the terms of the BSD License. The full license is in |
|
6 | 6 | # the file COPYING, distributed as part of this software. |
|
7 | 7 | #----------------------------------------------------------------------------- |
|
8 | 8 | |
|
9 | 9 | #----------------------------------------------------------------------------- |
|
10 | 10 | # Imports |
|
11 | 11 | #----------------------------------------------------------------------------- |
|
12 | 12 | |
|
13 | import warnings | |
|
14 | from contextlib import contextmanager | |
|
15 | ||
|
16 | import zmq | |
|
17 | ||
|
13 | 18 | from IPython.testing import decorators as testdec |
|
14 | 19 | from IPython.utils.traitlets import HasTraits, Any, Bool, List, Dict, Set, Int, Instance |
|
15 | 20 | |
|
16 | 21 | from IPython.external.decorator import decorator |
|
17 | 22 | |
|
18 | from .asyncresult import AsyncResult | |
|
19 | from .dependency import Dependency | |
|
23 | from . import map as Map | |
|
24 | from . import util | |
|
25 | from .asyncresult import AsyncResult, AsyncMapResult | |
|
26 | from .dependency import Dependency, dependent | |
|
20 | 27 | from .remotefunction import ParallelFunction, parallel, remote |
|
21 | 28 | |
|
22 | 29 | #----------------------------------------------------------------------------- |
|
23 | 30 | # Decorators |
|
24 | 31 | #----------------------------------------------------------------------------- |
|
25 | 32 | |
|
26 | 33 | @decorator |
|
27 | def myblock(f, self, *args, **kwargs): | |
|
28 | """override client.block with self.block during a call""" | |
|
29 | block = self.client.block | |
|
30 | self.client.block = self.block | |
|
31 | try: | |
|
32 | ret = f(self, *args, **kwargs) | |
|
33 | finally: | |
|
34 | self.client.block = block | |
|
35 | return ret | |
|
36 | ||
|
37 | @decorator | |
|
38 | 34 | def save_ids(f, self, *args, **kwargs): |
|
39 | 35 | """Keep our history and outstanding attributes up to date after a method call.""" |
|
40 | 36 | n_previous = len(self.client.history) |
|
37 | try: | |
|
41 | 38 | ret = f(self, *args, **kwargs) |
|
39 | finally: | |
|
42 | 40 | nmsgs = len(self.client.history) - n_previous |
|
43 | 41 | msg_ids = self.client.history[-nmsgs:] |
|
44 | 42 | self.history.extend(msg_ids) |
|
45 | 43 | map(self.outstanding.add, msg_ids) |
|
46 | 44 | return ret |
|
47 | 45 | |
|
48 | 46 | @decorator |
|
49 | 47 | def sync_results(f, self, *args, **kwargs): |
|
50 | 48 | """sync relevant results from self.client to our results attribute.""" |
|
51 | 49 | ret = f(self, *args, **kwargs) |
|
52 | 50 | delta = self.outstanding.difference(self.client.outstanding) |
|
53 | 51 | completed = self.outstanding.intersection(delta) |
|
54 | 52 | self.outstanding = self.outstanding.difference(completed) |
|
55 | 53 | for msg_id in completed: |
|
56 | 54 | self.results[msg_id] = self.client.results[msg_id] |
|
57 | 55 | return ret |
|
58 | 56 | |
|
59 | 57 | @decorator |
|
60 | 58 | def spin_after(f, self, *args, **kwargs): |
|
61 | 59 | """call spin after the method.""" |
|
62 | 60 | ret = f(self, *args, **kwargs) |
|
63 | 61 | self.spin() |
|
64 | 62 | return ret |
|
65 | 63 | |
|
66 | 64 | #----------------------------------------------------------------------------- |
|
67 | 65 | # Classes |
|
68 | 66 | #----------------------------------------------------------------------------- |
|
69 | 67 | |
|
70 | 68 | class View(HasTraits): |
|
71 | 69 | """Base View class for more convenint apply(f,*args,**kwargs) syntax via attributes. |
|
72 | 70 | |
|
73 | 71 | Don't use this class, use subclasses. |
|
72 | ||
|
73 | Methods | |
|
74 | ------- | |
|
75 | ||
|
76 | spin | |
|
77 | flushes incoming results and registration state changes | |
|
78 | control methods spin, and requesting `ids` also ensures up to date | |
|
79 | ||
|
80 | wait | |
|
81 | wait on one or more msg_ids | |
|
82 | ||
|
83 | execution methods | |
|
84 | apply | |
|
85 | legacy: execute, run | |
|
86 | ||
|
87 | data movement | |
|
88 | push, pull, scatter, gather | |
|
89 | ||
|
90 | query methods | |
|
91 | get_result, queue_status, purge_results, result_status | |
|
92 | ||
|
93 | control methods | |
|
94 | abort, shutdown | |
|
95 | ||
|
74 | 96 | """ |
|
75 | 97 | block=Bool(False) |
|
76 | bound=Bool(False) | |
|
77 | track=Bool(False) | |
|
98 | track=Bool(True) | |
|
78 | 99 | history=List() |
|
79 | 100 | outstanding = Set() |
|
80 | 101 | results = Dict() |
|
81 | 102 | client = Instance('IPython.zmq.parallel.client.Client') |
|
82 | 103 | |
|
104 | _socket = Instance('zmq.Socket') | |
|
83 | 105 | _ntargets = Int(1) |
|
84 | _balanced = Bool(False) | |
|
85 | _default_names = List(['block', 'bound', 'track']) | |
|
106 | _flag_names = List(['block', 'track']) | |
|
86 | 107 | _targets = Any() |
|
108 | _idents = Any() | |
|
87 | 109 | |
|
88 | def __init__(self, client=None, targets=None): | |
|
89 | super(View, self).__init__(client=client) | |
|
90 | self._targets = targets | |
|
110 | def __init__(self, client=None, socket=None, targets=None): | |
|
111 | super(View, self).__init__(client=client, _socket=socket) | |
|
91 | 112 | self._ntargets = 1 if isinstance(targets, (int,type(None))) else len(targets) |
|
92 | 113 | self.block = client.block |
|
93 | 114 | |
|
94 | for name in self._default_names: | |
|
115 | self._idents, self._targets = self.client._build_targets(targets) | |
|
116 | if targets is None or isinstance(targets, int): | |
|
117 | self._targets = targets | |
|
118 | for name in self._flag_names: | |
|
119 | # set flags, if they haven't been set yet | |
|
95 | 120 | setattr(self, name, getattr(self, name, None)) |
|
96 | 121 | |
|
97 | 122 | assert not self.__class__ is View, "Don't use base View objects, use subclasses" |
|
98 | 123 | |
|
99 | 124 | |
|
100 | 125 | def __repr__(self): |
|
101 | 126 | strtargets = str(self._targets) |
|
102 | 127 | if len(strtargets) > 16: |
|
103 | 128 | strtargets = strtargets[:12]+'...]' |
|
104 | 129 | return "<%s %s>"%(self.__class__.__name__, strtargets) |
|
105 | 130 | |
|
106 | 131 | @property |
|
107 | 132 | def targets(self): |
|
108 | 133 | return self._targets |
|
109 | 134 | |
|
110 | 135 | @targets.setter |
|
111 | 136 | def targets(self, value): |
|
112 | 137 | raise AttributeError("Cannot set View `targets` after construction!") |
|
113 | 138 | |
|
114 | @property | |
|
115 | def balanced(self): | |
|
116 | return self._balanced | |
|
117 | ||
|
118 | @balanced.setter | |
|
119 | def balanced(self, value): | |
|
120 | raise AttributeError("Cannot set View `balanced` after construction!") | |
|
121 | ||
|
122 | def _defaults(self, *excludes): | |
|
123 | """return dict of our default attributes, excluding names given.""" | |
|
124 | d = dict(balanced=self._balanced, targets=self._targets) | |
|
125 | for name in self._default_names: | |
|
126 | if name not in excludes: | |
|
127 | d[name] = getattr(self, name) | |
|
128 | return d | |
|
129 | ||
|
130 | 139 | def set_flags(self, **kwargs): |
|
131 | 140 | """set my attribute flags by keyword. |
|
132 | 141 | |
|
133 | A View is a wrapper for the Client's apply method, but | |
|
134 | with attributes that specify keyword arguments, those attributes | |
|
135 | can be set by keyword argument with this method. | |
|
142 | Views determine behavior with a few attributes (`block`, `track`, etc.). | |
|
143 | These attributes can be set all at once by name with this method. | |
|
136 | 144 | |
|
137 | 145 | Parameters |
|
138 | 146 | ---------- |
|
139 | 147 | |
|
140 | 148 | block : bool |
|
141 | 149 | whether to wait for results |
|
142 | bound : bool | |
|
143 | whether to pass the client's Namespace as the first argument | |
|
144 | to functions called via `apply`. | |
|
145 | 150 | track : bool |
|
146 | 151 | whether to create a MessageTracker to allow the user to |
|
147 | 152 | safely edit after arrays and buffers during non-copying |
|
148 | 153 | sends. |
|
149 | 154 | """ |
|
150 |
for |
|
|
151 |
if |
|
|
152 |
raise KeyError("Invalid name: %r"% |
|
|
153 | for name in ('block', 'bound'): | |
|
154 | if name in kwargs: | |
|
155 | setattr(self, name, kwargs[name]) | |
|
155 | for name, value in kwargs.iteritems(): | |
|
156 | if name not in self._flag_names: | |
|
157 | raise KeyError("Invalid name: %r"%name) | |
|
158 | else: | |
|
159 | setattr(self, name, value) | |
|
160 | ||
|
161 | @contextmanager | |
|
162 | def temp_flags(self, **kwargs): | |
|
163 | """temporarily set flags, for use in `with` statements. | |
|
164 | ||
|
165 | See set_flags for permanent setting of flags | |
|
166 | ||
|
167 | Examples | |
|
168 | -------- | |
|
169 | ||
|
170 | >>> view.track=False | |
|
171 | ... | |
|
172 | >>> with view.temp_flags(track=True): | |
|
173 | ... ar = view.apply(dostuff, my_big_array) | |
|
174 | ... ar.tracker.wait() # wait for send to finish | |
|
175 | >>> view.track | |
|
176 | False | |
|
177 | ||
|
178 | """ | |
|
179 | # preflight: save flags, and set temporaries | |
|
180 | saved_flags = {} | |
|
181 | for f in self._flag_names: | |
|
182 | saved_flags[f] = getattr(self, f) | |
|
183 | self.set_flags(**kwargs) | |
|
184 | # yield to the with-statement block | |
|
185 | yield | |
|
186 | # postflight: restore saved flags | |
|
187 | self.set_flags(**saved_flags) | |
|
188 | ||
|
156 | 189 | |
|
157 | 190 | #---------------------------------------------------------------- |
|
158 | # wrappers for client methods: | |
|
191 | # apply | |
|
159 | 192 | #---------------------------------------------------------------- |
|
160 | @sync_results | |
|
161 | def spin(self): | |
|
162 | """spin the client, and sync""" | |
|
163 | self.client.spin() | |
|
164 | 193 | |
|
165 | 194 | @sync_results |
|
166 | 195 | @save_ids |
|
196 | def _really_apply(self, f, args, kwargs, block=None, **options): | |
|
197 | """wrapper for client.send_apply_message""" | |
|
198 | raise NotImplementedError("Implement in subclasses") | |
|
199 | ||
|
167 | 200 | def apply(self, f, *args, **kwargs): |
|
168 | 201 | """calls f(*args, **kwargs) on remote engines, returning the result. |
|
169 | 202 | |
|
170 |
This method sets all |
|
|
171 | View's attributes. | |
|
203 | This method sets all apply flags via this View's attributes. | |
|
172 | 204 | |
|
173 | 205 | if self.block is False: |
|
174 | 206 | returns AsyncResult |
|
175 | 207 | else: |
|
176 | 208 | returns actual result of f(*args, **kwargs) |
|
177 | 209 | """ |
|
178 |
return self. |
|
|
210 | return self._really_apply(f, args, kwargs) | |
|
179 | 211 | |
|
180 | @save_ids | |
|
181 | 212 | def apply_async(self, f, *args, **kwargs): |
|
182 | 213 | """calls f(*args, **kwargs) on remote engines in a nonblocking manner. |
|
183 | 214 | |
|
184 | 215 | returns AsyncResult |
|
185 | 216 | """ |
|
186 | d = self._defaults('block', 'bound') | |
|
187 | return self.client.apply(f,args,kwargs, block=False, bound=False, **d) | |
|
217 | return self._really_apply(f, args, kwargs, block=False) | |
|
188 | 218 | |
|
189 | 219 | @spin_after |
|
190 | @save_ids | |
|
191 | 220 | def apply_sync(self, f, *args, **kwargs): |
|
192 | 221 | """calls f(*args, **kwargs) on remote engines in a blocking manner, |
|
193 | 222 | returning the result. |
|
194 | 223 | |
|
195 | 224 | returns: actual result of f(*args, **kwargs) |
|
196 | 225 | """ |
|
197 | d = self._defaults('block', 'bound', 'track') | |
|
198 | return self.client.apply(f,args,kwargs, block=True, bound=False, **d) | |
|
226 | return self._really_apply(f, args, kwargs, block=True) | |
|
199 | 227 | |
|
200 | # @sync_results | |
|
201 | # @save_ids | |
|
202 | # def apply_bound(self, f, *args, **kwargs): | |
|
203 | # """calls f(*args, **kwargs) bound to engine namespace(s). | |
|
204 | # | |
|
205 | # if self.block is False: | |
|
206 | # returns msg_id | |
|
207 | # else: | |
|
208 | # returns actual result of f(*args, **kwargs) | |
|
209 | # | |
|
210 | # This method has access to the targets' namespace via globals() | |
|
211 | # | |
|
212 | # """ | |
|
213 | # d = self._defaults('bound') | |
|
214 | # return self.client.apply(f, args, kwargs, bound=True, **d) | |
|
215 | # | |
|
228 | #---------------------------------------------------------------- | |
|
229 | # wrappers for client and control methods | |
|
230 | #---------------------------------------------------------------- | |
|
216 | 231 | @sync_results |
|
217 | @save_ids | |
|
218 | def apply_async_bound(self, f, *args, **kwargs): | |
|
219 | """calls f(*args, **kwargs) bound to engine namespace(s) | |
|
220 | in a nonblocking manner. | |
|
221 | ||
|
222 | The first argument to `f` will be the Engine's Namespace | |
|
223 | ||
|
224 | returns: AsyncResult | |
|
232 | def spin(self): | |
|
233 | """spin the client, and sync""" | |
|
234 | self.client.spin() | |
|
225 | 235 |
|
|
226 | """ | |
|
227 | d = self._defaults('block', 'bound') | |
|
228 | return self.client.apply(f, args, kwargs, block=False, bound=True, **d) | |
|
236 | @sync_results | |
|
237 | def wait(self, jobs=None, timeout=-1): | |
|
238 | """waits on one or more `jobs`, for up to `timeout` seconds. | |
|
229 | 239 | |
|
230 | @spin_after | |
|
231 | @save_ids | |
|
232 | def apply_sync_bound(self, f, *args, **kwargs): | |
|
233 | """calls f(*args, **kwargs) bound to engine namespace(s), waiting for the result. | |
|
240 | Parameters | |
|
241 | ---------- | |
|
234 | 242 | |
|
235 | The first argument to `f` will be the Engine's Namespace | |
|
243 | jobs : int, str, or list of ints and/or strs, or one or more AsyncResult objects | |
|
244 | ints are indices to self.history | |
|
245 | strs are msg_ids | |
|
246 | default: wait on all outstanding messages | |
|
247 | timeout : float | |
|
248 | a time in seconds, after which to give up. | |
|
249 | default is -1, which means no timeout | |
|
236 | 250 | |
|
237 | returns: actual result of f(*args, **kwargs) | |
|
251 | Returns | |
|
252 | ------- | |
|
238 | 253 | |
|
254 | True : when all msg_ids are done | |
|
255 | False : timeout reached, some msg_ids still outstanding | |
|
239 | 256 | """ |
|
240 | d = self._defaults('block', 'bound') | |
|
241 | return self.client.apply(f, args, kwargs, block=True, bound=True, **d) | |
|
257 | if jobs is None: | |
|
258 | jobs = self.history | |
|
259 | return self.client.wait(jobs, timeout) | |
|
242 | 260 | |
|
243 | 261 | def abort(self, jobs=None, block=None): |
|
244 | 262 | """Abort jobs on my engines. |
|
245 | 263 | |
|
246 | 264 | Parameters |
|
247 | 265 | ---------- |
|
248 | 266 | |
|
249 | 267 | jobs : None, str, list of strs, optional |
|
250 | 268 | if None: abort all jobs. |
|
251 | 269 | else: abort specific msg_id(s). |
|
252 | 270 | """ |
|
253 | 271 | block = block if block is not None else self.block |
|
254 | 272 | return self.client.abort(jobs=jobs, targets=self._targets, block=block) |
|
255 | 273 | |
|
256 | 274 | def queue_status(self, verbose=False): |
|
257 | 275 | """Fetch the Queue status of my engines""" |
|
258 | 276 | return self.client.queue_status(targets=self._targets, verbose=verbose) |
|
259 | 277 | |
|
260 | 278 | def purge_results(self, jobs=[], targets=[]): |
|
261 | 279 | """Instruct the controller to forget specific results.""" |
|
262 | 280 | if targets is None or targets == 'all': |
|
263 | 281 | targets = self._targets |
|
264 | 282 | return self.client.purge_results(jobs=jobs, targets=targets) |
|
265 | 283 | |
|
266 | 284 | @spin_after |
|
267 | 285 | def get_result(self, indices_or_msg_ids=None): |
|
268 | 286 | """return one or more results, specified by history index or msg_id. |
|
269 | 287 | |
|
270 | 288 | See client.get_result for details. |
|
271 | 289 | |
|
272 | 290 | """ |
|
273 | 291 | |
|
274 | 292 | if indices_or_msg_ids is None: |
|
275 | 293 | indices_or_msg_ids = -1 |
|
276 | 294 | if isinstance(indices_or_msg_ids, int): |
|
277 | 295 | indices_or_msg_ids = self.history[indices_or_msg_ids] |
|
278 | 296 | elif isinstance(indices_or_msg_ids, (list,tuple,set)): |
|
279 | 297 | indices_or_msg_ids = list(indices_or_msg_ids) |
|
280 | 298 | for i,index in enumerate(indices_or_msg_ids): |
|
281 | 299 | if isinstance(index, int): |
|
282 | 300 | indices_or_msg_ids[i] = self.history[index] |
|
283 | 301 | return self.client.get_result(indices_or_msg_ids) |
|
284 | 302 | |
|
285 | 303 | #------------------------------------------------------------------- |
|
286 | 304 | # Map |
|
287 | 305 | #------------------------------------------------------------------- |
|
288 | 306 | |
|
289 | 307 | def map(self, f, *sequences, **kwargs): |
|
290 | 308 | """override in subclasses""" |
|
291 | 309 | raise NotImplementedError |
|
292 | 310 | |
|
293 | 311 | def map_async(self, f, *sequences, **kwargs): |
|
294 | 312 | """Parallel version of builtin `map`, using this view's engines. |
|
295 | 313 | |
|
296 | 314 | This is equivalent to map(...block=False) |
|
297 | 315 | |
|
298 | 316 | See `self.map` for details. |
|
299 | 317 | """ |
|
300 | 318 | if 'block' in kwargs: |
|
301 | 319 | raise TypeError("map_async doesn't take a `block` keyword argument.") |
|
302 | 320 | kwargs['block'] = False |
|
303 | 321 | return self.map(f,*sequences,**kwargs) |
|
304 | 322 | |
|
305 | 323 | def map_sync(self, f, *sequences, **kwargs): |
|
306 | 324 | """Parallel version of builtin `map`, using this view's engines. |
|
307 | 325 | |
|
308 | 326 | This is equivalent to map(...block=True) |
|
309 | 327 | |
|
310 | 328 | See `self.map` for details. |
|
311 | 329 | """ |
|
312 | 330 | if 'block' in kwargs: |
|
313 | 331 | raise TypeError("map_sync doesn't take a `block` keyword argument.") |
|
314 | 332 | kwargs['block'] = True |
|
315 | 333 | return self.map(f,*sequences,**kwargs) |
|
316 | 334 | |
|
317 | 335 | def imap(self, f, *sequences, **kwargs): |
|
318 | 336 | """Parallel version of `itertools.imap`. |
|
319 | 337 | |
|
320 | 338 | See `self.map` for details. |
|
339 | ||
|
321 | 340 | """ |
|
322 | 341 | |
|
323 | 342 | return iter(self.map_async(f,*sequences, **kwargs)) |
|
324 | 343 | |
|
325 | 344 | #------------------------------------------------------------------- |
|
326 | 345 | # Decorators |
|
327 | 346 | #------------------------------------------------------------------- |
|
328 | 347 | |
|
329 |
def remote(self, |
|
|
348 | def remote(self, block=True, **flags): | |
|
330 | 349 | """Decorator for making a RemoteFunction""" |
|
331 | return remote(self.client, bound=bound, targets=self._targets, block=block, balanced=self._balanced) | |
|
350 | block = self.block if block is None else block | |
|
351 | return remote(self, block=block, **flags) | |
|
332 | 352 | |
|
333 |
def parallel(self, dist='b', |
|
|
353 | def parallel(self, dist='b', block=None, **flags): | |
|
334 | 354 | """Decorator for making a ParallelFunction""" |
|
335 | 355 | block = self.block if block is None else block |
|
336 |
return parallel(self |
|
|
356 | return parallel(self, dist=dist, block=block, **flags) | |
|
337 | 357 | |
|
338 | 358 | @testdec.skip_doctest |
|
339 | 359 | class DirectView(View): |
|
340 | 360 | """Direct Multiplexer View of one or more engines. |
|
341 | 361 | |
|
342 | 362 | These are created via indexed access to a client: |
|
343 | 363 | |
|
344 | 364 | >>> dv_1 = client[1] |
|
345 | 365 | >>> dv_all = client[:] |
|
346 | 366 | >>> dv_even = client[::2] |
|
347 | 367 | >>> dv_some = client[1:3] |
|
348 | 368 | |
|
349 | 369 | This object provides dictionary access to engine namespaces: |
|
350 | 370 | |
|
351 | 371 | # push a=5: |
|
352 | 372 | >>> dv['a'] = 5 |
|
353 | 373 | # pull 'foo': |
|
354 | 374 | >>> db['foo'] |
|
355 | 375 | |
|
356 | 376 | """ |
|
357 | 377 | |
|
358 | def __init__(self, client=None, targets=None): | |
|
359 | super(DirectView, self).__init__(client=client, targets=targets) | |
|
360 | self._balanced = False | |
|
378 | def __init__(self, client=None, socket=None, targets=None): | |
|
379 | super(DirectView, self).__init__(client=client, socket=socket, targets=targets) | |
|
361 | 380 | |
|
362 | @spin_after | |
|
381 | ||
|
382 | @sync_results | |
|
363 | 383 | @save_ids |
|
384 | def _really_apply(self, f, args=None, kwargs=None, block=None, track=None): | |
|
385 | """calls f(*args, **kwargs) on remote engines, returning the result. | |
|
386 | ||
|
387 | This method sets all of `apply`'s flags via this View's attributes. | |
|
388 | ||
|
389 | Parameters | |
|
390 | ---------- | |
|
391 | ||
|
392 | f : callable | |
|
393 | ||
|
394 | args : list [default: empty] | |
|
395 | ||
|
396 | kwargs : dict [default: empty] | |
|
397 | ||
|
398 | block : bool [default: self.block] | |
|
399 | whether to block | |
|
400 | track : bool [default: self.track] | |
|
401 | whether to ask zmq to track the message, for safe non-copying sends | |
|
402 | ||
|
403 | Returns | |
|
404 | ------- | |
|
405 | ||
|
406 | if self.block is False: | |
|
407 | returns AsyncResult | |
|
408 | else: | |
|
409 | returns actual result of f(*args, **kwargs) on the engine(s) | |
|
410 | This will be a list of self.targets is also a list (even length 1), or | |
|
411 | the single result if self.targets is an integer engine id | |
|
412 | """ | |
|
413 | args = [] if args is None else args | |
|
414 | kwargs = {} if kwargs is None else kwargs | |
|
415 | block = self.block if block is None else block | |
|
416 | track = self.track if track is None else track | |
|
417 | msg_ids = [] | |
|
418 | trackers = [] | |
|
419 | for ident in self._idents: | |
|
420 | msg = self.client.send_apply_message(self._socket, f, args, kwargs, track=track, | |
|
421 | ident=ident) | |
|
422 | if track: | |
|
423 | trackers.append(msg['tracker']) | |
|
424 | msg_ids.append(msg['msg_id']) | |
|
425 | tracker = None if track is False else zmq.MessageTracker(*trackers) | |
|
426 | ar = AsyncResult(self.client, msg_ids, fname=f.__name__, targets=self._targets, tracker=tracker) | |
|
427 | if block: | |
|
428 | try: | |
|
429 | return ar.get() | |
|
430 | except KeyboardInterrupt: | |
|
431 | pass | |
|
432 | return ar | |
|
433 | ||
|
434 | @spin_after | |
|
364 | 435 | def map(self, f, *sequences, **kwargs): |
|
365 |
"""view.map(f, *sequences, block=self.block |
|
|
436 | """view.map(f, *sequences, block=self.block) => list|AsyncMapResult | |
|
366 | 437 | |
|
367 | 438 | Parallel version of builtin `map`, using this View's `targets`. |
|
368 | 439 | |
|
369 | 440 | There will be one task per target, so work will be chunked |
|
370 | 441 | if the sequences are longer than `targets`. |
|
371 | 442 | |
|
372 | 443 | Results can be iterated as they are ready, but will become available in chunks. |
|
373 | 444 | |
|
374 | 445 | Parameters |
|
375 | 446 | ---------- |
|
376 | 447 | |
|
377 | 448 | f : callable |
|
378 | 449 | function to be mapped |
|
379 | 450 | *sequences: one or more sequences of matching length |
|
380 | 451 | the sequences to be distributed and passed to `f` |
|
381 | 452 | block : bool |
|
382 | 453 | whether to wait for the result or not [default self.block] |
|
383 | bound : bool | |
|
384 | whether to pass the client's Namespace as the first argument to `f` | |
|
385 | 454 | |
|
386 | 455 | Returns |
|
387 | 456 | ------- |
|
388 | 457 | |
|
389 | 458 | if block=False: |
|
390 | 459 | AsyncMapResult |
|
391 | 460 | An object like AsyncResult, but which reassembles the sequence of results |
|
392 | 461 | into a single list. AsyncMapResults can be iterated through before all |
|
393 | 462 | results are complete. |
|
394 | 463 | else: |
|
395 | 464 | list |
|
396 | 465 | the result of map(f,*sequences) |
|
397 | 466 | """ |
|
398 | 467 | |
|
399 |
block = kwargs. |
|
|
400 | bound = kwargs.get('bound', self.bound) | |
|
468 | block = kwargs.pop('block', self.block) | |
|
401 | 469 | for k in kwargs.keys(): |
|
402 |
if k not in ['block', ' |
|
|
470 | if k not in ['block', 'track']: | |
|
403 | 471 | raise TypeError("invalid keyword arg, %r"%k) |
|
404 | 472 | |
|
405 | 473 | assert len(sequences) > 0, "must have some sequences to map onto!" |
|
406 |
pf = ParallelFunction(self |
|
|
407 | targets=self._targets, balanced=False) | |
|
474 | pf = ParallelFunction(self, f, block=block, **kwargs) | |
|
408 | 475 | return pf.map(*sequences) |
|
409 | 476 | |
|
410 | @sync_results | |
|
411 | @save_ids | |
|
412 | 477 | def execute(self, code, block=None): |
|
413 | """execute some code on my targets.""" | |
|
478 | """Executes `code` on `targets` in blocking or nonblocking manner. | |
|
414 | 479 |
|
|
415 | block = block if block is not None else self.block | |
|
480 | ``execute`` is always `bound` (affects engine namespace) | |
|
416 | 481 |
|
|
417 | return self.client.execute(code, block=block, targets=self._targets) | |
|
482 | Parameters | |
|
483 | ---------- | |
|
418 | 484 | |
|
419 | @sync_results | |
|
420 | @save_ids | |
|
421 | def run(self, fname, block=None): | |
|
422 | """execute the code in a file on my targets.""" | |
|
485 | code : str | |
|
486 | the code string to be executed | |
|
487 | block : bool | |
|
488 | whether or not to wait until done to return | |
|
489 | default: self.block | |
|
490 | """ | |
|
491 | return self._really_apply(util._execute, args=(code,), block=block) | |
|
423 | 492 |
|
|
424 | block = block if block is not None else self.block | |
|
493 | def run(self, filename, block=None): | |
|
494 | """Execute contents of `filename` on my engine(s). | |
|
495 | ||
|
496 | This simply reads the contents of the file and calls `execute`. | |
|
497 | ||
|
498 | Parameters | |
|
499 | ---------- | |
|
500 | ||
|
501 | filename : str | |
|
502 | The path to the file | |
|
503 | targets : int/str/list of ints/strs | |
|
504 | the engines on which to execute | |
|
505 | default : all | |
|
506 | block : bool | |
|
507 | whether or not to wait until done | |
|
508 | default: self.block | |
|
425 | 509 |
|
|
426 | return self.client.run(fname, block=block, targets=self._targets) | |
|
510 | """ | |
|
511 | with open(filename, 'r') as f: | |
|
512 | # add newline in case of trailing indented whitespace | |
|
513 | # which will cause SyntaxError | |
|
514 | code = f.read()+'\n' | |
|
515 | return self.execute(code, block=block) | |
|
427 | 516 | |
|
428 | 517 | def update(self, ns): |
|
429 |
"""update remote namespace with dict `ns` |
|
|
430 | return self.client.push(ns, targets=self._targets, block=self.block) | |
|
518 | """update remote namespace with dict `ns` | |
|
431 | 519 | |
|
432 | def push(self, ns, block=None): | |
|
433 | """update remote namespace with dict `ns`""" | |
|
520 | See `push` for details. | |
|
521 | """ | |
|
522 | return self.push(ns, block=self.block, track=self.track) | |
|
434 | 523 |
|
|
435 | block = block if block is not None else self.block | |
|
524 | def push(self, ns, block=None, track=None): | |
|
525 | """update remote namespace with dict `ns` | |
|
436 | 526 |
|
|
437 | return self.client.push(ns, targets=self._targets, block=block) | |
|
527 | Parameters | |
|
528 | ---------- | |
|
529 | ||
|
530 | ns : dict | |
|
531 | dict of keys with which to update engine namespace(s) | |
|
532 | block : bool [default : self.block] | |
|
533 | whether to wait to be notified of engine receipt | |
|
534 | ||
|
535 | """ | |
|
536 | ||
|
537 | block = block if block is not None else self.block | |
|
538 | track = track if track is not None else self.track | |
|
539 | # applier = self.apply_sync if block else self.apply_async | |
|
540 | if not isinstance(ns, dict): | |
|
541 | raise TypeError("Must be a dict, not %s"%type(ns)) | |
|
542 | return self._really_apply(util._push, (ns,),block=block, track=track) | |
|
438 | 543 | |
|
439 | 544 | def get(self, key_s): |
|
440 | 545 | """get object(s) by `key_s` from remote namespace |
|
441 | will return one object if it is a key. | |
|
442 | It also takes a list of keys, and will return a list of objects.""" | |
|
546 | ||
|
547 | see `pull` for details. | |
|
548 | """ | |
|
443 | 549 | # block = block if block is not None else self.block |
|
444 |
return self |
|
|
550 | return self.pull(key_s, block=True) | |
|
551 | ||
|
552 | def pull(self, names, block=True): | |
|
553 | """get object(s) by `name` from remote namespace | |
|
445 | 554 | |
|
446 | @sync_results | |
|
447 | @save_ids | |
|
448 | def pull(self, key_s, block=True): | |
|
449 | """get object(s) by `key_s` from remote namespace | |
|
450 | 555 | will return one object if it is a key. |
|
451 |
|
|
|
556 | can also take a list of keys, in which case it will return a list of objects. | |
|
557 | """ | |
|
452 | 558 | block = block if block is not None else self.block |
|
453 | return self.client.pull(key_s, block=block, targets=self._targets) | |
|
559 | applier = self.apply_sync if block else self.apply_async | |
|
560 | if isinstance(names, basestring): | |
|
561 | pass | |
|
562 | elif isinstance(names, (list,tuple,set)): | |
|
563 | for key in names: | |
|
564 | if not isinstance(key, basestring): | |
|
565 | raise TypeError("keys must be str, not type %r"%type(key)) | |
|
566 | else: | |
|
567 | raise TypeError("names must be strs, not %r"%names) | |
|
568 | return applier(util._pull, names) | |
|
454 | 569 | |
|
455 | def scatter(self, key, seq, dist='b', flatten=False, block=None): | |
|
570 | def scatter(self, key, seq, dist='b', flatten=False, block=None, track=None): | |
|
456 | 571 | """ |
|
457 | 572 | Partition a Python sequence and send the partitions to a set of engines. |
|
458 | 573 | """ |
|
459 | 574 | block = block if block is not None else self.block |
|
575 | track = track if track is not None else self.track | |
|
576 | targets = self._targets | |
|
577 | mapObject = Map.dists[dist]() | |
|
578 | nparts = len(targets) | |
|
579 | msg_ids = [] | |
|
580 | trackers = [] | |
|
581 | for index, engineid in enumerate(targets): | |
|
582 | push = self.client[engineid].push | |
|
583 | partition = mapObject.getPartition(seq, index, nparts) | |
|
584 | if flatten and len(partition) == 1: | |
|
585 | r = push({key: partition[0]}, block=False, track=track) | |
|
586 | else: | |
|
587 | r = push({key: partition},block=False, track=track) | |
|
588 | msg_ids.extend(r.msg_ids) | |
|
589 | if track: | |
|
590 | trackers.append(r._tracker) | |
|
460 | 591 | |
|
461 | return self.client.scatter(key, seq, dist=dist, flatten=flatten, | |
|
462 | targets=self._targets, block=block) | |
|
592 | if track: | |
|
593 | tracker = zmq.MessageTracker(*trackers) | |
|
594 | else: | |
|
595 | tracker = None | |
|
596 | ||
|
597 | r = AsyncResult(self.client, msg_ids, fname='scatter', targets=targets, tracker=tracker) | |
|
598 | if block: | |
|
599 | r.wait() | |
|
600 | else: | |
|
601 | return r | |
|
463 | 602 | |
|
464 | 603 | @sync_results |
|
465 | 604 | @save_ids |
|
466 | 605 | def gather(self, key, dist='b', block=None): |
|
467 | 606 | """ |
|
468 | 607 | Gather a partitioned sequence on a set of engines as a single local seq. |
|
469 | 608 | """ |
|
470 | 609 | block = block if block is not None else self.block |
|
610 | mapObject = Map.dists[dist]() | |
|
611 | msg_ids = [] | |
|
612 | for index, engineid in enumerate(self._targets): | |
|
471 | 613 | |
|
472 | return self.client.gather(key, dist=dist, targets=self._targets, block=block) | |
|
614 | msg_ids.extend(self.client[engineid].pull(key, block=False).msg_ids) | |
|
615 | ||
|
616 | r = AsyncMapResult(self.client, msg_ids, mapObject, fname='gather') | |
|
617 | ||
|
618 | if block: | |
|
619 | try: | |
|
620 | return r.get() | |
|
621 | except KeyboardInterrupt: | |
|
622 | pass | |
|
623 | return r | |
|
473 | 624 | |
|
474 | 625 | def __getitem__(self, key): |
|
475 | 626 | return self.get(key) |
|
476 | 627 | |
|
477 | 628 | def __setitem__(self,key, value): |
|
478 | 629 | self.update({key:value}) |
|
479 | 630 | |
|
480 | 631 | def clear(self, block=False): |
|
481 | 632 | """Clear the remote namespaces on my engines.""" |
|
482 | 633 | block = block if block is not None else self.block |
|
483 | 634 | return self.client.clear(targets=self._targets, block=block) |
|
484 | 635 | |
|
485 | 636 | def kill(self, block=True): |
|
486 | 637 | """Kill my engines.""" |
|
487 | 638 | block = block if block is not None else self.block |
|
488 | 639 | return self.client.kill(targets=self._targets, block=block) |
|
489 | 640 | |
|
490 | 641 | #---------------------------------------- |
|
491 | 642 | # activate for %px,%autopx magics |
|
492 | 643 | #---------------------------------------- |
|
493 | 644 | def activate(self): |
|
494 | 645 | """Make this `View` active for parallel magic commands. |
|
495 | 646 | |
|
496 | 647 | IPython has a magic command syntax to work with `MultiEngineClient` objects. |
|
497 | 648 | In a given IPython session there is a single active one. While |
|
498 | 649 | there can be many `Views` created and used by the user, |
|
499 | 650 | there is only one active one. The active `View` is used whenever |
|
500 | 651 | the magic commands %px and %autopx are used. |
|
501 | 652 | |
|
502 | 653 | The activate() method is called on a given `View` to make it |
|
503 | 654 | active. Once this has been done, the magic commands can be used. |
|
504 | 655 | """ |
|
505 | 656 | |
|
506 | 657 | try: |
|
507 | 658 | # This is injected into __builtins__. |
|
508 | 659 | ip = get_ipython() |
|
509 | 660 | except NameError: |
|
510 | 661 | print "The IPython parallel magics (%result, %px, %autopx) only work within IPython." |
|
511 | 662 | else: |
|
512 | 663 | pmagic = ip.plugin_manager.get_plugin('parallelmagic') |
|
513 | 664 | if pmagic is not None: |
|
514 | 665 | pmagic.active_multiengine_client = self |
|
515 | 666 | else: |
|
516 | 667 | print "You must first load the parallelmagic extension " \ |
|
517 | 668 | "by doing '%load_ext parallelmagic'" |
|
518 | 669 | |
|
519 | 670 | |
|
520 | 671 | @testdec.skip_doctest |
|
521 | 672 | class LoadBalancedView(View): |
|
522 | 673 | """An load-balancing View that only executes via the Task scheduler. |
|
523 | 674 | |
|
524 | 675 | Load-balanced views can be created with the client's `view` method: |
|
525 | 676 | |
|
526 |
>>> v = client. |
|
|
677 | >>> v = client.load_balanced_view() | |
|
527 | 678 | |
|
528 | 679 | or targets can be specified, to restrict the potential destinations: |
|
529 | 680 | |
|
530 |
>>> v = client. |
|
|
681 | >>> v = client.client.load_balanced_view(([1,3]) | |
|
531 | 682 | |
|
532 | 683 | which would restrict loadbalancing to between engines 1 and 3. |
|
533 | 684 | |
|
534 | 685 | """ |
|
535 | 686 | |
|
536 |
_ |
|
|
687 | _flag_names = ['block', 'track', 'follow', 'after', 'timeout'] | |
|
537 | 688 | |
|
538 | def __init__(self, client=None, targets=None): | |
|
539 | super(LoadBalancedView, self).__init__(client=client, targets=targets) | |
|
689 | def __init__(self, client=None, socket=None, targets=None): | |
|
690 | super(LoadBalancedView, self).__init__(client=client, socket=socket, targets=targets) | |
|
540 | 691 | self._ntargets = 1 |
|
541 | self._balanced = True | |
|
692 | self._task_scheme=client._task_scheme | |
|
693 | if targets is None: | |
|
694 | self._targets = None | |
|
695 | self._idents=[] | |
|
542 | 696 | |
|
543 | 697 | def _validate_dependency(self, dep): |
|
544 | 698 | """validate a dependency. |
|
545 | 699 | |
|
546 | 700 | For use in `set_flags`. |
|
547 | 701 | """ |
|
548 | 702 | if dep is None or isinstance(dep, (str, AsyncResult, Dependency)): |
|
549 | 703 | return True |
|
550 | 704 | elif isinstance(dep, (list,set, tuple)): |
|
551 | 705 | for d in dep: |
|
552 | if not isinstance(d, str, AsyncResult): | |
|
706 | if not isinstance(d, (str, AsyncResult)): | |
|
553 | 707 | return False |
|
554 | 708 | elif isinstance(dep, dict): |
|
555 | 709 | if set(dep.keys()) != set(Dependency().as_dict().keys()): |
|
556 | 710 | return False |
|
557 | 711 | if not isinstance(dep['msg_ids'], list): |
|
558 | 712 | return False |
|
559 | 713 | for d in dep['msg_ids']: |
|
560 | 714 | if not isinstance(d, str): |
|
561 | 715 | return False |
|
562 | 716 | else: |
|
563 | 717 | return False |
|
564 | 718 |
|
|
719 | return True | |
|
720 | ||
|
721 | def _render_dependency(self, dep): | |
|
722 | """helper for building jsonable dependencies from various input forms.""" | |
|
723 | if isinstance(dep, Dependency): | |
|
724 | return dep.as_dict() | |
|
725 | elif isinstance(dep, AsyncResult): | |
|
726 | return dep.msg_ids | |
|
727 | elif dep is None: | |
|
728 | return [] | |
|
729 | else: | |
|
730 | # pass to Dependency constructor | |
|
731 | return list(Dependency(dep)) | |
|
732 | ||
|
565 | 733 | def set_flags(self, **kwargs): |
|
566 | 734 | """set my attribute flags by keyword. |
|
567 | 735 | |
|
568 | 736 | A View is a wrapper for the Client's apply method, but with attributes |
|
569 | 737 | that specify keyword arguments, those attributes can be set by keyword |
|
570 | 738 | argument with this method. |
|
571 | 739 | |
|
572 | 740 | Parameters |
|
573 | 741 | ---------- |
|
574 | 742 | |
|
575 | 743 | block : bool |
|
576 | 744 | whether to wait for results |
|
577 | bound : bool | |
|
578 | whether to pass the client's Namespace as the first argument | |
|
579 | to functions called via `apply`. | |
|
580 | 745 | track : bool |
|
581 | 746 | whether to create a MessageTracker to allow the user to |
|
582 | 747 | safely edit after arrays and buffers during non-copying |
|
583 | 748 | sends. |
|
584 | follow : Dependency, list, msg_id, AsyncResult | |
|
585 | the location dependencies of tasks | |
|
586 | after : Dependency, list, msg_id, AsyncResult | |
|
587 | the time dependencies of tasks | |
|
588 | timeout : int,None | |
|
589 | the timeout to be used for tasks | |
|
749 | # | |
|
750 | after : Dependency or collection of msg_ids | |
|
751 | Only for load-balanced execution (targets=None) | |
|
752 | Specify a list of msg_ids as a time-based dependency. | |
|
753 | This job will only be run *after* the dependencies | |
|
754 | have been met. | |
|
755 | ||
|
756 | follow : Dependency or collection of msg_ids | |
|
757 | Only for load-balanced execution (targets=None) | |
|
758 | Specify a list of msg_ids as a location-based dependency. | |
|
759 | This job will only be run on an engine where this dependency | |
|
760 | is met. | |
|
761 | ||
|
762 | timeout : float/int or None | |
|
763 | Only for load-balanced execution (targets=None) | |
|
764 | Specify an amount of time (in seconds) for the scheduler to | |
|
765 | wait for dependencies to be met before failing with a | |
|
766 | DependencyTimeout. | |
|
590 | 767 | """ |
|
591 | 768 | |
|
592 | 769 | super(LoadBalancedView, self).set_flags(**kwargs) |
|
593 | 770 | for name in ('follow', 'after'): |
|
594 | 771 | if name in kwargs: |
|
595 | 772 | value = kwargs[name] |
|
596 | 773 | if self._validate_dependency(value): |
|
597 | 774 | setattr(self, name, value) |
|
598 | 775 | else: |
|
599 | 776 | raise ValueError("Invalid dependency: %r"%value) |
|
600 | 777 | if 'timeout' in kwargs: |
|
601 | 778 | t = kwargs['timeout'] |
|
602 | if not isinstance(t, (int, long, float, None)): | |
|
779 | if not isinstance(t, (int, long, float, type(None))): | |
|
603 | 780 | raise TypeError("Invalid type for timeout: %r"%type(t)) |
|
604 | 781 | if t is not None: |
|
605 | 782 | if t < 0: |
|
606 | 783 | raise ValueError("Invalid timeout: %s"%t) |
|
607 | 784 | self.timeout = t |
|
608 | 785 | |
|
786 | @sync_results | |
|
787 | @save_ids | |
|
788 | def _really_apply(self, f, args=None, kwargs=None, block=None, track=None, | |
|
789 | after=None, follow=None, timeout=None): | |
|
790 | """calls f(*args, **kwargs) on a remote engine, returning the result. | |
|
791 | ||
|
792 | This method temporarily sets all of `apply`'s flags for a single call. | |
|
793 | ||
|
794 | Parameters | |
|
795 | ---------- | |
|
796 | ||
|
797 | f : callable | |
|
798 | ||
|
799 | args : list [default: empty] | |
|
800 | ||
|
801 | kwargs : dict [default: empty] | |
|
802 | ||
|
803 | block : bool [default: self.block] | |
|
804 | whether to block | |
|
805 | track : bool [default: self.track] | |
|
806 | whether to ask zmq to track the message, for safe non-copying sends | |
|
807 | ||
|
808 | !!!!!! TODO: THE REST HERE !!!! | |
|
809 | ||
|
810 | Returns | |
|
811 | ------- | |
|
812 | ||
|
813 | if self.block is False: | |
|
814 | returns AsyncResult | |
|
815 | else: | |
|
816 | returns actual result of f(*args, **kwargs) on the engine(s) | |
|
817 | This will be a list of self.targets is also a list (even length 1), or | |
|
818 | the single result if self.targets is an integer engine id | |
|
819 | """ | |
|
820 | ||
|
821 | # validate whether we can run | |
|
822 | if self._socket.closed: | |
|
823 | msg = "Task farming is disabled" | |
|
824 | if self._task_scheme == 'pure': | |
|
825 | msg += " because the pure ZMQ scheduler cannot handle" | |
|
826 | msg += " disappearing engines." | |
|
827 | raise RuntimeError(msg) | |
|
828 | ||
|
829 | if self._task_scheme == 'pure': | |
|
830 | # pure zmq scheme doesn't support dependencies | |
|
831 | msg = "Pure ZMQ scheduler doesn't support dependencies" | |
|
832 | if (follow or after): | |
|
833 | # hard fail on DAG dependencies | |
|
834 | raise RuntimeError(msg) | |
|
835 | if isinstance(f, dependent): | |
|
836 | # soft warn on functional dependencies | |
|
837 | warnings.warn(msg, RuntimeWarning) | |
|
838 | ||
|
839 | # build args | |
|
840 | args = [] if args is None else args | |
|
841 | kwargs = {} if kwargs is None else kwargs | |
|
842 | block = self.block if block is None else block | |
|
843 | track = self.track if track is None else track | |
|
844 | after = self.after if after is None else after | |
|
845 | follow = self.follow if follow is None else follow | |
|
846 | timeout = self.timeout if timeout is None else timeout | |
|
847 | after = self._render_dependency(after) | |
|
848 | follow = self._render_dependency(follow) | |
|
849 | subheader = dict(after=after, follow=follow, timeout=timeout, targets=self._idents) | |
|
850 | ||
|
851 | msg = self.client.send_apply_message(self._socket, f, args, kwargs, track=track, | |
|
852 | subheader=subheader) | |
|
853 | tracker = None if track is False else msg['tracker'] | |
|
854 | ||
|
855 | ar = AsyncResult(self.client, msg['msg_id'], fname=f.__name__, targets=None, tracker=tracker) | |
|
856 | ||
|
857 | if block: | |
|
858 | try: | |
|
859 | return ar.get() | |
|
860 | except KeyboardInterrupt: | |
|
861 | pass | |
|
862 | return ar | |
|
863 | ||
|
609 | 864 | @spin_after |
|
610 | 865 | @save_ids |
|
611 | 866 | def map(self, f, *sequences, **kwargs): |
|
612 |
"""view.map(f, *sequences, block=self.block, |
|
|
867 | """view.map(f, *sequences, block=self.block, chunksize=1) => list|AsyncMapResult | |
|
613 | 868 | |
|
614 | 869 | Parallel version of builtin `map`, load-balanced by this View. |
|
615 | 870 | |
|
616 |
`block`, |
|
|
871 | `block`, and `chunksize` can be specified by keyword only. | |
|
617 | 872 | |
|
618 |
Each `chunk |
|
|
873 | Each `chunksize` elements will be a separate task, and will be | |
|
619 | 874 | load-balanced. This lets individual elements be available for iteration |
|
620 | 875 | as soon as they arrive. |
|
621 | 876 | |
|
622 | 877 | Parameters |
|
623 | 878 | ---------- |
|
624 | 879 | |
|
625 | 880 | f : callable |
|
626 | 881 | function to be mapped |
|
627 | 882 | *sequences: one or more sequences of matching length |
|
628 | 883 | the sequences to be distributed and passed to `f` |
|
629 | 884 | block : bool |
|
630 | 885 | whether to wait for the result or not [default self.block] |
|
631 | bound : bool | |
|
632 | whether to pass the client's Namespace as the first argument to `f` | |
|
633 | 886 | track : bool |
|
634 | 887 | whether to create a MessageTracker to allow the user to |
|
635 | 888 | safely edit after arrays and buffers during non-copying |
|
636 | 889 | sends. |
|
637 |
chunk |
|
|
890 | chunksize : int | |
|
638 | 891 | how many elements should be in each task [default 1] |
|
639 | 892 | |
|
640 | 893 | Returns |
|
641 | 894 | ------- |
|
642 | 895 | |
|
643 | 896 | if block=False: |
|
644 | 897 | AsyncMapResult |
|
645 | 898 | An object like AsyncResult, but which reassembles the sequence of results |
|
646 | 899 | into a single list. AsyncMapResults can be iterated through before all |
|
647 | 900 | results are complete. |
|
648 | 901 | else: |
|
649 | 902 | the result of map(f,*sequences) |
|
650 | 903 | |
|
651 | 904 | """ |
|
652 | 905 | |
|
653 | 906 | # default |
|
654 | 907 | block = kwargs.get('block', self.block) |
|
655 |
|
|
|
656 | chunk_size = kwargs.get('chunk_size', 1) | |
|
908 | chunksize = kwargs.get('chunksize', 1) | |
|
657 | 909 | |
|
658 | 910 | keyset = set(kwargs.keys()) |
|
659 |
extra_keys = keyset.difference_update(set(['block', ' |
|
|
911 | extra_keys = keyset.difference_update(set(['block', 'chunksize'])) | |
|
660 | 912 | if extra_keys: |
|
661 | 913 | raise TypeError("Invalid kwargs: %s"%list(extra_keys)) |
|
662 | 914 | |
|
663 | 915 | assert len(sequences) > 0, "must have some sequences to map onto!" |
|
664 | 916 | |
|
665 |
pf = ParallelFunction(self |
|
|
666 | targets=self._targets, balanced=True, | |
|
667 | chunk_size=chunk_size) | |
|
917 | pf = ParallelFunction(self, f, block=block, chunksize=chunksize) | |
|
668 | 918 | return pf.map(*sequences) |
|
669 | 919 | |
|
670 | 920 | __all__ = ['LoadBalancedView', 'DirectView'] No newline at end of file |
@@ -1,119 +1,120 b'' | |||
|
1 | 1 | """Example for generating an arbitrary DAG as a dependency map. |
|
2 | 2 | |
|
3 | 3 | This demo uses networkx to generate the graph. |
|
4 | 4 | |
|
5 | 5 | Authors |
|
6 | 6 | ------- |
|
7 | 7 | * MinRK |
|
8 | 8 | """ |
|
9 | 9 | import networkx as nx |
|
10 | 10 | from random import randint, random |
|
11 | 11 | from IPython.zmq.parallel import client as cmod |
|
12 | 12 | |
|
13 | 13 | def randomwait(): |
|
14 | 14 | import time |
|
15 | 15 | from random import random |
|
16 | 16 | time.sleep(random()) |
|
17 | 17 | return time.time() |
|
18 | 18 | |
|
19 | 19 | |
|
20 | 20 | def random_dag(nodes, edges): |
|
21 | 21 | """Generate a random Directed Acyclic Graph (DAG) with a given number of nodes and edges.""" |
|
22 | 22 | G = nx.DiGraph() |
|
23 | 23 | for i in range(nodes): |
|
24 | 24 | G.add_node(i) |
|
25 | 25 | while edges > 0: |
|
26 | 26 | a = randint(0,nodes-1) |
|
27 | 27 | b=a |
|
28 | 28 | while b==a: |
|
29 | 29 | b = randint(0,nodes-1) |
|
30 | 30 | G.add_edge(a,b) |
|
31 | 31 | if nx.is_directed_acyclic_graph(G): |
|
32 | 32 | edges -= 1 |
|
33 | 33 | else: |
|
34 | 34 | # we closed a loop! |
|
35 | 35 | G.remove_edge(a,b) |
|
36 | 36 | return G |
|
37 | 37 | |
|
38 | 38 | def add_children(G, parent, level, n=2): |
|
39 | 39 | """Add children recursively to a binary tree.""" |
|
40 | 40 | if level == 0: |
|
41 | 41 | return |
|
42 | 42 | for i in range(n): |
|
43 | 43 | child = parent+str(i) |
|
44 | 44 | G.add_node(child) |
|
45 | 45 | G.add_edge(parent,child) |
|
46 | 46 | add_children(G, child, level-1, n) |
|
47 | 47 | |
|
48 | 48 | def make_bintree(levels): |
|
49 | 49 | """Make a symmetrical binary tree with @levels""" |
|
50 | 50 | G = nx.DiGraph() |
|
51 | 51 | root = '0' |
|
52 | 52 | G.add_node(root) |
|
53 | 53 | add_children(G, root, levels, 2) |
|
54 | 54 | return G |
|
55 | 55 | |
|
56 |
def submit_jobs( |
|
|
56 | def submit_jobs(view, G, jobs): | |
|
57 | 57 | """Submit jobs via client where G describes the time dependencies.""" |
|
58 | 58 | results = {} |
|
59 | 59 | for node in nx.topological_sort(G): |
|
60 |
|
|
|
61 |
results[node] = |
|
|
60 | with view.temp_flags(after=[ results[n] for n in G.predecessors(node) ]): | |
|
61 | results[node] = view.apply(jobs[node]) | |
|
62 | 62 | return results |
|
63 | 63 | |
|
64 | 64 | def validate_tree(G, results): |
|
65 | 65 | """Validate that jobs executed after their dependencies.""" |
|
66 | 66 | for node in G: |
|
67 | 67 | started = results[node].metadata.started |
|
68 | 68 | for parent in G.predecessors(node): |
|
69 | 69 | finished = results[parent].metadata.completed |
|
70 | 70 | assert started > finished, "%s should have happened after %s"%(node, parent) |
|
71 | 71 | |
|
72 | 72 | def main(nodes, edges): |
|
73 | 73 | """Generate a random graph, submit jobs, then validate that the |
|
74 | 74 | dependency order was enforced. |
|
75 | 75 | Finally, plot the graph, with time on the x-axis, and |
|
76 | 76 | in-degree on the y (just for spread). All arrows must |
|
77 | 77 | point at least slightly to the right if the graph is valid. |
|
78 | 78 | """ |
|
79 | import pylab | |
|
79 | from matplotlib import pyplot as plt | |
|
80 | 80 | from matplotlib.dates import date2num |
|
81 | 81 | from matplotlib.cm import gist_rainbow |
|
82 | 82 | print "building DAG" |
|
83 | 83 | G = random_dag(nodes, edges) |
|
84 | 84 | jobs = {} |
|
85 | 85 | pos = {} |
|
86 | 86 | colors = {} |
|
87 | 87 | for node in G: |
|
88 | 88 | jobs[node] = randomwait |
|
89 | 89 | |
|
90 | 90 | client = cmod.Client() |
|
91 | view = client.load_balanced_view() | |
|
91 | 92 | print "submitting %i tasks with %i dependencies"%(nodes,edges) |
|
92 |
results = submit_jobs( |
|
|
93 | results = submit_jobs(view, G, jobs) | |
|
93 | 94 | print "waiting for results" |
|
94 | client.barrier() | |
|
95 | view.wait() | |
|
95 | 96 | print "done" |
|
96 | 97 | for node in G: |
|
97 | 98 | md = results[node].metadata |
|
98 | 99 | start = date2num(md.started) |
|
99 | 100 | runtime = date2num(md.completed) - start |
|
100 | 101 | pos[node] = (start, runtime) |
|
101 | 102 | colors[node] = md.engine_id |
|
102 | 103 | validate_tree(G, results) |
|
103 | 104 | nx.draw(G, pos, node_list=colors.keys(), node_color=colors.values(), cmap=gist_rainbow, |
|
104 | 105 | with_labels=False) |
|
105 | 106 | x,y = zip(*pos.values()) |
|
106 | 107 | xmin,ymin = map(min, (x,y)) |
|
107 | 108 | xmax,ymax = map(max, (x,y)) |
|
108 | 109 | xscale = xmax-xmin |
|
109 | 110 | yscale = ymax-ymin |
|
110 |
p |
|
|
111 |
p |
|
|
111 | plt.xlim(xmin-xscale*.1,xmax+xscale*.1) | |
|
112 | plt.ylim(ymin-yscale*.1,ymax+yscale*.1) | |
|
112 | 113 | return G,results |
|
113 | 114 | |
|
114 | 115 | if __name__ == '__main__': |
|
115 | import pylab | |
|
116 | from matplotlib import pyplot as plt | |
|
116 | 117 | # main(5,10) |
|
117 | 118 | main(32,96) |
|
118 |
p |
|
|
119 | plt.show() | |
|
119 | 120 | No newline at end of file |
@@ -1,118 +1,130 b'' | |||
|
1 | 1 | from IPython.zmq.parallel import error |
|
2 | 2 | from IPython.zmq.parallel.dependency import Dependency |
|
3 | 3 | from IPython.zmq.parallel.client import * |
|
4 | 4 | |
|
5 | 5 | client = Client() |
|
6 | 6 | |
|
7 | 7 | # this will only run on machines that can import numpy: |
|
8 | 8 | @require('numpy') |
|
9 | 9 | def norm(A): |
|
10 | 10 | from numpy.linalg import norm |
|
11 | 11 | return norm(A,2) |
|
12 | 12 | |
|
13 | 13 | def checkpid(pid): |
|
14 | 14 | """return the pid of the engine""" |
|
15 | 15 | import os |
|
16 | 16 | return os.getpid() == pid |
|
17 | 17 | |
|
18 | 18 | def checkhostname(host): |
|
19 | 19 | import socket |
|
20 | 20 | return socket.gethostname() == host |
|
21 | 21 | |
|
22 | 22 | def getpid(): |
|
23 | 23 | import os |
|
24 | 24 | return os.getpid() |
|
25 | 25 | |
|
26 | 26 | pid0 = client[0].apply_sync(getpid) |
|
27 | 27 | |
|
28 | 28 | # this will depend on the pid being that of target 0: |
|
29 | 29 | @depend(checkpid, pid0) |
|
30 | 30 | def getpid2(): |
|
31 | 31 | import os |
|
32 | 32 | return os.getpid() |
|
33 | 33 | |
|
34 |
view = client |
|
|
34 | view = client.load_balanced_view() | |
|
35 | 35 | view.block=True |
|
36 | 36 | |
|
37 | 37 | # will run on anything: |
|
38 | 38 | pids1 = [ view.apply(getpid) for i in range(len(client.ids)) ] |
|
39 | 39 | print pids1 |
|
40 | 40 | # will only run on e0: |
|
41 | 41 | pids2 = [ view.apply(getpid2) for i in range(len(client.ids)) ] |
|
42 | 42 | print pids2 |
|
43 | 43 | |
|
44 | 44 | print "now test some dependency behaviors" |
|
45 | 45 | |
|
46 | 46 | def wait(t): |
|
47 | 47 | import time |
|
48 | 48 | time.sleep(t) |
|
49 | 49 | return t |
|
50 | 50 | |
|
51 | 51 | # fail after some time: |
|
52 | 52 | def wait_and_fail(t): |
|
53 | 53 | import time |
|
54 | 54 | time.sleep(t) |
|
55 | 55 | return 1/0 |
|
56 | 56 | |
|
57 | 57 | successes = [ view.apply_async(wait, 1).msg_ids[0] for i in range(len(client.ids)) ] |
|
58 | 58 | failures = [ view.apply_async(wait_and_fail, 1).msg_ids[0] for i in range(len(client.ids)) ] |
|
59 | 59 | |
|
60 | 60 | mixed = [failures[0],successes[0]] |
|
61 |
d1a = Dependency(mixed, |
|
|
62 |
d1b = Dependency(mixed, |
|
|
63 |
d2a = Dependency(mixed, |
|
|
64 |
d2b = Dependency(mixed, |
|
|
65 |
d3 = Dependency(failures, |
|
|
66 |
d4 = Dependency(failures, |
|
|
67 |
d5 = Dependency(failures, |
|
|
68 |
d6 = Dependency(successes, |
|
|
69 | ||
|
70 |
|
|
|
71 | ||
|
72 | r1a = client.apply(getpid, after=d1a) | |
|
73 |
|
|
|
74 | r2a = client.apply(getpid, after=d2b, follow=d2a) | |
|
75 | r2b = client.apply(getpid, after=d2a, follow=d2b) | |
|
76 | r3 = client.apply(getpid, after=d3) | |
|
77 |
|
|
|
78 | r4b = client.apply(getpid, follow=d4) | |
|
79 | r4c = client.apply(getpid, after=d3, follow=d4) | |
|
80 | r5 = client.apply(getpid, after=d5) | |
|
81 | r5b = client.apply(getpid, follow=d5, after=d3) | |
|
82 | r6 = client.apply(getpid, follow=d6) | |
|
83 | r6b = client.apply(getpid, after=d6, follow=d2b) | |
|
61 | d1a = Dependency(mixed, all=False, failure=True) # yes | |
|
62 | d1b = Dependency(mixed, all=False) # yes | |
|
63 | d2a = Dependency(mixed, all=True, failure=True) # yes after / no follow | |
|
64 | d2b = Dependency(mixed, all=True) # no | |
|
65 | d3 = Dependency(failures, all=False) # no | |
|
66 | d4 = Dependency(failures, all=False, failure=True) # yes | |
|
67 | d5 = Dependency(failures, all=True, failure=True) # yes after / no follow | |
|
68 | d6 = Dependency(successes, all=True, failure=True) # yes after / no follow | |
|
69 | ||
|
70 | view.block = False | |
|
71 | flags = view.temp_flags | |
|
72 | with flags(after=d1a): | |
|
73 | r1a = view.apply(getpid) | |
|
74 | with flags(follow=d1b): | |
|
75 | r1b = view.apply(getpid) | |
|
76 | with flags(after=d2b, follow=d2a): | |
|
77 | r2a = view.apply(getpid) | |
|
78 | with flags(after=d2a, follow=d2b): | |
|
79 | r2b = view.apply(getpid) | |
|
80 | with flags(after=d3): | |
|
81 | r3 = view.apply(getpid) | |
|
82 | with flags(after=d4): | |
|
83 | r4a = view.apply(getpid) | |
|
84 | with flags(follow=d4): | |
|
85 | r4b = view.apply(getpid) | |
|
86 | with flags(after=d3, follow=d4): | |
|
87 | r4c = view.apply(getpid) | |
|
88 | with flags(after=d5): | |
|
89 | r5 = view.apply(getpid) | |
|
90 | with flags(follow=d5, after=d3): | |
|
91 | r5b = view.apply(getpid) | |
|
92 | with flags(follow=d6): | |
|
93 | r6 = view.apply(getpid) | |
|
94 | with flags(after=d6, follow=d2b): | |
|
95 | r6b = view.apply(getpid) | |
|
84 | 96 | |
|
85 | 97 | def should_fail(f): |
|
86 | 98 | try: |
|
87 | 99 | f() |
|
88 | 100 | except error.KernelError: |
|
89 | 101 | pass |
|
90 | 102 | else: |
|
91 | 103 | print 'should have raised' |
|
92 | 104 | # raise Exception("should have raised") |
|
93 | 105 | |
|
94 | 106 | # print r1a.msg_ids |
|
95 | 107 | r1a.get() |
|
96 | 108 | # print r1b.msg_ids |
|
97 | 109 | r1b.get() |
|
98 | 110 | # print r2a.msg_ids |
|
99 | 111 | should_fail(r2a.get) |
|
100 | 112 | # print r2b.msg_ids |
|
101 | 113 | should_fail(r2b.get) |
|
102 | 114 | # print r3.msg_ids |
|
103 | 115 | should_fail(r3.get) |
|
104 | 116 | # print r4a.msg_ids |
|
105 | 117 | r4a.get() |
|
106 | 118 | # print r4b.msg_ids |
|
107 | 119 | r4b.get() |
|
108 | 120 | # print r4c.msg_ids |
|
109 | 121 | should_fail(r4c.get) |
|
110 | 122 | # print r5.msg_ids |
|
111 | 123 | r5.get() |
|
112 | 124 | # print r5b.msg_ids |
|
113 | 125 | should_fail(r5b.get) |
|
114 | 126 | # print r6.msg_ids |
|
115 | 127 | should_fail(r6.get) # assuming > 1 engine |
|
116 | 128 | # print r6b.msg_ids |
|
117 | 129 | should_fail(r6b.get) |
|
118 | 130 | print 'done' |
@@ -1,36 +1,37 b'' | |||
|
1 | 1 | from IPython.zmq.parallel.client import * |
|
2 | 2 | |
|
3 | 3 | client = Client() |
|
4 | view = client[:] | |
|
4 | 5 | |
|
5 |
@remote( |
|
|
6 | @view.remote(block=True) | |
|
6 | 7 | def square(a): |
|
7 | 8 | """return square of a number""" |
|
8 | 9 | return a*a |
|
9 | 10 | |
|
10 | 11 | squares = map(square, range(42)) |
|
11 | 12 | |
|
12 | 13 | # but that blocked between each result; not exactly useful |
|
13 | 14 | |
|
14 | 15 | square.block = False |
|
15 | 16 | |
|
16 | 17 | arlist = map(square, range(42)) |
|
17 | 18 | # submitted very fast |
|
18 | 19 | |
|
19 | 20 | # wait for the results: |
|
20 | 21 | squares2 = [ r.get() for r in arlist ] |
|
21 | 22 | |
|
22 | 23 | # now the more convenient @parallel decorator, which has a map method: |
|
23 | 24 | |
|
24 |
@parallel( |
|
|
25 | @view.parallel(block=False) | |
|
25 | 26 | def psquare(a): |
|
26 | 27 | """return square of a number""" |
|
27 | 28 | return a*a |
|
28 | 29 | |
|
29 | 30 | # this chunks the data into n-negines jobs, not 42 jobs: |
|
30 | 31 | ar = psquare.map(range(42)) |
|
31 | 32 | |
|
32 | 33 | # wait for the results to be done: |
|
33 | 34 | squares3 = ar.get() |
|
34 | 35 | |
|
35 | 36 | print squares == squares2, squares3==squares |
|
36 | 37 | # True No newline at end of file |
@@ -1,15 +1,15 b'' | |||
|
1 | 1 | from IPython.zmq.parallel.client import * |
|
2 | 2 | |
|
3 | 3 | client = Client() |
|
4 | 4 | |
|
5 | 5 | for id in client.ids: |
|
6 |
client.push(dict(ids=id*id) |
|
|
6 | client[id].push(dict(ids=id*id)) | |
|
7 | 7 | |
|
8 |
|
|
|
9 |
|
|
|
8 | v = client[0] | |
|
9 | v['a'] = 5 | |
|
10 | 10 | |
|
11 |
print |
|
|
11 | print v['a'] | |
|
12 | 12 | |
|
13 | 13 | remotes = client[:] |
|
14 | 14 | |
|
15 | 15 | print remotes['ids'] No newline at end of file |
@@ -1,144 +1,144 b'' | |||
|
1 | 1 | #!/usr/bin/env python |
|
2 | 2 | """Run a Monte-Carlo options pricer in parallel.""" |
|
3 | 3 | |
|
4 | 4 | #----------------------------------------------------------------------------- |
|
5 | 5 | # Imports |
|
6 | 6 | #----------------------------------------------------------------------------- |
|
7 | 7 | |
|
8 | 8 | import sys |
|
9 | 9 | import time |
|
10 | 10 | from IPython.zmq.parallel import client |
|
11 | 11 | import numpy as np |
|
12 | 12 | from mcpricer import price_options |
|
13 | 13 | from matplotlib import pyplot as plt |
|
14 | 14 | |
|
15 | 15 | #----------------------------------------------------------------------------- |
|
16 | 16 | # Setup parameters for the run |
|
17 | 17 | #----------------------------------------------------------------------------- |
|
18 | 18 | |
|
19 | 19 | def ask_question(text, the_type, default): |
|
20 | 20 | s = '%s [%r]: ' % (text, the_type(default)) |
|
21 | 21 | result = raw_input(s) |
|
22 | 22 | if result: |
|
23 | 23 | return the_type(result) |
|
24 | 24 | else: |
|
25 | 25 | return the_type(default) |
|
26 | 26 | |
|
27 | 27 | cluster_profile = ask_question("Cluster profile", str, "default") |
|
28 | 28 | price = ask_question("Initial price", float, 100.0) |
|
29 | 29 | rate = ask_question("Interest rate", float, 0.05) |
|
30 | 30 | days = ask_question("Days to expiration", int, 260) |
|
31 | 31 | paths = ask_question("Number of MC paths", int, 10000) |
|
32 | 32 | n_strikes = ask_question("Number of strike values", int, 5) |
|
33 | 33 | min_strike = ask_question("Min strike price", float, 90.0) |
|
34 | 34 | max_strike = ask_question("Max strike price", float, 110.0) |
|
35 | 35 | n_sigmas = ask_question("Number of volatility values", int, 5) |
|
36 | 36 | min_sigma = ask_question("Min volatility", float, 0.1) |
|
37 | 37 | max_sigma = ask_question("Max volatility", float, 0.4) |
|
38 | 38 | |
|
39 | 39 | strike_vals = np.linspace(min_strike, max_strike, n_strikes) |
|
40 | 40 | sigma_vals = np.linspace(min_sigma, max_sigma, n_sigmas) |
|
41 | 41 | |
|
42 | 42 | #----------------------------------------------------------------------------- |
|
43 | 43 | # Setup for parallel calculation |
|
44 | 44 | #----------------------------------------------------------------------------- |
|
45 | 45 | |
|
46 | 46 | # The Client is used to setup the calculation and works with all |
|
47 | 47 | # engines. |
|
48 | 48 | c = client.Client(profile=cluster_profile) |
|
49 | 49 | |
|
50 | 50 | # A LoadBalancedView is an interface to the engines that provides dynamic load |
|
51 | 51 | # balancing at the expense of not knowing which engine will execute the code. |
|
52 | view = c.view() | |
|
52 | view = c.load_balanced_view() | |
|
53 | 53 | |
|
54 | 54 | # Initialize the common code on the engines. This Python module has the |
|
55 | 55 | # price_options function that prices the options. |
|
56 | 56 | |
|
57 | 57 | #----------------------------------------------------------------------------- |
|
58 | 58 | # Perform parallel calculation |
|
59 | 59 | #----------------------------------------------------------------------------- |
|
60 | 60 | |
|
61 | 61 | print "Running parallel calculation over strike prices and volatilities..." |
|
62 | 62 | print "Strike prices: ", strike_vals |
|
63 | 63 | print "Volatilities: ", sigma_vals |
|
64 | 64 | sys.stdout.flush() |
|
65 | 65 | |
|
66 | 66 | # Submit tasks to the TaskClient for each (strike, sigma) pair as a MapTask. |
|
67 | 67 | t1 = time.time() |
|
68 | 68 | async_results = [] |
|
69 | 69 | for strike in strike_vals: |
|
70 | 70 | for sigma in sigma_vals: |
|
71 | 71 | ar = view.apply_async(price_options, price, strike, sigma, rate, days, paths) |
|
72 | 72 | async_results.append(ar) |
|
73 | 73 | |
|
74 | 74 | print "Submitted tasks: ", len(async_results) |
|
75 | 75 | sys.stdout.flush() |
|
76 | 76 | |
|
77 | 77 | # Block until all tasks are completed. |
|
78 |
c. |
|
|
78 | c.wait(async_results) | |
|
79 | 79 | t2 = time.time() |
|
80 | 80 | t = t2-t1 |
|
81 | 81 | |
|
82 | 82 | print "Parallel calculation completed, time = %s s" % t |
|
83 | 83 | print "Collecting results..." |
|
84 | 84 | |
|
85 | 85 | # Get the results using TaskClient.get_task_result. |
|
86 | 86 | results = [ar.get() for ar in async_results] |
|
87 | 87 | |
|
88 | 88 | # Assemble the result into a structured NumPy array. |
|
89 | 89 | prices = np.empty(n_strikes*n_sigmas, |
|
90 | 90 | dtype=[('ecall',float),('eput',float),('acall',float),('aput',float)] |
|
91 | 91 | ) |
|
92 | 92 | |
|
93 | 93 | for i, price in enumerate(results): |
|
94 | 94 | prices[i] = tuple(price) |
|
95 | 95 | |
|
96 | 96 | prices.shape = (n_strikes, n_sigmas) |
|
97 | 97 | strike_mesh, sigma_mesh = np.meshgrid(strike_vals, sigma_vals) |
|
98 | 98 | |
|
99 | 99 | print "Results are available: strike_mesh, sigma_mesh, prices" |
|
100 | 100 | print "To plot results type 'plot_options(sigma_mesh, strike_mesh, prices)'" |
|
101 | 101 | |
|
102 | 102 | #----------------------------------------------------------------------------- |
|
103 | 103 | # Utilities |
|
104 | 104 | #----------------------------------------------------------------------------- |
|
105 | 105 | |
|
106 | 106 | def plot_options(sigma_mesh, strike_mesh, prices): |
|
107 | 107 | """ |
|
108 | 108 | Make a contour plot of the option price in (sigma, strike) space. |
|
109 | 109 | """ |
|
110 | 110 | plt.figure(1) |
|
111 | 111 | |
|
112 | 112 | plt.subplot(221) |
|
113 | 113 | plt.contourf(sigma_mesh, strike_mesh, prices['ecall']) |
|
114 | 114 | plt.axis('tight') |
|
115 | 115 | plt.colorbar() |
|
116 | 116 | plt.title('European Call') |
|
117 | 117 | plt.ylabel("Strike Price") |
|
118 | 118 | |
|
119 | 119 | plt.subplot(222) |
|
120 | 120 | plt.contourf(sigma_mesh, strike_mesh, prices['acall']) |
|
121 | 121 | plt.axis('tight') |
|
122 | 122 | plt.colorbar() |
|
123 | 123 | plt.title("Asian Call") |
|
124 | 124 | |
|
125 | 125 | plt.subplot(223) |
|
126 | 126 | plt.contourf(sigma_mesh, strike_mesh, prices['eput']) |
|
127 | 127 | plt.axis('tight') |
|
128 | 128 | plt.colorbar() |
|
129 | 129 | plt.title("European Put") |
|
130 | 130 | plt.xlabel("Volatility") |
|
131 | 131 | plt.ylabel("Strike Price") |
|
132 | 132 | |
|
133 | 133 | plt.subplot(224) |
|
134 | 134 | plt.contourf(sigma_mesh, strike_mesh, prices['aput']) |
|
135 | 135 | plt.axis('tight') |
|
136 | 136 | plt.colorbar() |
|
137 | 137 | plt.title("Asian Put") |
|
138 | 138 | plt.xlabel("Volatility") |
|
139 | 139 | |
|
140 | 140 | |
|
141 | 141 | |
|
142 | 142 | |
|
143 | 143 | |
|
144 | 144 |
@@ -1,64 +1,64 b'' | |||
|
1 | 1 | """Calculate statistics on the digits of pi in parallel. |
|
2 | 2 | |
|
3 | 3 | This program uses the functions in :file:`pidigits.py` to calculate |
|
4 | 4 | the frequencies of 2 digit sequences in the digits of pi. The |
|
5 | 5 | results are plotted using matplotlib. |
|
6 | 6 | |
|
7 | 7 | To run, text files from http://www.super-computing.org/ |
|
8 | 8 | must be installed in the working directory of the IPython engines. |
|
9 | 9 | The actual filenames to be used can be set with the ``filestring`` |
|
10 | 10 | variable below. |
|
11 | 11 | |
|
12 | 12 | The dataset we have been using for this is the 200 million digit one here: |
|
13 | 13 | ftp://pi.super-computing.org/.2/pi200m/ |
|
14 | 14 | |
|
15 | 15 | and the files used will be downloaded if they are not in the working directory |
|
16 | 16 | of the IPython engines. |
|
17 | 17 | """ |
|
18 | 18 | |
|
19 | 19 | from IPython.zmq.parallel import client |
|
20 | 20 | from matplotlib import pyplot as plt |
|
21 | 21 | import numpy as np |
|
22 | 22 | from pidigits import * |
|
23 | 23 | from timeit import default_timer as clock |
|
24 | 24 | |
|
25 | 25 | # Files with digits of pi (10m digits each) |
|
26 | 26 | filestring = 'pi200m.ascii.%(i)02dof20' |
|
27 | 27 | files = [filestring % {'i':i} for i in range(1,16)] |
|
28 | 28 | |
|
29 | 29 | # Connect to the IPython cluster |
|
30 |
c = client.Client( |
|
|
31 | c.run('pidigits.py') | |
|
30 | c = client.Client() | |
|
31 | c[:].run('pidigits.py') | |
|
32 | 32 | |
|
33 | 33 | # the number of engines |
|
34 | 34 | n = len(c) |
|
35 | 35 | id0 = c.ids[0] |
|
36 | 36 | v = c[:] |
|
37 | v.set_flags(bound=True,block=True) | |
|
37 | v.block=True | |
|
38 | 38 | # fetch the pi-files |
|
39 | 39 | print "downloading %i files of pi"%n |
|
40 | 40 | v.map(fetch_pi_file, files[:n]) |
|
41 | 41 | print "done" |
|
42 | 42 | |
|
43 | 43 | # Run 10m digits on 1 engine |
|
44 | 44 | t1 = clock() |
|
45 | 45 | freqs10m = c[id0].apply_sync(compute_two_digit_freqs, files[0]) |
|
46 | 46 | t2 = clock() |
|
47 | 47 | digits_per_second1 = 10.0e6/(t2-t1) |
|
48 | 48 | print "Digits per second (1 core, 10m digits): ", digits_per_second1 |
|
49 | 49 | |
|
50 | 50 | |
|
51 | 51 | # Run n*10m digits on all engines |
|
52 | 52 | t1 = clock() |
|
53 | 53 | freqs_all = v.map(compute_two_digit_freqs, files[:n]) |
|
54 | 54 | freqs150m = reduce_freqs(freqs_all) |
|
55 | 55 | t2 = clock() |
|
56 | 56 | digits_per_second8 = n*10.0e6/(t2-t1) |
|
57 | 57 | print "Digits per second (%i engines, %i0m digits): "%(n,n), digits_per_second8 |
|
58 | 58 | |
|
59 | 59 | print "Speedup: ", digits_per_second8/digits_per_second1 |
|
60 | 60 | |
|
61 | 61 | plot_two_digit_freqs(freqs150m) |
|
62 | 62 | plt.title("2 digit sequences in %i0m digits of pi"%n) |
|
63 | 63 | plt.show() |
|
64 | 64 |
@@ -1,203 +1,205 b'' | |||
|
1 | 1 | #!/usr/bin/env python |
|
2 | 2 | """ |
|
3 | 3 | A simple python program of solving a 2D wave equation in parallel. |
|
4 | 4 | Domain partitioning and inter-processor communication |
|
5 | 5 | are done by an object of class MPIRectPartitioner2D |
|
6 | 6 | (which is a subclass of RectPartitioner2D and uses MPI via mpi4py) |
|
7 | 7 | |
|
8 | 8 | An example of running the program is (8 processors, 4x2 partition, |
|
9 | 9 | 400x100 grid cells):: |
|
10 | 10 | |
|
11 | 11 | $ ipclusterz start --profile mpi -n 8 # start 8 engines (assuming mpi profile has been configured) |
|
12 | 12 | $ ./parallelwave-mpi.py --grid 400 100 --partition 4 2 --profile mpi |
|
13 | 13 | |
|
14 | 14 | See also parallelwave-mpi, which runs the same program, but uses MPI |
|
15 | 15 | (via mpi4py) for the inter-engine communication. |
|
16 | 16 | |
|
17 | 17 | Authors |
|
18 | 18 | ------- |
|
19 | 19 | |
|
20 | 20 | * Xing Cai |
|
21 | 21 | * Min Ragan-Kelley |
|
22 | 22 | |
|
23 | 23 | """ |
|
24 | 24 | |
|
25 | 25 | import sys |
|
26 | 26 | import time |
|
27 | 27 | |
|
28 | 28 | from numpy import exp, zeros, newaxis, sqrt |
|
29 | 29 | |
|
30 | 30 | from IPython.external import argparse |
|
31 | 31 | from IPython.zmq.parallel.client import Client, Reference |
|
32 | 32 | |
|
33 |
def setup_partitioner( |
|
|
33 | def setup_partitioner(index, num_procs, gnum_cells, parts): | |
|
34 | 34 | """create a partitioner in the engine namespace""" |
|
35 | global partitioner | |
|
35 | 36 | p = MPIRectPartitioner2D(my_id=index, num_procs=num_procs) |
|
36 | 37 | p.redim(global_num_cells=gnum_cells, num_parts=parts) |
|
37 | 38 | p.prepare_communication() |
|
38 | 39 | # put the partitioner into the global namespace: |
|
39 |
|
|
|
40 | partitioner=p | |
|
40 | 41 | |
|
41 |
def setup_solver( |
|
|
42 | def setup_solver(*args, **kwargs): | |
|
42 | 43 | """create a WaveSolver in the engine namespace""" |
|
43 | ns.solver = WaveSolver(*args, **kwargs) | |
|
44 | global solver | |
|
45 | solver = WaveSolver(*args, **kwargs) | |
|
44 | 46 | |
|
45 | 47 | def wave_saver(u, x, y, t): |
|
46 | 48 | """save the wave log""" |
|
47 | 49 | global u_hist |
|
48 | 50 | global t_hist |
|
49 | 51 | t_hist.append(t) |
|
50 | 52 | u_hist.append(1.0*u) |
|
51 | 53 | |
|
52 | 54 | |
|
53 | 55 | # main program: |
|
54 | 56 | if __name__ == '__main__': |
|
55 | 57 | |
|
56 | 58 | parser = argparse.ArgumentParser() |
|
57 | 59 | paa = parser.add_argument |
|
58 | 60 | paa('--grid', '-g', |
|
59 | 61 | type=int, nargs=2, default=[100,100], dest='grid', |
|
60 | 62 | help="Cells in the grid, e.g. --grid 100 200") |
|
61 | 63 | paa('--partition', '-p', |
|
62 | 64 | type=int, nargs=2, default=None, |
|
63 | 65 | help="Process partition grid, e.g. --partition 4 2 for 4x2") |
|
64 | 66 | paa('-c', |
|
65 | 67 | type=float, default=1., |
|
66 | 68 | help="Wave speed (I think)") |
|
67 | 69 | paa('-Ly', |
|
68 | 70 | type=float, default=1., |
|
69 | 71 | help="system size (in y)") |
|
70 | 72 | paa('-Lx', |
|
71 | 73 | type=float, default=1., |
|
72 | 74 | help="system size (in x)") |
|
73 | 75 | paa('-t', '--tstop', |
|
74 | 76 | type=float, default=1., |
|
75 | 77 | help="Time units to run") |
|
76 | 78 | paa('--profile', |
|
77 | 79 | type=unicode, default=u'default', |
|
78 | 80 | help="Specify the ipcluster profile for the client to connect to.") |
|
79 | 81 | paa('--save', |
|
80 | 82 | action='store_true', |
|
81 | 83 | help="Add this flag to save the time/wave history during the run.") |
|
82 | 84 | paa('--scalar', |
|
83 | 85 | action='store_true', |
|
84 | 86 | help="Also run with scalar interior implementation, to see vector speedup.") |
|
85 | 87 | |
|
86 | 88 | ns = parser.parse_args() |
|
87 | 89 | # set up arguments |
|
88 | 90 | grid = ns.grid |
|
89 | 91 | partition = ns.partition |
|
90 | 92 | Lx = ns.Lx |
|
91 | 93 | Ly = ns.Ly |
|
92 | 94 | c = ns.c |
|
93 | 95 | tstop = ns.tstop |
|
94 | 96 | if ns.save: |
|
95 | 97 | user_action = wave_saver |
|
96 | 98 | else: |
|
97 | 99 | user_action = None |
|
98 | 100 | |
|
99 | 101 | num_cells = 1.0*(grid[0]-1)*(grid[1]-1) |
|
100 | 102 | final_test = True |
|
101 | 103 | |
|
102 | 104 | # create the Client |
|
103 | 105 | rc = Client(profile=ns.profile) |
|
104 | 106 | num_procs = len(rc.ids) |
|
105 | 107 | |
|
106 | 108 | if partition is None: |
|
107 | 109 | partition = [1,num_procs] |
|
108 | 110 | |
|
109 | 111 | assert partition[0]*partition[1] == num_procs, "can't map partition %s to %i engines"%(partition, num_procs) |
|
110 | 112 | |
|
111 | 113 | view = rc[:] |
|
112 | 114 | print "Running %s system on %s processes until %f"%(grid, partition, tstop) |
|
113 | 115 | |
|
114 | 116 | # functions defining initial/boundary/source conditions |
|
115 | 117 | def I(x,y): |
|
116 | 118 | from numpy import exp |
|
117 | 119 | return 1.5*exp(-100*((x-0.5)**2+(y-0.5)**2)) |
|
118 | 120 | def f(x,y,t): |
|
119 | 121 | return 0.0 |
|
120 | 122 | # from numpy import exp,sin |
|
121 | 123 | # return 10*exp(-(x - sin(100*t))**2) |
|
122 | 124 | def bc(x,y,t): |
|
123 | 125 | return 0.0 |
|
124 | 126 | |
|
125 | 127 | # initial imports, setup rank |
|
126 | 128 | view.execute('\n'.join([ |
|
127 | 129 | "from mpi4py import MPI", |
|
128 | 130 | "import numpy", |
|
129 | 131 | "mpi = MPI.COMM_WORLD", |
|
130 | 132 | "my_id = MPI.COMM_WORLD.Get_rank()"]), block=True) |
|
131 | 133 | |
|
132 | 134 | # initialize t_hist/u_hist for saving the state at each step (optional) |
|
133 | 135 | view['t_hist'] = [] |
|
134 | 136 | view['u_hist'] = [] |
|
135 | 137 | |
|
136 | 138 | # set vector/scalar implementation details |
|
137 | 139 | impl = {} |
|
138 | 140 | impl['ic'] = 'vectorized' |
|
139 | 141 | impl['inner'] = 'scalar' |
|
140 | 142 | impl['bc'] = 'vectorized' |
|
141 | 143 | |
|
142 | 144 | # execute some files so that the classes we need will be defined on the engines: |
|
143 | 145 | view.run('RectPartitioner.py') |
|
144 | 146 | view.run('wavesolver.py') |
|
145 | 147 | |
|
146 | 148 | # setup remote partitioner |
|
147 | 149 | # note that Reference means that the argument passed to setup_partitioner will be the |
|
148 | 150 | # object named 'my_id' in the engine's namespace |
|
149 |
view.apply_sync |
|
|
151 | view.apply_sync(setup_partitioner, Reference('my_id'), num_procs, grid, partition) | |
|
150 | 152 | # wait for initial communication to complete |
|
151 | 153 | view.execute('mpi.barrier()') |
|
152 | 154 | # setup remote solvers |
|
153 |
view.apply_sync |
|
|
155 | view.apply_sync(setup_solver, I,f,c,bc,Lx,Ly,partitioner=Reference('partitioner'), dt=0,implementation=impl) | |
|
154 | 156 | |
|
155 | 157 | # lambda for calling solver.solve: |
|
156 | 158 | _solve = lambda *args, **kwargs: solver.solve(*args, **kwargs) |
|
157 | 159 | |
|
158 | 160 | if ns.scalar: |
|
159 | 161 | impl['inner'] = 'scalar' |
|
160 | 162 | # run first with element-wise Python operations for each cell |
|
161 | 163 | t0 = time.time() |
|
162 | 164 | ar = view.apply_async(_solve, tstop, dt=0, verbose=True, final_test=final_test, user_action=user_action) |
|
163 | 165 | if final_test: |
|
164 | 166 | # this sum is performed element-wise as results finish |
|
165 | 167 | s = sum(ar) |
|
166 | 168 | # the L2 norm (RMS) of the result: |
|
167 | 169 | norm = sqrt(s/num_cells) |
|
168 | 170 | else: |
|
169 | 171 | norm = -1 |
|
170 | 172 | t1 = time.time() |
|
171 | 173 | print 'scalar inner-version, Wtime=%g, norm=%g'%(t1-t0, norm) |
|
172 | 174 | |
|
173 | 175 | impl['inner'] = 'vectorized' |
|
174 | 176 | # setup new solvers |
|
175 |
view.apply_sync |
|
|
177 | view.apply_sync(setup_solver, I,f,c,bc,Lx,Ly,partitioner=Reference('partitioner'), dt=0,implementation=impl) | |
|
176 | 178 | view.execute('mpi.barrier()') |
|
177 | 179 | |
|
178 | 180 | # run again with numpy vectorized inner-implementation |
|
179 | 181 | t0 = time.time() |
|
180 | 182 | ar = view.apply_async(_solve, tstop, dt=0, verbose=True, final_test=final_test)#, user_action=wave_saver) |
|
181 | 183 | if final_test: |
|
182 | 184 | # this sum is performed element-wise as results finish |
|
183 | 185 | s = sum(ar) |
|
184 | 186 | # the L2 norm (RMS) of the result: |
|
185 | 187 | norm = sqrt(s/num_cells) |
|
186 | 188 | else: |
|
187 | 189 | norm = -1 |
|
188 | 190 | t1 = time.time() |
|
189 | 191 | print 'vector inner-version, Wtime=%g, norm=%g'%(t1-t0, norm) |
|
190 | 192 | |
|
191 | 193 | # if ns.save is True, then u_hist stores the history of u as a list |
|
192 | 194 | # If the partion scheme is Nx1, then u can be reconstructed via 'gather': |
|
193 | 195 | if ns.save and partition[-1] == 1: |
|
194 | 196 | import pylab |
|
195 | 197 | view.execute('u_last=u_hist[-1]') |
|
196 | 198 | # map mpi IDs to IPython IDs, which may not match |
|
197 | 199 | ranks = view['my_id'] |
|
198 | 200 | targets = range(len(ranks)) |
|
199 | 201 | for idx in range(len(ranks)): |
|
200 | 202 | targets[idx] = ranks.index(idx) |
|
201 | 203 | u_last = rc[targets].gather('u_last', block=True) |
|
202 | 204 | pylab.pcolor(u_last) |
|
203 | 205 | pylab.show() |
@@ -1,207 +1,209 b'' | |||
|
1 | 1 | #!/usr/bin/env python |
|
2 | 2 | """ |
|
3 | 3 | A simple python program of solving a 2D wave equation in parallel. |
|
4 | 4 | Domain partitioning and inter-processor communication |
|
5 | 5 | are done by an object of class ZMQRectPartitioner2D |
|
6 | 6 | (which is a subclass of RectPartitioner2D and uses 0MQ via pyzmq) |
|
7 | 7 | |
|
8 | 8 | An example of running the program is (8 processors, 4x2 partition, |
|
9 | 9 | 200x200 grid cells):: |
|
10 | 10 | |
|
11 | 11 | $ ipclusterz start -n 8 # start 8 engines |
|
12 | 12 | $ ./parallelwave.py --grid 200 200 --partition 4 2 |
|
13 | 13 | |
|
14 | 14 | See also parallelwave-mpi, which runs the same program, but uses MPI |
|
15 | 15 | (via mpi4py) for the inter-engine communication. |
|
16 | 16 | |
|
17 | 17 | Authors |
|
18 | 18 | ------- |
|
19 | 19 | |
|
20 | 20 | * Xing Cai |
|
21 | 21 | * Min Ragan-Kelley |
|
22 | 22 | |
|
23 | 23 | """ |
|
24 | 24 | # |
|
25 | 25 | import sys |
|
26 | 26 | import time |
|
27 | 27 | |
|
28 | 28 | from numpy import exp, zeros, newaxis, sqrt |
|
29 | 29 | |
|
30 | 30 | from IPython.external import argparse |
|
31 | 31 | from IPython.zmq.parallel.client import Client, Reference |
|
32 | 32 | |
|
33 |
def setup_partitioner( |
|
|
33 | def setup_partitioner(comm, addrs, index, num_procs, gnum_cells, parts): | |
|
34 | 34 | """create a partitioner in the engine namespace""" |
|
35 | global partitioner | |
|
35 | 36 | p = ZMQRectPartitioner2D(comm, addrs, my_id=index, num_procs=num_procs) |
|
36 | 37 | p.redim(global_num_cells=gnum_cells, num_parts=parts) |
|
37 | 38 | p.prepare_communication() |
|
38 | 39 | # put the partitioner into the global namespace: |
|
39 |
|
|
|
40 | partitioner=p | |
|
40 | 41 | |
|
41 |
def setup_solver( |
|
|
42 | def setup_solver(*args, **kwargs): | |
|
42 | 43 | """create a WaveSolver in the engine namespace.""" |
|
43 | ns.solver = WaveSolver(*args, **kwargs) | |
|
44 | global solver | |
|
45 | solver = WaveSolver(*args, **kwargs) | |
|
44 | 46 | |
|
45 | 47 | def wave_saver(u, x, y, t): |
|
46 | 48 | """save the wave state for each timestep.""" |
|
47 | 49 | global u_hist |
|
48 | 50 | global t_hist |
|
49 | 51 | t_hist.append(t) |
|
50 | 52 | u_hist.append(1.0*u) |
|
51 | 53 | |
|
52 | 54 | |
|
53 | 55 | # main program: |
|
54 | 56 | if __name__ == '__main__': |
|
55 | 57 | |
|
56 | 58 | parser = argparse.ArgumentParser() |
|
57 | 59 | paa = parser.add_argument |
|
58 | 60 | paa('--grid', '-g', |
|
59 | 61 | type=int, nargs=2, default=[100,100], dest='grid', |
|
60 | 62 | help="Cells in the grid, e.g. --grid 100 200") |
|
61 | 63 | paa('--partition', '-p', |
|
62 | 64 | type=int, nargs=2, default=None, |
|
63 | 65 | help="Process partition grid, e.g. --partition 4 2 for 4x2") |
|
64 | 66 | paa('-c', |
|
65 | 67 | type=float, default=1., |
|
66 | 68 | help="Wave speed (I think)") |
|
67 | 69 | paa('-Ly', |
|
68 | 70 | type=float, default=1., |
|
69 | 71 | help="system size (in y)") |
|
70 | 72 | paa('-Lx', |
|
71 | 73 | type=float, default=1., |
|
72 | 74 | help="system size (in x)") |
|
73 | 75 | paa('-t', '--tstop', |
|
74 | 76 | type=float, default=1., |
|
75 | 77 | help="Time units to run") |
|
76 | 78 | paa('--profile', |
|
77 | 79 | type=unicode, default=u'default', |
|
78 | 80 | help="Specify the ipcluster profile for the client to connect to.") |
|
79 | 81 | paa('--save', |
|
80 | 82 | action='store_true', |
|
81 | 83 | help="Add this flag to save the time/wave history during the run.") |
|
82 | 84 | paa('--scalar', |
|
83 | 85 | action='store_true', |
|
84 | 86 | help="Also run with scalar interior implementation, to see vector speedup.") |
|
85 | 87 | |
|
86 | 88 | ns = parser.parse_args() |
|
87 | 89 | # set up arguments |
|
88 | 90 | grid = ns.grid |
|
89 | 91 | partition = ns.partition |
|
90 | 92 | Lx = ns.Lx |
|
91 | 93 | Ly = ns.Ly |
|
92 | 94 | c = ns.c |
|
93 | 95 | tstop = ns.tstop |
|
94 | 96 | if ns.save: |
|
95 | 97 | user_action = wave_saver |
|
96 | 98 | else: |
|
97 | 99 | user_action = None |
|
98 | 100 | |
|
99 | 101 | num_cells = 1.0*(grid[0]-1)*(grid[1]-1) |
|
100 | 102 | final_test = True |
|
101 | 103 | |
|
102 | 104 | # create the Client |
|
103 | 105 | rc = Client(profile=ns.profile) |
|
104 | 106 | num_procs = len(rc.ids) |
|
105 | 107 | |
|
106 | 108 | if partition is None: |
|
107 | 109 | partition = [num_procs,1] |
|
108 | 110 | else: |
|
109 | 111 | num_procs = min(num_procs, partition[0]*partition[1]) |
|
110 | 112 | |
|
111 | 113 | assert partition[0]*partition[1] == num_procs, "can't map partition %s to %i engines"%(partition, num_procs) |
|
112 | 114 | |
|
113 | 115 | # construct the View: |
|
114 | 116 | view = rc[:num_procs] |
|
115 | 117 | print "Running %s system on %s processes until %f"%(grid, partition, tstop) |
|
116 | 118 | |
|
117 | 119 | # functions defining initial/boundary/source conditions |
|
118 | 120 | def I(x,y): |
|
119 | 121 | from numpy import exp |
|
120 | 122 | return 1.5*exp(-100*((x-0.5)**2+(y-0.5)**2)) |
|
121 | 123 | def f(x,y,t): |
|
122 | 124 | return 0.0 |
|
123 | 125 | # from numpy import exp,sin |
|
124 | 126 | # return 10*exp(-(x - sin(100*t))**2) |
|
125 | 127 | def bc(x,y,t): |
|
126 | 128 | return 0.0 |
|
127 | 129 | |
|
128 | 130 | # initialize t_hist/u_hist for saving the state at each step (optional) |
|
129 | 131 | view['t_hist'] = [] |
|
130 | 132 | view['u_hist'] = [] |
|
131 | 133 | |
|
132 | 134 | # set vector/scalar implementation details |
|
133 | 135 | impl = {} |
|
134 | 136 | impl['ic'] = 'vectorized' |
|
135 | 137 | impl['inner'] = 'scalar' |
|
136 | 138 | impl['bc'] = 'vectorized' |
|
137 | 139 | |
|
138 | 140 | # execute some files so that the classes we need will be defined on the engines: |
|
139 | 141 | view.execute('import numpy') |
|
140 | 142 | view.run('communicator.py') |
|
141 | 143 | view.run('RectPartitioner.py') |
|
142 | 144 | view.run('wavesolver.py') |
|
143 | 145 | |
|
144 | 146 | # scatter engine IDs |
|
145 | 147 | view.scatter('my_id', range(num_procs), flatten=True) |
|
146 | 148 | |
|
147 | 149 | # create the engine connectors |
|
148 | 150 | view.execute('com = EngineCommunicator()') |
|
149 | 151 | |
|
150 | 152 | # gather the connection information into a single dict |
|
151 | 153 | ar = view.apply_async(lambda : com.info) |
|
152 | 154 | peers = ar.get_dict() |
|
153 | 155 | # print peers |
|
154 | 156 | # this is a dict, keyed by engine ID, of the connection info for the EngineCommunicators |
|
155 | 157 | |
|
156 | 158 | # setup remote partitioner |
|
157 | 159 | # note that Reference means that the argument passed to setup_partitioner will be the |
|
158 | 160 | # object named 'com' in the engine's namespace |
|
159 |
view.apply_sync |
|
|
161 | view.apply_sync(setup_partitioner, Reference('com'), peers, Reference('my_id'), num_procs, grid, partition) | |
|
160 | 162 | time.sleep(1) |
|
161 | 163 | # convenience lambda to call solver.solve: |
|
162 | 164 | _solve = lambda *args, **kwargs: solver.solve(*args, **kwargs) |
|
163 | 165 | |
|
164 | 166 | if ns.scalar: |
|
165 | 167 | impl['inner'] = 'scalar' |
|
166 | 168 | # setup remote solvers |
|
167 |
view.apply_sync |
|
|
169 | view.apply_sync(setup_solver, I,f,c,bc,Lx,Ly, partitioner=Reference('partitioner'), dt=0,implementation=impl) | |
|
168 | 170 | |
|
169 | 171 | # run first with element-wise Python operations for each cell |
|
170 | 172 | t0 = time.time() |
|
171 | 173 | ar = view.apply_async(_solve, tstop, dt=0, verbose=True, final_test=final_test, user_action=user_action) |
|
172 | 174 | if final_test: |
|
173 | 175 | # this sum is performed element-wise as results finish |
|
174 | 176 | s = sum(ar) |
|
175 | 177 | # the L2 norm (RMS) of the result: |
|
176 | 178 | norm = sqrt(s/num_cells) |
|
177 | 179 | else: |
|
178 | 180 | norm = -1 |
|
179 | 181 | t1 = time.time() |
|
180 | 182 | print 'scalar inner-version, Wtime=%g, norm=%g'%(t1-t0, norm) |
|
181 | 183 | |
|
182 | 184 | # run again with faster numpy-vectorized inner implementation: |
|
183 | 185 | impl['inner'] = 'vectorized' |
|
184 | 186 | # setup remote solvers |
|
185 |
view.apply_sync |
|
|
187 | view.apply_sync(setup_solver, I,f,c,bc,Lx,Ly,partitioner=Reference('partitioner'), dt=0,implementation=impl) | |
|
186 | 188 | |
|
187 | 189 | t0 = time.time() |
|
188 | 190 | |
|
189 | 191 | ar = view.apply_async(_solve, tstop, dt=0, verbose=True, final_test=final_test)#, user_action=wave_saver) |
|
190 | 192 | if final_test: |
|
191 | 193 | # this sum is performed element-wise as results finish |
|
192 | 194 | s = sum(ar) |
|
193 | 195 | # the L2 norm (RMS) of the result: |
|
194 | 196 | norm = sqrt(s/num_cells) |
|
195 | 197 | else: |
|
196 | 198 | norm = -1 |
|
197 | 199 | t1 = time.time() |
|
198 | 200 | print 'vector inner-version, Wtime=%g, norm=%g'%(t1-t0, norm) |
|
199 | 201 | |
|
200 | 202 | # if ns.save is True, then u_hist stores the history of u as a list |
|
201 | 203 | # If the partion scheme is Nx1, then u can be reconstructed via 'gather': |
|
202 | 204 | if ns.save and partition[-1] == 1: |
|
203 | 205 | import pylab |
|
204 | 206 | view.execute('u_last=u_hist[-1]') |
|
205 | 207 | u_last = view.gather('u_last', block=True) |
|
206 | 208 | pylab.pcolor(u_last) |
|
207 | 209 | pylab.show() No newline at end of file |
@@ -1,34 +1,34 b'' | |||
|
1 | 1 | ===================== |
|
2 | 2 | IPython Documentation |
|
3 | 3 | ===================== |
|
4 | 4 | |
|
5 | 5 | .. htmlonly:: |
|
6 | 6 | |
|
7 | 7 | :Release: |release| |
|
8 | 8 | :Date: |today| |
|
9 | 9 | |
|
10 | 10 | Welcome to the official IPython documentation. |
|
11 | 11 | |
|
12 | 12 | Contents |
|
13 | 13 | ======== |
|
14 | 14 | |
|
15 | 15 | .. toctree:: |
|
16 | 16 | :maxdepth: 1 |
|
17 | 17 | |
|
18 | 18 | overview.txt |
|
19 | 19 | whatsnew/index.txt |
|
20 | 20 | install/index.txt |
|
21 | 21 | interactive/index.txt |
|
22 | parallel/index.txt | |
|
22 | .. parallel/index.txt | |
|
23 | 23 | parallelz/index.txt |
|
24 | 24 | config/index.txt |
|
25 | 25 | development/index.txt |
|
26 | 26 | api/index.txt |
|
27 | 27 | faq.txt |
|
28 | 28 | about/index.txt |
|
29 | 29 | |
|
30 | 30 | .. htmlonly:: |
|
31 | 31 | * :ref:`genindex` |
|
32 | 32 | * :ref:`modindex` |
|
33 | 33 | * :ref:`search` |
|
34 | 34 |
@@ -1,172 +1,173 b'' | |||
|
1 | 1 | .. _dag_dependencies: |
|
2 | 2 | |
|
3 | 3 | ================ |
|
4 | 4 | DAG Dependencies |
|
5 | 5 | ================ |
|
6 | 6 | |
|
7 | 7 | Often, parallel workflow is described in terms of a `Directed Acyclic Graph |
|
8 | 8 | <http://en.wikipedia.org/wiki/Directed_acyclic_graph>`_ or DAG. A popular library |
|
9 | 9 | for working with Graphs is NetworkX_. Here, we will walk through a demo mapping |
|
10 | 10 | a nx DAG to task dependencies. |
|
11 | 11 | |
|
12 | 12 | The full script that runs this demo can be found in |
|
13 | 13 | :file:`docs/examples/newparallel/dagdeps.py`. |
|
14 | 14 | |
|
15 | 15 | Why are DAGs good for task dependencies? |
|
16 | 16 | ---------------------------------------- |
|
17 | 17 | |
|
18 | 18 | The 'G' in DAG is 'Graph'. A Graph is a collection of **nodes** and **edges** that connect |
|
19 | 19 | the nodes. For our purposes, each node would be a task, and each edge would be a |
|
20 | 20 | dependency. The 'D' in DAG stands for 'Directed'. This means that each edge has a |
|
21 | 21 | direction associated with it. So we can interpret the edge (a,b) as meaning that b depends |
|
22 | 22 | on a, whereas the edge (b,a) would mean a depends on b. The 'A' is 'Acyclic', meaning that |
|
23 | 23 | there must not be any closed loops in the graph. This is important for dependencies, |
|
24 | 24 | because if a loop were closed, then a task could ultimately depend on itself, and never be |
|
25 | 25 | able to run. If your workflow can be described as a DAG, then it is impossible for your |
|
26 | 26 | dependencies to cause a deadlock. |
|
27 | 27 | |
|
28 | 28 | A Sample DAG |
|
29 | 29 | ------------ |
|
30 | 30 | |
|
31 | 31 | Here, we have a very simple 5-node DAG: |
|
32 | 32 | |
|
33 | 33 | .. figure:: simpledag.* |
|
34 | 34 | |
|
35 | 35 | With NetworkX, an arrow is just a fattened bit on the edge. Here, we can see that task 0 |
|
36 | 36 | depends on nothing, and can run immediately. 1 and 2 depend on 0; 3 depends on |
|
37 | 37 | 1 and 2; and 4 depends only on 1. |
|
38 | 38 | |
|
39 | 39 | A possible sequence of events for this workflow: |
|
40 | 40 | |
|
41 | 41 | 0. Task 0 can run right away |
|
42 | 42 | 1. 0 finishes, so 1,2 can start |
|
43 | 43 | 2. 1 finishes, 3 is still waiting on 2, but 4 can start right away |
|
44 | 44 | 3. 2 finishes, and 3 can finally start |
|
45 | 45 | |
|
46 | 46 | |
|
47 | 47 | Further, taking failures into account, assuming all dependencies are run with the default |
|
48 |
`success |
|
|
48 | `success=True,failure=False`, the following cases would occur for each node's failure: | |
|
49 | 49 | |
|
50 | 50 | 0. fails: all other tasks fail as Impossible |
|
51 | 51 | 1. 2 can still succeed, but 3,4 are unreachable |
|
52 | 52 | 2. 3 becomes unreachable, but 4 is unaffected |
|
53 | 53 | 3. and 4. are terminal, and can have no effect on other nodes |
|
54 | 54 | |
|
55 | 55 | The code to generate the simple DAG: |
|
56 | 56 | |
|
57 | 57 | .. sourcecode:: python |
|
58 | 58 | |
|
59 | 59 | import networkx as nx |
|
60 | 60 | |
|
61 | 61 | G = nx.DiGraph() |
|
62 | 62 | |
|
63 | 63 | # add 5 nodes, labeled 0-4: |
|
64 | 64 | map(G.add_node, range(5)) |
|
65 | 65 | # 1,2 depend on 0: |
|
66 | 66 | G.add_edge(0,1) |
|
67 | 67 | G.add_edge(0,2) |
|
68 | 68 | # 3 depends on 1,2 |
|
69 | 69 | G.add_edge(1,3) |
|
70 | 70 | G.add_edge(2,3) |
|
71 | 71 | # 4 depends on 1 |
|
72 | 72 | G.add_edge(1,4) |
|
73 | 73 | |
|
74 | 74 | # now draw the graph: |
|
75 | 75 | pos = { 0 : (0,0), 1 : (1,1), 2 : (-1,1), |
|
76 | 76 | 3 : (0,2), 4 : (2,2)} |
|
77 | 77 | nx.draw(G, pos, edge_color='r') |
|
78 | 78 | |
|
79 | 79 | |
|
80 | 80 | For demonstration purposes, we have a function that generates a random DAG with a given |
|
81 | 81 | number of nodes and edges. |
|
82 | 82 | |
|
83 | 83 | .. literalinclude:: ../../examples/newparallel/dagdeps.py |
|
84 | 84 | :language: python |
|
85 | 85 | :lines: 20-36 |
|
86 | 86 | |
|
87 | 87 | So first, we start with a graph of 32 nodes, with 128 edges: |
|
88 | 88 | |
|
89 | 89 | .. sourcecode:: ipython |
|
90 | 90 | |
|
91 | 91 | In [2]: G = random_dag(32,128) |
|
92 | 92 | |
|
93 | 93 | Now, we need to build our dict of jobs corresponding to the nodes on the graph: |
|
94 | 94 | |
|
95 | 95 | .. sourcecode:: ipython |
|
96 | 96 | |
|
97 | 97 | In [3]: jobs = {} |
|
98 | 98 | |
|
99 | 99 | # in reality, each job would presumably be different |
|
100 | 100 | # randomwait is just a function that sleeps for a random interval |
|
101 | 101 | In [4]: for node in G: |
|
102 | 102 | ...: jobs[node] = randomwait |
|
103 | 103 | |
|
104 | 104 | Once we have a dict of jobs matching the nodes on the graph, we can start submitting jobs, |
|
105 | 105 | and linking up the dependencies. Since we don't know a job's msg_id until it is submitted, |
|
106 | 106 | which is necessary for building dependencies, it is critical that we don't submit any jobs |
|
107 | 107 | before other jobs it may depend on. Fortunately, NetworkX provides a |
|
108 | 108 | :meth:`topological_sort` method which ensures exactly this. It presents an iterable, that |
|
109 | 109 | guarantees that when you arrive at a node, you have already visited all the nodes it |
|
110 | 110 | on which it depends: |
|
111 | 111 | |
|
112 | 112 | .. sourcecode:: ipython |
|
113 | 113 | |
|
114 | In [5]: c = client.Client() | |
|
114 | In [5]: rc = client.Client() | |
|
115 | In [5]: view = rc.load_balanced_view() | |
|
115 | 116 | |
|
116 | 117 | In [6]: results = {} |
|
117 | 118 | |
|
118 | 119 | In [7]: for node in G.topological_sort(): |
|
119 | 120 | ...: # get list of AsyncResult objects from nodes |
|
120 | 121 | ...: # leading into this one as dependencies |
|
121 | 122 | ...: deps = [ results[n] for n in G.predecessors(node) ] |
|
122 | 123 | ...: # submit and store AsyncResult object |
|
123 |
...: results[node] = |
|
|
124 | ...: results[node] = view.apply_with_flags(jobs[node], after=deps, block=False) | |
|
124 | 125 | |
|
125 | 126 | Now that we have submitted all the jobs, we can wait for the results: |
|
126 | 127 | |
|
127 | 128 | .. sourcecode:: ipython |
|
128 | 129 | |
|
129 |
In [8]: |
|
|
130 | In [8]: view.wait(results.values()) | |
|
130 | 131 | |
|
131 | 132 | Now, at least we know that all the jobs ran and did not fail (``r.get()`` would have |
|
132 | 133 | raised an error if a task failed). But we don't know that the ordering was properly |
|
133 | 134 | respected. For this, we can use the :attr:`metadata` attribute of each AsyncResult. |
|
134 | 135 | |
|
135 | 136 | These objects store a variety of metadata about each task, including various timestamps. |
|
136 | 137 | We can validate that the dependencies were respected by checking that each task was |
|
137 | 138 | started after all of its predecessors were completed: |
|
138 | 139 | |
|
139 | 140 | .. literalinclude:: ../../examples/newparallel/dagdeps.py |
|
140 | 141 | :language: python |
|
141 | 142 | :lines: 64-70 |
|
142 | 143 | |
|
143 | 144 | We can also validate the graph visually. By drawing the graph with each node's x-position |
|
144 | 145 | as its start time, all arrows must be pointing to the right if dependencies were respected. |
|
145 | 146 | For spreading, the y-position will be the runtime of the task, so long tasks |
|
146 | 147 | will be at the top, and quick, small tasks will be at the bottom. |
|
147 | 148 | |
|
148 | 149 | .. sourcecode:: ipython |
|
149 | 150 | |
|
150 | 151 | In [10]: from matplotlib.dates import date2num |
|
151 | 152 | |
|
152 | 153 | In [11]: from matplotlib.cm import gist_rainbow |
|
153 | 154 | |
|
154 | 155 | In [12]: pos = {}; colors = {} |
|
155 | 156 | |
|
156 | 157 | In [12]: for node in G: |
|
157 | 158 | ...: md = results[node].metadata |
|
158 | 159 | ...: start = date2num(md.started) |
|
159 | 160 | ...: runtime = date2num(md.completed) - start |
|
160 | 161 | ...: pos[node] = (start, runtime) |
|
161 | 162 | ...: colors[node] = md.engine_id |
|
162 | 163 | |
|
163 | 164 | In [13]: nx.draw(G, pos, node_list=colors.keys(), node_color=colors.values(), |
|
164 | 165 | ...: cmap=gist_rainbow) |
|
165 | 166 | |
|
166 | 167 | .. figure:: dagdeps.* |
|
167 | 168 | |
|
168 | 169 | Time started on x, runtime on y, and color-coded by engine-id (in this case there |
|
169 | 170 | were four engines). Edges denote dependencies. |
|
170 | 171 | |
|
171 | 172 | |
|
172 | 173 | .. _NetworkX: http://networkx.lanl.gov/ |
@@ -1,21 +1,22 b'' | |||
|
1 | 1 | .. _parallelz_index: |
|
2 | 2 | |
|
3 | 3 | ========================================== |
|
4 | 4 | Using IPython for parallel computing (ZMQ) |
|
5 | 5 | ========================================== |
|
6 | 6 | |
|
7 | 7 | .. toctree:: |
|
8 | 8 | :maxdepth: 2 |
|
9 | 9 | |
|
10 | 10 | parallel_intro.txt |
|
11 | 11 | parallel_process.txt |
|
12 | 12 | parallel_multiengine.txt |
|
13 | 13 | parallel_task.txt |
|
14 | 14 | parallel_mpi.txt |
|
15 | 15 | parallel_security.txt |
|
16 | 16 | parallel_winhpc.txt |
|
17 | 17 | parallel_demos.txt |
|
18 | 18 | dag_dependencies.txt |
|
19 | 19 | parallel_details.txt |
|
20 | parallel_transition.txt | |
|
20 | 21 | |
|
21 | 22 |
@@ -1,284 +1,284 b'' | |||
|
1 | 1 | ================= |
|
2 | 2 | Parallel examples |
|
3 | 3 | ================= |
|
4 | 4 | |
|
5 | 5 | .. note:: |
|
6 | 6 | |
|
7 | 7 | Performance numbers from ``IPython.kernel``, not newparallel. |
|
8 | 8 | |
|
9 | 9 | In this section we describe two more involved examples of using an IPython |
|
10 | 10 | cluster to perform a parallel computation. In these examples, we will be using |
|
11 | 11 | IPython's "pylab" mode, which enables interactive plotting using the |
|
12 | 12 | Matplotlib package. IPython can be started in this mode by typing:: |
|
13 | 13 | |
|
14 | 14 | ipython --pylab |
|
15 | 15 | |
|
16 | 16 | at the system command line. |
|
17 | 17 | |
|
18 | 18 | 150 million digits of pi |
|
19 | 19 | ======================== |
|
20 | 20 | |
|
21 | 21 | In this example we would like to study the distribution of digits in the |
|
22 | 22 | number pi (in base 10). While it is not known if pi is a normal number (a |
|
23 | 23 | number is normal in base 10 if 0-9 occur with equal likelihood) numerical |
|
24 | 24 | investigations suggest that it is. We will begin with a serial calculation on |
|
25 | 25 | 10,000 digits of pi and then perform a parallel calculation involving 150 |
|
26 | 26 | million digits. |
|
27 | 27 | |
|
28 | 28 | In both the serial and parallel calculation we will be using functions defined |
|
29 | 29 | in the :file:`pidigits.py` file, which is available in the |
|
30 | 30 | :file:`docs/examples/newparallel` directory of the IPython source distribution. |
|
31 | 31 | These functions provide basic facilities for working with the digits of pi and |
|
32 | 32 | can be loaded into IPython by putting :file:`pidigits.py` in your current |
|
33 | 33 | working directory and then doing: |
|
34 | 34 | |
|
35 | 35 | .. sourcecode:: ipython |
|
36 | 36 | |
|
37 | 37 | In [1]: run pidigits.py |
|
38 | 38 | |
|
39 | 39 | Serial calculation |
|
40 | 40 | ------------------ |
|
41 | 41 | |
|
42 | 42 | For the serial calculation, we will use `SymPy <http://www.sympy.org>`_ to |
|
43 | 43 | calculate 10,000 digits of pi and then look at the frequencies of the digits |
|
44 | 44 | 0-9. Out of 10,000 digits, we expect each digit to occur 1,000 times. While |
|
45 | 45 | SymPy is capable of calculating many more digits of pi, our purpose here is to |
|
46 | 46 | set the stage for the much larger parallel calculation. |
|
47 | 47 | |
|
48 | 48 | In this example, we use two functions from :file:`pidigits.py`: |
|
49 | 49 | :func:`one_digit_freqs` (which calculates how many times each digit occurs) |
|
50 | 50 | and :func:`plot_one_digit_freqs` (which uses Matplotlib to plot the result). |
|
51 | 51 | Here is an interactive IPython session that uses these functions with |
|
52 | 52 | SymPy: |
|
53 | 53 | |
|
54 | 54 | .. sourcecode:: ipython |
|
55 | 55 | |
|
56 | 56 | In [7]: import sympy |
|
57 | 57 | |
|
58 | 58 | In [8]: pi = sympy.pi.evalf(40) |
|
59 | 59 | |
|
60 | 60 | In [9]: pi |
|
61 | 61 | Out[9]: 3.141592653589793238462643383279502884197 |
|
62 | 62 | |
|
63 | 63 | In [10]: pi = sympy.pi.evalf(10000) |
|
64 | 64 | |
|
65 | 65 | In [11]: digits = (d for d in str(pi)[2:]) # create a sequence of digits |
|
66 | 66 | |
|
67 | 67 | In [12]: run pidigits.py # load one_digit_freqs/plot_one_digit_freqs |
|
68 | 68 | |
|
69 | 69 | In [13]: freqs = one_digit_freqs(digits) |
|
70 | 70 | |
|
71 | 71 | In [14]: plot_one_digit_freqs(freqs) |
|
72 | 72 | Out[14]: [<matplotlib.lines.Line2D object at 0x18a55290>] |
|
73 | 73 | |
|
74 | 74 | The resulting plot of the single digit counts shows that each digit occurs |
|
75 | 75 | approximately 1,000 times, but that with only 10,000 digits the |
|
76 | 76 | statistical fluctuations are still rather large: |
|
77 | 77 | |
|
78 | 78 | .. image:: ../parallel/single_digits.* |
|
79 | 79 | |
|
80 | 80 | It is clear that to reduce the relative fluctuations in the counts, we need |
|
81 | 81 | to look at many more digits of pi. That brings us to the parallel calculation. |
|
82 | 82 | |
|
83 | 83 | Parallel calculation |
|
84 | 84 | -------------------- |
|
85 | 85 | |
|
86 | 86 | Calculating many digits of pi is a challenging computational problem in itself. |
|
87 | 87 | Because we want to focus on the distribution of digits in this example, we |
|
88 | 88 | will use pre-computed digit of pi from the website of Professor Yasumasa |
|
89 | 89 | Kanada at the University of Tokyo (http://www.super-computing.org). These |
|
90 | 90 | digits come in a set of text files (ftp://pi.super-computing.org/.2/pi200m/) |
|
91 | 91 | that each have 10 million digits of pi. |
|
92 | 92 | |
|
93 | 93 | For the parallel calculation, we have copied these files to the local hard |
|
94 | 94 | drives of the compute nodes. A total of 15 of these files will be used, for a |
|
95 | 95 | total of 150 million digits of pi. To make things a little more interesting we |
|
96 | 96 | will calculate the frequencies of all 2 digits sequences (00-99) and then plot |
|
97 | 97 | the result using a 2D matrix in Matplotlib. |
|
98 | 98 | |
|
99 | 99 | The overall idea of the calculation is simple: each IPython engine will |
|
100 | 100 | compute the two digit counts for the digits in a single file. Then in a final |
|
101 | 101 | step the counts from each engine will be added up. To perform this |
|
102 | 102 | calculation, we will need two top-level functions from :file:`pidigits.py`: |
|
103 | 103 | |
|
104 | 104 | .. literalinclude:: ../../examples/newparallel/pidigits.py |
|
105 | 105 | :language: python |
|
106 | 106 | :lines: 41-56 |
|
107 | 107 | |
|
108 | 108 | We will also use the :func:`plot_two_digit_freqs` function to plot the |
|
109 | 109 | results. The code to run this calculation in parallel is contained in |
|
110 | 110 | :file:`docs/examples/newparallel/parallelpi.py`. This code can be run in parallel |
|
111 | 111 | using IPython by following these steps: |
|
112 | 112 | |
|
113 | 113 | 1. Use :command:`ipclusterz` to start 15 engines. We used an 8 core (2 quad |
|
114 | 114 | core CPUs) cluster with hyperthreading enabled which makes the 8 cores |
|
115 | 115 | looks like 16 (1 controller + 15 engines) in the OS. However, the maximum |
|
116 | 116 | speedup we can observe is still only 8x. |
|
117 | 117 | 2. With the file :file:`parallelpi.py` in your current working directory, open |
|
118 | 118 | up IPython in pylab mode and type ``run parallelpi.py``. This will download |
|
119 | 119 | the pi files via ftp the first time you run it, if they are not |
|
120 | 120 | present in the Engines' working directory. |
|
121 | 121 | |
|
122 | 122 | When run on our 8 core cluster, we observe a speedup of 7.7x. This is slightly |
|
123 | 123 | less than linear scaling (8x) because the controller is also running on one of |
|
124 | 124 | the cores. |
|
125 | 125 | |
|
126 | 126 | To emphasize the interactive nature of IPython, we now show how the |
|
127 | 127 | calculation can also be run by simply typing the commands from |
|
128 | 128 | :file:`parallelpi.py` interactively into IPython: |
|
129 | 129 | |
|
130 | 130 | .. sourcecode:: ipython |
|
131 | 131 | |
|
132 | 132 | In [1]: from IPython.zmq.parallel import client |
|
133 | 133 | |
|
134 | 134 | # The Client allows us to use the engines interactively. |
|
135 | 135 | # We simply pass Client the name of the cluster profile we |
|
136 | 136 | # are using. |
|
137 | 137 | In [2]: c = client.Client(profile='mycluster') |
|
138 |
In [3]: view = c. |
|
|
138 | In [3]: view = c.load_balanced_view() | |
|
139 | 139 | |
|
140 | 140 | In [3]: c.ids |
|
141 | 141 | Out[3]: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14] |
|
142 | 142 | |
|
143 | 143 | In [4]: run pidigits.py |
|
144 | 144 | |
|
145 | 145 | In [5]: filestring = 'pi200m.ascii.%(i)02dof20' |
|
146 | 146 | |
|
147 | 147 | # Create the list of files to process. |
|
148 | 148 | In [6]: files = [filestring % {'i':i} for i in range(1,16)] |
|
149 | 149 | |
|
150 | 150 | In [7]: files |
|
151 | 151 | Out[7]: |
|
152 | 152 | ['pi200m.ascii.01of20', |
|
153 | 153 | 'pi200m.ascii.02of20', |
|
154 | 154 | 'pi200m.ascii.03of20', |
|
155 | 155 | 'pi200m.ascii.04of20', |
|
156 | 156 | 'pi200m.ascii.05of20', |
|
157 | 157 | 'pi200m.ascii.06of20', |
|
158 | 158 | 'pi200m.ascii.07of20', |
|
159 | 159 | 'pi200m.ascii.08of20', |
|
160 | 160 | 'pi200m.ascii.09of20', |
|
161 | 161 | 'pi200m.ascii.10of20', |
|
162 | 162 | 'pi200m.ascii.11of20', |
|
163 | 163 | 'pi200m.ascii.12of20', |
|
164 | 164 | 'pi200m.ascii.13of20', |
|
165 | 165 | 'pi200m.ascii.14of20', |
|
166 | 166 | 'pi200m.ascii.15of20'] |
|
167 | 167 | |
|
168 | 168 | # download the data files if they don't already exist: |
|
169 |
In [8]: |
|
|
169 | In [8]: v.map(fetch_pi_file, files) | |
|
170 | 170 | |
|
171 | 171 | # This is the parallel calculation using the Client.map method |
|
172 | 172 | # which applies compute_two_digit_freqs to each file in files in parallel. |
|
173 |
In [9]: freqs_all = |
|
|
173 | In [9]: freqs_all = v.map(compute_two_digit_freqs, files) | |
|
174 | 174 | |
|
175 | 175 | # Add up the frequencies from each engine. |
|
176 | 176 | In [10]: freqs = reduce_freqs(freqs_all) |
|
177 | 177 | |
|
178 | 178 | In [11]: plot_two_digit_freqs(freqs) |
|
179 | 179 | Out[11]: <matplotlib.image.AxesImage object at 0x18beb110> |
|
180 | 180 | |
|
181 | 181 | In [12]: plt.title('2 digit counts of 150m digits of pi') |
|
182 | 182 | Out[12]: <matplotlib.text.Text object at 0x18d1f9b0> |
|
183 | 183 | |
|
184 | 184 | The resulting plot generated by Matplotlib is shown below. The colors indicate |
|
185 | 185 | which two digit sequences are more (red) or less (blue) likely to occur in the |
|
186 | 186 | first 150 million digits of pi. We clearly see that the sequence "41" is |
|
187 | 187 | most likely and that "06" and "07" are least likely. Further analysis would |
|
188 | 188 | show that the relative size of the statistical fluctuations have decreased |
|
189 | 189 | compared to the 10,000 digit calculation. |
|
190 | 190 | |
|
191 | 191 | .. image:: ../parallel/two_digit_counts.* |
|
192 | 192 | |
|
193 | 193 | |
|
194 | 194 | Parallel options pricing |
|
195 | 195 | ======================== |
|
196 | 196 | |
|
197 | 197 | An option is a financial contract that gives the buyer of the contract the |
|
198 | 198 | right to buy (a "call") or sell (a "put") a secondary asset (a stock for |
|
199 | 199 | example) at a particular date in the future (the expiration date) for a |
|
200 | 200 | pre-agreed upon price (the strike price). For this right, the buyer pays the |
|
201 | 201 | seller a premium (the option price). There are a wide variety of flavors of |
|
202 | 202 | options (American, European, Asian, etc.) that are useful for different |
|
203 | 203 | purposes: hedging against risk, speculation, etc. |
|
204 | 204 | |
|
205 | 205 | Much of modern finance is driven by the need to price these contracts |
|
206 | 206 | accurately based on what is known about the properties (such as volatility) of |
|
207 | 207 | the underlying asset. One method of pricing options is to use a Monte Carlo |
|
208 | 208 | simulation of the underlying asset price. In this example we use this approach |
|
209 | 209 | to price both European and Asian (path dependent) options for various strike |
|
210 | 210 | prices and volatilities. |
|
211 | 211 | |
|
212 | 212 | The code for this example can be found in the :file:`docs/examples/newparallel` |
|
213 | 213 | directory of the IPython source. The function :func:`price_options` in |
|
214 | 214 | :file:`mcpricer.py` implements the basic Monte Carlo pricing algorithm using |
|
215 | 215 | the NumPy package and is shown here: |
|
216 | 216 | |
|
217 | 217 | .. literalinclude:: ../../examples/newparallel/mcpricer.py |
|
218 | 218 | :language: python |
|
219 | 219 | |
|
220 | 220 | To run this code in parallel, we will use IPython's :class:`LoadBalancedView` class, |
|
221 | 221 | which distributes work to the engines using dynamic load balancing. This |
|
222 | 222 | view is a wrapper of the :class:`Client` class shown in |
|
223 | 223 | the previous example. The parallel calculation using :class:`LoadBalancedView` can |
|
224 | 224 | be found in the file :file:`mcpricer.py`. The code in this file creates a |
|
225 | 225 | :class:`TaskClient` instance and then submits a set of tasks using |
|
226 | 226 | :meth:`TaskClient.run` that calculate the option prices for different |
|
227 | 227 | volatilities and strike prices. The results are then plotted as a 2D contour |
|
228 | 228 | plot using Matplotlib. |
|
229 | 229 | |
|
230 | 230 | .. literalinclude:: ../../examples/newparallel/mcdriver.py |
|
231 | 231 | :language: python |
|
232 | 232 | |
|
233 | 233 | To use this code, start an IPython cluster using :command:`ipclusterz`, open |
|
234 | 234 | IPython in the pylab mode with the file :file:`mcdriver.py` in your current |
|
235 | 235 | working directory and then type: |
|
236 | 236 | |
|
237 | 237 | .. sourcecode:: ipython |
|
238 | 238 | |
|
239 | 239 | In [7]: run mcdriver.py |
|
240 | 240 | Submitted tasks: [0, 1, 2, ...] |
|
241 | 241 | |
|
242 | 242 | Once all the tasks have finished, the results can be plotted using the |
|
243 | 243 | :func:`plot_options` function. Here we make contour plots of the Asian |
|
244 | 244 | call and Asian put options as function of the volatility and strike price: |
|
245 | 245 | |
|
246 | 246 | .. sourcecode:: ipython |
|
247 | 247 | |
|
248 | 248 | In [8]: plot_options(sigma_vals, K_vals, prices['acall']) |
|
249 | 249 | |
|
250 | 250 | In [9]: plt.figure() |
|
251 | 251 | Out[9]: <matplotlib.figure.Figure object at 0x18c178d0> |
|
252 | 252 | |
|
253 | 253 | In [10]: plot_options(sigma_vals, K_vals, prices['aput']) |
|
254 | 254 | |
|
255 | 255 | These results are shown in the two figures below. On a 8 core cluster the |
|
256 | 256 | entire calculation (10 strike prices, 10 volatilities, 100,000 paths for each) |
|
257 | 257 | took 30 seconds in parallel, giving a speedup of 7.7x, which is comparable |
|
258 | 258 | to the speedup observed in our previous example. |
|
259 | 259 | |
|
260 | 260 | .. image:: ../parallel/asian_call.* |
|
261 | 261 | |
|
262 | 262 | .. image:: ../parallel/asian_put.* |
|
263 | 263 | |
|
264 | 264 | Conclusion |
|
265 | 265 | ========== |
|
266 | 266 | |
|
267 | 267 | To conclude these examples, we summarize the key features of IPython's |
|
268 | 268 | parallel architecture that have been demonstrated: |
|
269 | 269 | |
|
270 | 270 | * Serial code can be parallelized often with only a few extra lines of code. |
|
271 | 271 | We have used the :class:`DirectView` and :class:`LoadBalancedView` classes |
|
272 | 272 | for this purpose. |
|
273 | 273 | * The resulting parallel code can be run without ever leaving the IPython's |
|
274 | 274 | interactive shell. |
|
275 | 275 | * Any data computed in parallel can be explored interactively through |
|
276 | 276 | visualization or further numerical calculations. |
|
277 | 277 | * We have run these examples on a cluster running Windows HPC Server 2008. |
|
278 | 278 | IPython's built in support for the Windows HPC job scheduler makes it |
|
279 | 279 | easy to get started with IPython's parallel capabilities. |
|
280 | 280 | |
|
281 | 281 | .. note:: |
|
282 | 282 | |
|
283 | 283 | The newparallel code has never been run on Windows HPC Server, so the last |
|
284 | 284 | conclusion is untested. |
@@ -1,438 +1,493 b'' | |||
|
1 | 1 | .. _parallel_details: |
|
2 | 2 | |
|
3 | 3 | ========================================== |
|
4 | 4 | Details of Parallel Computing with IPython |
|
5 | 5 | ========================================== |
|
6 | 6 | |
|
7 | 7 | .. note:: |
|
8 | 8 | |
|
9 | 9 | There are still many sections to fill out |
|
10 | 10 | |
|
11 | 11 | |
|
12 | 12 | Caveats |
|
13 | 13 | ======= |
|
14 | 14 | |
|
15 | 15 | First, some caveats about the detailed workings of parallel computing with 0MQ and IPython. |
|
16 | 16 | |
|
17 | 17 | Non-copying sends and numpy arrays |
|
18 | 18 | ---------------------------------- |
|
19 | 19 | |
|
20 | 20 | When numpy arrays are passed as arguments to apply or via data-movement methods, they are not |
|
21 |
copied. |
|
|
22 |
PyZMQ does allow you to track when a message has been sent so you can know when it is safe |
|
|
23 | IPython only allows for this. | |
|
21 | copied. This means that you must be careful if you are sending an array that you intend to work | |
|
22 | on. PyZMQ does allow you to track when a message has been sent so you can know when it is safe | |
|
23 | to edit the buffer, but IPython only allows for this. | |
|
24 | 24 | |
|
25 | 25 |
It is also important to note that the non-copying receive of a message is *read-only*. |
|
26 |
means that if you intend to work in-place on an array that you have sent or received, you must |
|
|
27 |
it. |
|
|
26 | means that if you intend to work in-place on an array that you have sent or received, you must | |
|
27 | copy it. This is true for both numpy arrays sent to engines and numpy arrays retrieved as | |
|
28 | results. | |
|
28 | 29 | |
|
29 | 30 | The following will fail: |
|
30 | 31 | |
|
31 | 32 | .. sourcecode:: ipython |
|
32 | 33 | |
|
33 | 34 | In [3]: A = numpy.zeros(2) |
|
34 | 35 | |
|
35 | 36 | In [4]: def setter(a): |
|
36 | 37 | ...: a[0]=1 |
|
37 | 38 | ...: return a |
|
38 | 39 | |
|
39 | 40 | In [5]: rc[0].apply_sync(setter, A) |
|
40 | 41 | --------------------------------------------------------------------------- |
|
41 | 42 | RemoteError Traceback (most recent call last) |
|
42 | 43 | ... |
|
43 | 44 | RemoteError: RuntimeError(array is not writeable) |
|
44 | 45 | Traceback (most recent call last): |
|
45 | 46 | File "/Users/minrk/dev/ip/mine/IPython/zmq/parallel/streamkernel.py", line 329, in apply_request |
|
46 | 47 | exec code in working, working |
|
47 | 48 | File "<string>", line 1, in <module> |
|
48 | 49 | File "<ipython-input-14-736187483856>", line 2, in setter |
|
49 | 50 | RuntimeError: array is not writeable |
|
50 | 51 | |
|
51 | 52 | If you do need to edit the array in-place, just remember to copy the array if it's read-only. |
|
52 | 53 | The :attr:`ndarray.flags.writeable` flag will tell you if you can write to an array. |
|
53 | 54 | |
|
54 | 55 | .. sourcecode:: ipython |
|
55 | 56 | |
|
56 | 57 | In [3]: A = numpy.zeros(2) |
|
57 | 58 | |
|
58 | 59 | In [4]: def setter(a): |
|
59 | 60 | ...: """only copy read-only arrays""" |
|
60 | 61 | ...: if not a.flags.writeable: |
|
61 | 62 | ...: a=a.copy() |
|
62 | 63 | ...: a[0]=1 |
|
63 | 64 | ...: return a |
|
64 | 65 | |
|
65 | 66 | In [5]: rc[0].apply_sync(setter, A) |
|
66 | 67 | Out[5]: array([ 1., 0.]) |
|
67 | 68 | |
|
68 | 69 | # note that results will also be read-only: |
|
69 | 70 | In [6]: _.flags.writeable |
|
70 | 71 | Out[6]: False |
|
71 | 72 | |
|
73 | If you want to safely edit an array in-place after *sending* it, you must use the `track=True` flag. IPython always performs non-copying sends of arrays, which return immediately. You | |
|
74 | must instruct IPython track those messages *at send time* in order to know for sure that the send has completed. AsyncResults have a :attr:`sent` property, and :meth:`wait_on_send` method | |
|
75 | for checking and waiting for 0MQ to finish with a buffer. | |
|
76 | ||
|
77 | .. sourcecode:: ipython | |
|
78 | ||
|
79 | In [5]: A = numpy.random.random((1024,1024)) | |
|
80 | ||
|
81 | In [6]: view.track=True | |
|
82 | ||
|
83 | In [7]: ar = view.apply_async(lambda x: 2*x, A) | |
|
84 | ||
|
85 | In [8]: ar.sent | |
|
86 | Out[8]: False | |
|
87 | ||
|
88 | In [9]: ar.wait_on_send() # blocks until sent is True | |
|
89 | ||
|
90 | ||
|
72 | 91 | What is sendable? |
|
73 | 92 | ----------------- |
|
74 | 93 | |
|
75 | 94 | If IPython doesn't know what to do with an object, it will pickle it. There is a short list of |
|
76 | 95 | objects that are not pickled: ``buffers``, ``str/bytes`` objects, and ``numpy`` |
|
77 | 96 | arrays. These are handled specially by IPython in order to prevent the copying of data. Sending |
|
78 | 97 | bytes or numpy arrays will result in exactly zero in-memory copies of your data (unless the data |
|
79 | 98 | is very small). |
|
80 | 99 | |
|
81 | 100 | If you have an object that provides a Python buffer interface, then you can always send that |
|
82 | 101 | buffer without copying - and reconstruct the object on the other side in your own code. It is |
|
83 | 102 | possible that the object reconstruction will become extensible, so you can add your own |
|
84 | 103 | non-copying types, but this does not yet exist. |
|
85 | 104 | |
|
105 | Closures | |
|
106 | ******** | |
|
107 | ||
|
108 | Just about anything in Python is pickleable. The one notable exception is objects (generally | |
|
109 | functions) with *closures*. Closures can be a complicated topic, but the basic principal is that | |
|
110 | functions that refer to variables in their parent scope have closures. | |
|
111 | ||
|
112 | An example of a function that uses a closure: | |
|
113 | ||
|
114 | .. sourcecode:: python | |
|
115 | ||
|
116 | def f(a): | |
|
117 | def inner(): | |
|
118 | # inner will have a closure | |
|
119 | return a | |
|
120 | return echo | |
|
121 | ||
|
122 | f1 = f(1) | |
|
123 | f2 = f(2) | |
|
124 | f1() # returns 1 | |
|
125 | f2() # returns 2 | |
|
126 | ||
|
127 | f1 and f2 will have closures referring to the scope in which `inner` was defined, because they | |
|
128 | use the variable 'a'. As a result, you would not be able to send ``f1`` or ``f2`` with IPython. | |
|
129 | Note that you *would* be able to send `f`. This is only true for interactively defined | |
|
130 | functions (as are often used in decorators), and only when there are variables used inside the | |
|
131 | inner function, that are defined in the outer function. If the names are *not* in the outer | |
|
132 | function, then there will not be a closure, and the generated function will look in | |
|
133 | ``globals()`` for the name: | |
|
134 | ||
|
135 | .. sourcecode:: python | |
|
136 | ||
|
137 | def g(b): | |
|
138 | # note that `b` is not referenced in inner's scope | |
|
139 | def inner(): | |
|
140 | # this inner will *not* have a closure | |
|
141 | return a | |
|
142 | return echo | |
|
143 | g1 = g(1) | |
|
144 | g2 = g(2) | |
|
145 | g1() # raises NameError on 'a' | |
|
146 | a=5 | |
|
147 | g2() # returns 5 | |
|
148 | ||
|
149 | `g1` and `g2` *will* be sendable with IPython, and will treat the engine's namespace as | |
|
150 | globals(). The :meth:`pull` method is implemented based on this principal. If we did not | |
|
151 | provide pull, you could implement it yourself with `apply`, by simply returning objects out | |
|
152 | of the global namespace: | |
|
153 | ||
|
154 | .. sourcecode:: ipython | |
|
155 | ||
|
156 | In [10]: view.apply(lambda : a) | |
|
157 | ||
|
158 | # is equivalent to | |
|
159 | In [11]: view.pull('a') | |
|
86 | 160 | |
|
87 | 161 | Running Code |
|
88 | 162 | ============ |
|
89 | 163 | |
|
90 | 164 | There are two principal units of execution in Python: strings of Python code (e.g. 'a=5'), |
|
91 | 165 | and Python functions. IPython is designed around the use of functions via the core |
|
92 | 166 | Client method, called `apply`. |
|
93 | 167 | |
|
94 | 168 | Apply |
|
95 | 169 | ----- |
|
96 | 170 | |
|
97 |
The principal method of remote execution is :meth:`apply`, of |
|
|
171 | The principal method of remote execution is :meth:`apply`, of View objects. The Client provides | |
|
172 | the full execution and communication API for engines via its low-level | |
|
173 | :meth:`send_apply_message` method. | |
|
98 | 174 | |
|
99 | 175 | f : function |
|
100 | 176 | The fuction to be called remotely |
|
101 | 177 | args : tuple/list |
|
102 | 178 | The positional arguments passed to `f` |
|
103 | 179 | kwargs : dict |
|
104 | 180 | The keyword arguments passed to `f` |
|
105 | bound : bool (default: False) | |
|
106 | Whether to pass the Engine(s) Namespace as the first argument to `f`. | |
|
107 | 181 | block : bool (default: self.block) |
|
108 | 182 | Whether to wait for the result, or return immediately. |
|
109 | 183 | False: |
|
110 | 184 | returns AsyncResult |
|
111 | 185 | True: |
|
112 | 186 | returns actual result(s) of f(*args, **kwargs) |
|
113 | 187 | if multiple targets: |
|
114 | 188 | list of results, matching `targets` |
|
115 | 189 | track : bool |
|
116 | 190 | whether to track non-copying sends. |
|
117 | 191 | [default False] |
|
118 | 192 | |
|
119 | 193 | targets : int,list of ints, 'all', None |
|
120 | 194 | Specify the destination of the job. |
|
121 | 195 | if None: |
|
122 | 196 | Submit via Task queue for load-balancing. |
|
123 | 197 | if 'all': |
|
124 | 198 | Run on all active engines |
|
125 | 199 | if list: |
|
126 | 200 | Run on each specified engine |
|
127 | 201 | if int: |
|
128 | 202 | Run on single engine |
|
129 | 203 | Not eht |
|
130 | 204 | |
|
131 | 205 | balanced : bool, default None |
|
132 | 206 | whether to load-balance. This will default to True |
|
133 | 207 | if targets is unspecified, or False if targets is specified. |
|
134 | 208 | |
|
135 | 209 | If `balanced` and `targets` are both specified, the task will |
|
136 | 210 | be assigne to *one* of the targets by the scheduler. |
|
137 | 211 | |
|
138 | The following arguments are only used when balanced is True: | |
|
139 | ||
|
140 | 212 | after : Dependency or collection of msg_ids |
|
141 | 213 | Only for load-balanced execution (targets=None) |
|
142 | 214 | Specify a list of msg_ids as a time-based dependency. |
|
143 | 215 | This job will only be run *after* the dependencies |
|
144 | 216 | have been met. |
|
145 | 217 | |
|
146 | 218 | follow : Dependency or collection of msg_ids |
|
147 | 219 | Only for load-balanced execution (targets=None) |
|
148 | 220 | Specify a list of msg_ids as a location-based dependency. |
|
149 | 221 | This job will only be run on an engine where this dependency |
|
150 | 222 | is met. |
|
151 | 223 | |
|
152 | 224 | timeout : float/int or None |
|
153 | 225 | Only for load-balanced execution (targets=None) |
|
154 | 226 | Specify an amount of time (in seconds) for the scheduler to |
|
155 | 227 | wait for dependencies to be met before failing with a |
|
156 | 228 | DependencyTimeout. |
|
157 | 229 | |
|
158 | 230 | execute and run |
|
159 | 231 | --------------- |
|
160 | 232 | |
|
161 |
For executing strings of Python code, |
|
|
162 |
method, which rather than take functions and arguments, take simple strings. |
|
|
163 |
takes a string of Python code to execute, and sends it to the Engine(s). `run` |
|
|
164 |
`execute`, but for a *file*, rather than a string. It is simply a wrapper that |
|
|
165 | very similar to ``execute(open(f).read())``. | |
|
233 | For executing strings of Python code, :class:`DirectView`s also provide an :meth:`execute` and a | |
|
234 | :meth:`run` method, which rather than take functions and arguments, take simple strings. | |
|
235 | `execute` simply takes a string of Python code to execute, and sends it to the Engine(s). `run` | |
|
236 | is the same as `execute`, but for a *file*, rather than a string. It is simply a wrapper that | |
|
237 | does something very similar to ``execute(open(f).read())``. | |
|
166 | 238 | |
|
167 | 239 | .. note:: |
|
168 | 240 | |
|
169 | 241 | TODO: Example |
|
170 | 242 | |
|
171 | 243 | Views |
|
172 | 244 | ===== |
|
173 | 245 | |
|
174 | 246 | The principal extension of the :class:`~parallel.client.Client` is the |
|
175 |
:class:`~parallel.view.View` class. The client |
|
|
176 | execution patterns, where you must specify everything about the execution as keywords to each | |
|
177 | call to :meth:`apply`. For users who want to more conveniently specify various options for | |
|
178 | several similar calls, we have the :class:`~parallel.view.View` objects. The basic principle of | |
|
179 | the views is to encapsulate the keyword arguments to :meth:`client.apply` as attributes, | |
|
180 | allowing users to specify them once and apply to any subsequent calls until the attribute is | |
|
181 | changed. | |
|
247 | :class:`~parallel.view.View` class. The client | |
|
182 | 248 | |
|
183 | 249 | Two of apply's keyword arguments are set at the construction of the View, and are immutable for |
|
184 | 250 | a given View: `balanced` and `targets`. `balanced` determines whether the View will be a |
|
185 | 251 | :class:`.LoadBalancedView` or a :class:`.DirectView`, and `targets` will be the View's `targets` |
|
186 | 252 | attribute. Attempts to change this will raise errors. |
|
187 | 253 | |
|
188 |
Views are cached by targets |
|
|
254 | Views are cached by targets/class, so requesting a view multiple times will always return the | |
|
255 | *same object*, not create a new one: | |
|
189 | 256 | |
|
190 | 257 | .. sourcecode:: ipython |
|
191 | 258 | |
|
192 |
In [3]: v1 = rc.view([1,2,3] |
|
|
193 |
In [4]: v2 = rc.view([1,2,3] |
|
|
259 | In [3]: v1 = rc.load_balanced_view([1,2,3]) | |
|
260 | In [4]: v2 = rc.load_balanced_view([1,2,3]) | |
|
194 | 261 | |
|
195 | 262 | In [5]: v2 is v1 |
|
196 | 263 | Out[5]: True |
|
197 | 264 | |
|
198 | 265 | |
|
199 | A :class:`View` always uses its `targets` attribute, and it will use its `bound` | |
|
200 | and `block` attributes in its :meth:`apply` method, but the suffixed :meth:`apply_x` | |
|
201 | methods allow overriding `bound` and `block` for a single call. | |
|
202 | ||
|
203 | ================== ========== ========== | |
|
204 | method block bound | |
|
205 | ================== ========== ========== | |
|
206 | apply self.block self.bound | |
|
207 | apply_sync True False | |
|
208 | apply_async False False | |
|
209 | apply_sync_bound True True | |
|
210 | apply_async_bound False True | |
|
211 | ================== ========== ========== | |
|
212 | ||
|
213 | 266 | DirectView |
|
214 | 267 | ---------- |
|
215 | 268 | |
|
216 | 269 | The :class:`.DirectView` is the class for the IPython :ref:`Multiplexing Interface |
|
217 | 270 | <parallel_multiengine>`. |
|
218 | 271 | |
|
219 | 272 | Creating a DirectView |
|
220 | 273 | ********************* |
|
221 | 274 | |
|
222 | 275 | DirectViews can be created in two ways, by index access to a client, or by a client's |
|
223 | 276 | :meth:`view` method. Index access to a Client works in a few ways. First, you can create |
|
224 | 277 | DirectViews to single engines simply by accessing the client by engine id: |
|
225 | 278 | |
|
226 | 279 | .. sourcecode:: ipython |
|
227 | 280 | |
|
228 | 281 | In [2]: rc[0] |
|
229 | 282 | Out[2]: <DirectView 0> |
|
230 | 283 | |
|
231 | 284 | You can also create a DirectView with a list of engines: |
|
232 | 285 | |
|
233 | 286 | .. sourcecode:: ipython |
|
234 | 287 | |
|
235 | 288 | In [2]: rc[0,1,2] |
|
236 | 289 | Out[2]: <DirectView [0,1,2]> |
|
237 | 290 | |
|
238 | 291 | Other methods for accessing elements, such as slicing and negative indexing, work by passing |
|
239 | 292 | the index directly to the client's :attr:`ids` list, so: |
|
240 | 293 | |
|
241 | 294 | .. sourcecode:: ipython |
|
242 | 295 | |
|
243 | 296 | # negative index |
|
244 | 297 | In [2]: rc[-1] |
|
245 | 298 | Out[2]: <DirectView 3> |
|
246 | 299 | |
|
247 | 300 | # or slicing: |
|
248 | 301 | In [3]: rc[::2] |
|
249 | 302 | Out[3]: <DirectView [0,2]> |
|
250 | 303 | |
|
251 | 304 | are always the same as: |
|
252 | 305 | |
|
253 | 306 | .. sourcecode:: ipython |
|
254 | 307 | |
|
255 | 308 | In [2]: rc[rc.ids[-1]] |
|
256 | 309 | Out[2]: <DirectView 3> |
|
257 | 310 | |
|
258 | 311 | In [3]: rc[rc.ids[::2]] |
|
259 | 312 | Out[3]: <DirectView [0,2]> |
|
260 | 313 | |
|
261 | 314 | Also note that the slice is evaluated at the time of construction of the DirectView, so the |
|
262 | 315 | targets will not change over time if engines are added/removed from the cluster. Requesting |
|
263 | 316 | two views with the same slice at different times will *not* necessarily return the same View |
|
264 | 317 | if the number of engines has changed. |
|
265 | 318 | |
|
266 | 319 | Execution via DirectView |
|
267 | 320 | ************************ |
|
268 | 321 | |
|
269 | 322 | The DirectView is the simplest way to work with one or more engines directly (hence the name). |
|
270 | 323 | |
|
271 | 324 | |
|
272 | 325 | Data movement via DirectView |
|
273 | 326 | **************************** |
|
274 | 327 | |
|
275 | 328 | Since a Python namespace is just a :class:`dict`, :class:`DirectView` objects provide |
|
276 | 329 | dictionary-style access by key and methods such as :meth:`get` and |
|
277 | 330 | :meth:`update` for convenience. This make the remote namespaces of the engines |
|
278 | 331 | appear as a local dictionary. Underneath, these methods call :meth:`apply`: |
|
279 | 332 | |
|
280 | 333 | .. sourcecode:: ipython |
|
281 | 334 | |
|
282 | 335 | In [51]: dview['a']=['foo','bar'] |
|
283 | 336 | |
|
284 | 337 | In [52]: dview['a'] |
|
285 | 338 | Out[52]: [ ['foo', 'bar'], ['foo', 'bar'], ['foo', 'bar'], ['foo', 'bar'] ] |
|
286 | 339 | |
|
287 | 340 | Scatter and gather |
|
288 | 341 | ------------------ |
|
289 | 342 | |
|
290 | 343 | Sometimes it is useful to partition a sequence and push the partitions to |
|
291 | 344 | different engines. In MPI language, this is know as scatter/gather and we |
|
292 | 345 | follow that terminology. However, it is important to remember that in |
|
293 | 346 | IPython's :class:`Client` class, :meth:`scatter` is from the |
|
294 | 347 | interactive IPython session to the engines and :meth:`gather` is from the |
|
295 | 348 | engines back to the interactive IPython session. For scatter/gather operations |
|
296 | 349 | between engines, MPI should be used: |
|
297 | 350 | |
|
298 | 351 | .. sourcecode:: ipython |
|
299 | 352 | |
|
300 | 353 | In [58]: dview.scatter('a',range(16)) |
|
301 | 354 | Out[58]: [None,None,None,None] |
|
302 | 355 | |
|
303 | 356 | In [59]: dview['a'] |
|
304 | 357 | Out[59]: [ [0, 1, 2, 3], [4, 5, 6, 7], [8, 9, 10, 11], [12, 13, 14, 15] ] |
|
305 | 358 | |
|
306 | 359 | In [60]: dview.gather('a') |
|
307 | 360 | Out[60]: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15] |
|
308 | 361 | |
|
309 | 362 | |
|
310 | 363 | |
|
311 | 364 | LoadBalancedView |
|
312 | 365 | ---------------- |
|
313 | 366 | |
|
314 | 367 | The :class:`.LoadBalancedView` |
|
315 | 368 | |
|
316 | 369 | |
|
317 | 370 | Data Movement |
|
318 | 371 | ============= |
|
319 | 372 | |
|
320 | 373 | push |
|
321 | 374 | |
|
322 | 375 | pull |
|
323 | 376 | |
|
324 | 377 | Reference |
|
325 | 378 | |
|
326 | 379 | Results |
|
327 | 380 | ======= |
|
328 | 381 | |
|
329 | 382 | AsyncResults are the primary class |
|
330 | 383 | |
|
331 | 384 | get_result |
|
332 | 385 | |
|
333 | 386 | results,metadata |
|
334 | 387 | |
|
335 | 388 | Querying the Hub |
|
336 | 389 | ================ |
|
337 | 390 | |
|
338 | 391 | The Hub sees all traffic that may pass through the schedulers between engines and clients. |
|
339 | 392 | It does this so that it can track state, allowing multiple clients to retrieve results of |
|
340 | 393 | computations submitted by their peers, as well as persisting the state to a database. |
|
341 | 394 | |
|
342 | 395 | queue_status |
|
343 | 396 | |
|
344 | 397 | You can check the status of the queues of the engines with this command. |
|
345 | 398 | |
|
346 | 399 | result_status |
|
347 | 400 | |
|
348 | 401 | purge_results |
|
349 | 402 | |
|
350 | 403 | Controlling the Engines |
|
351 | 404 | ======================= |
|
352 | 405 | |
|
353 | 406 | There are a few actions you can do with Engines that do not involve execution. These |
|
354 | 407 | messages are sent via the Control socket, and bypass any long queues of waiting execution |
|
355 | 408 | jobs |
|
356 | 409 | |
|
357 | 410 | abort |
|
358 | 411 | |
|
359 | 412 | Sometimes you may want to prevent a job you have submitted from actually running. The method |
|
360 | 413 | for this is :meth:`abort`. It takes a container of msg_ids, and instructs the Engines to not |
|
361 | 414 | run the jobs if they arrive. The jobs will then fail with an AbortedTask error. |
|
362 | 415 | |
|
363 | 416 | clear |
|
364 | 417 | |
|
365 | 418 | You may want to purge the Engine(s) namespace of any data you have left in it. After |
|
366 | 419 | running `clear`, there will be no names in the Engine's namespace |
|
367 | 420 | |
|
368 | 421 | shutdown |
|
369 | 422 | |
|
370 | 423 | You can also instruct engines (and the Controller) to terminate from a Client. This |
|
371 | 424 | can be useful when a job is finished, since you can shutdown all the processes with a |
|
372 | 425 | single command. |
|
373 | 426 | |
|
374 | 427 | Synchronization |
|
375 | 428 | =============== |
|
376 | 429 | |
|
377 | 430 | Since the Client is a synchronous object, events do not automatically trigger in your |
|
378 | 431 | interactive session - you must poll the 0MQ sockets for incoming messages. Note that |
|
379 | 432 | this polling *does not* actually make any network requests. It simply performs a `select` |
|
380 | 433 | operation, to check if messages are already in local memory, waiting to be handled. |
|
381 | 434 | |
|
382 |
The method that handles incoming messages is :meth:`spin`. |
|
|
435 | The method that handles incoming messages is :meth:`spin`. This method flushes any waiting | |
|
436 | messages on the various incoming sockets, and updates the state of the Client. | |
|
383 | 437 | |
|
384 |
If you need to wait for particular results to finish, you can use the :meth:` |
|
|
438 | If you need to wait for particular results to finish, you can use the :meth:`wait` method, | |
|
385 | 439 | which will call :meth:`spin` until the messages are no longer outstanding. Anything that |
|
386 | 440 | represents a collection of messages, such as a list of msg_ids or one or more AsyncResult |
|
387 |
objects, can be passed as argument to |
|
|
388 |
the |
|
|
441 | objects, can be passed as argument to wait. A timeout can be specified, which will prevent | |
|
442 | the call from blocking for more than a specified time, but the default behavior is to wait | |
|
389 | 443 | forever. |
|
390 | 444 | |
|
391 | 445 | |
|
392 | 446 | |
|
393 | 447 | The client also has an `outstanding` attribute - a ``set`` of msg_ids that are awaiting replies. |
|
394 |
This is the default if |
|
|
448 | This is the default if wait is called with no arguments - i.e. wait on *all* outstanding | |
|
449 | messages. | |
|
395 | 450 | |
|
396 | 451 | |
|
397 | 452 | .. note:: |
|
398 | 453 | |
|
399 |
TODO |
|
|
454 | TODO wait example | |
|
400 | 455 | |
|
401 | 456 | Map |
|
402 | 457 | === |
|
403 | 458 | |
|
404 | 459 | Many parallel computing problems can be expressed as a `map`, or running a single program with a |
|
405 | 460 | variety of different inputs. Python has a built-in :py-func:`map`, which does exactly this, and |
|
406 | 461 | many parallel execution tools in Python, such as the built-in :py-class:`multiprocessing.Pool` |
|
407 | 462 | object provide implementations of `map`. All View objects provide a :meth:`map` method as well, |
|
408 | 463 | but the load-balanced and direct implementations differ. |
|
409 | 464 | |
|
410 | 465 | Views' map methods can be called on any number of sequences, but they can also take the `block` |
|
411 | 466 | and `bound` keyword arguments, just like :meth:`~client.apply`, but *only as keywords*. |
|
412 | 467 | |
|
413 | 468 | .. sourcecode:: python |
|
414 | 469 | |
|
415 | 470 | dview.map(*sequences, block=None) |
|
416 | 471 | |
|
417 | 472 | |
|
418 | 473 | * iter, map_async, reduce |
|
419 | 474 | |
|
420 | 475 | Decorators and RemoteFunctions |
|
421 | 476 | ============================== |
|
422 | 477 | |
|
423 | 478 | @parallel |
|
424 | 479 | |
|
425 | 480 | @remote |
|
426 | 481 | |
|
427 | 482 | RemoteFunction |
|
428 | 483 | |
|
429 | 484 | ParallelFunction |
|
430 | 485 | |
|
431 | 486 | Dependencies |
|
432 | 487 | ============ |
|
433 | 488 | |
|
434 | 489 | @depend |
|
435 | 490 | |
|
436 | 491 | @require |
|
437 | 492 | |
|
438 | 493 | Dependency |
@@ -1,243 +1,244 b'' | |||
|
1 | 1 | .. _ip1par: |
|
2 | 2 | |
|
3 | 3 | ============================ |
|
4 | 4 | Overview and getting started |
|
5 | 5 | ============================ |
|
6 | 6 | |
|
7 | 7 | Introduction |
|
8 | 8 | ============ |
|
9 | 9 | |
|
10 | 10 | This section gives an overview of IPython's sophisticated and powerful |
|
11 | 11 | architecture for parallel and distributed computing. This architecture |
|
12 | 12 | abstracts out parallelism in a very general way, which enables IPython to |
|
13 | 13 | support many different styles of parallelism including: |
|
14 | 14 | |
|
15 | 15 | * Single program, multiple data (SPMD) parallelism. |
|
16 | 16 | * Multiple program, multiple data (MPMD) parallelism. |
|
17 | 17 | * Message passing using MPI. |
|
18 | 18 | * Task farming. |
|
19 | 19 | * Data parallel. |
|
20 | 20 | * Combinations of these approaches. |
|
21 | 21 | * Custom user defined approaches. |
|
22 | 22 | |
|
23 | 23 | Most importantly, IPython enables all types of parallel applications to |
|
24 | 24 | be developed, executed, debugged and monitored *interactively*. Hence, |
|
25 | 25 | the ``I`` in IPython. The following are some example usage cases for IPython: |
|
26 | 26 | |
|
27 | 27 | * Quickly parallelize algorithms that are embarrassingly parallel |
|
28 | 28 | using a number of simple approaches. Many simple things can be |
|
29 | 29 | parallelized interactively in one or two lines of code. |
|
30 | 30 | |
|
31 | 31 | * Steer traditional MPI applications on a supercomputer from an |
|
32 | 32 | IPython session on your laptop. |
|
33 | 33 | |
|
34 | 34 | * Analyze and visualize large datasets (that could be remote and/or |
|
35 | 35 | distributed) interactively using IPython and tools like |
|
36 | 36 | matplotlib/TVTK. |
|
37 | 37 | |
|
38 | 38 | * Develop, test and debug new parallel algorithms |
|
39 | 39 | (that may use MPI) interactively. |
|
40 | 40 | |
|
41 | 41 | * Tie together multiple MPI jobs running on different systems into |
|
42 | 42 | one giant distributed and parallel system. |
|
43 | 43 | |
|
44 | 44 | * Start a parallel job on your cluster and then have a remote |
|
45 | 45 | collaborator connect to it and pull back data into their |
|
46 | 46 | local IPython session for plotting and analysis. |
|
47 | 47 | |
|
48 | 48 | * Run a set of tasks on a set of CPUs using dynamic load balancing. |
|
49 | 49 | |
|
50 | 50 | Architecture overview |
|
51 | 51 | ===================== |
|
52 | 52 | |
|
53 | 53 | The IPython architecture consists of four components: |
|
54 | 54 | |
|
55 | 55 | * The IPython engine. |
|
56 | 56 | * The IPython hub. |
|
57 | 57 | * The IPython schedulers. |
|
58 | 58 | * The controller client. |
|
59 | 59 | |
|
60 | 60 | These components live in the :mod:`IPython.zmq.parallel` package and are |
|
61 | 61 | installed with IPython. They do, however, have additional dependencies |
|
62 | 62 | that must be installed. For more information, see our |
|
63 | 63 | :ref:`installation documentation <install_index>`. |
|
64 | 64 | |
|
65 | 65 | .. TODO: include zmq in install_index |
|
66 | 66 | |
|
67 | 67 | IPython engine |
|
68 | 68 | --------------- |
|
69 | 69 | |
|
70 | 70 | The IPython engine is a Python instance that takes Python commands over a |
|
71 | 71 | network connection. Eventually, the IPython engine will be a full IPython |
|
72 | 72 | interpreter, but for now, it is a regular Python interpreter. The engine |
|
73 | 73 | can also handle incoming and outgoing Python objects sent over a network |
|
74 | 74 | connection. When multiple engines are started, parallel and distributed |
|
75 | 75 | computing becomes possible. An important feature of an IPython engine is |
|
76 | 76 | that it blocks while user code is being executed. Read on for how the |
|
77 | 77 | IPython controller solves this problem to expose a clean asynchronous API |
|
78 | 78 | to the user. |
|
79 | 79 | |
|
80 | 80 | IPython controller |
|
81 | 81 | ------------------ |
|
82 | 82 | |
|
83 | 83 | The IPython controller processes provide an interface for working with a set of engines. |
|
84 | 84 | At a general level, the controller is a collection of processes to which IPython engines |
|
85 | 85 | and clients can connect. The controller is composed of a :class:`Hub` and a collection of |
|
86 | 86 | :class:`Schedulers`. These Schedulers are typically run in separate processes but on the |
|
87 | 87 | same machine as the Hub, but can be run anywhere from local threads or on remote machines. |
|
88 | 88 | |
|
89 | 89 | The controller also provides a single point of contact for users who wish to |
|
90 | 90 | utilize the engines connected to the controller. There are different ways of |
|
91 | 91 | working with a controller. In IPython, all of these models are implemented via |
|
92 |
the client's :meth:`. |
|
|
92 | the client's :meth:`.View.apply` method, with various arguments, or | |
|
93 | 93 | constructing :class:`.View` objects to represent subsets of engines. The two |
|
94 | 94 | primary models for interacting with engines are: |
|
95 | 95 | |
|
96 | 96 | * A **Direct** interface, where engines are addressed explicitly. |
|
97 | 97 | * A **LoadBalanced** interface, where the Scheduler is trusted with assigning work to |
|
98 | 98 | appropriate engines. |
|
99 | 99 | |
|
100 | 100 | Advanced users can readily extend the View models to enable other |
|
101 | 101 | styles of parallelism. |
|
102 | 102 | |
|
103 | 103 | .. note:: |
|
104 | 104 | |
|
105 | 105 | A single controller and set of engines can be used with multiple models |
|
106 | 106 | simultaneously. This opens the door for lots of interesting things. |
|
107 | 107 | |
|
108 | 108 | |
|
109 | 109 | The Hub |
|
110 | 110 | ******* |
|
111 | 111 | |
|
112 | 112 | The center of an IPython cluster is the Hub. This is the process that keeps |
|
113 | 113 | track of engine connections, schedulers, clients, as well as all task requests and |
|
114 | 114 | results. The primary role of the Hub is to facilitate queries of the cluster state, and |
|
115 | 115 | minimize the necessary information required to establish the many connections involved in |
|
116 | 116 | connecting new clients and engines. |
|
117 | 117 | |
|
118 | 118 | |
|
119 | 119 | Schedulers |
|
120 | 120 | ********** |
|
121 | 121 | |
|
122 | 122 | All actions that can be performed on the engine go through a Scheduler. While the engines |
|
123 | 123 | themselves block when user code is run, the schedulers hide that from the user to provide |
|
124 | 124 | a fully asynchronous interface to a set of engines. |
|
125 | 125 | |
|
126 | 126 | |
|
127 | IPython client | |
|
128 | -------------- | |
|
127 | IPython client and views | |
|
128 | ------------------------ | |
|
129 | 129 | |
|
130 | There is one primary object, the :class:`~.parallel.client.Client`, for connecting to a | |
|
131 |
|
|
|
132 |
interact with a set of engines through the interface. Here are the two default |
|
|
130 | There is one primary object, the :class:`~.parallel.client.Client`, for connecting to a cluster. | |
|
131 | For each execution model, there is a corresponding :class:`~.parallel.view.View`. These views | |
|
132 | allow users to interact with a set of engines through the interface. Here are the two default | |
|
133 | views: | |
|
133 | 134 | |
|
134 | 135 | * The :class:`DirectView` class for explicit addressing. |
|
135 | 136 | * The :class:`LoadBalancedView` class for destination-agnostic scheduling. |
|
136 | 137 | |
|
137 | 138 | Security |
|
138 | 139 | -------- |
|
139 | 140 | |
|
140 | 141 | IPython uses ZeroMQ for networking, which has provided many advantages, but |
|
141 | 142 | one of the setbacks is its utter lack of security [ZeroMQ]_. By default, no IPython |
|
142 | 143 | connections are encrypted, but open ports only listen on localhost. The only |
|
143 | 144 | source of security for IPython is via ssh-tunnel. IPython supports both shell |
|
144 | 145 | (`openssh`) and `paramiko` based tunnels for connections. There is a key necessary |
|
145 | 146 | to submit requests, but due to the lack of encryption, it does not provide |
|
146 | 147 | significant security if loopback traffic is compromised. |
|
147 | 148 | |
|
148 | 149 | In our architecture, the controller is the only process that listens on |
|
149 | 150 | network ports, and is thus the main point of vulnerability. The standard model |
|
150 | 151 | for secure connections is to designate that the controller listen on |
|
151 | 152 | localhost, and use ssh-tunnels to connect clients and/or |
|
152 | 153 | engines. |
|
153 | 154 | |
|
154 | 155 | To connect and authenticate to the controller an engine or client needs |
|
155 | 156 | some information that the controller has stored in a JSON file. |
|
156 | 157 | Thus, the JSON files need to be copied to a location where |
|
157 | 158 | the clients and engines can find them. Typically, this is the |
|
158 | 159 | :file:`~/.ipython/clusterz_default/security` directory on the host where the |
|
159 | 160 | client/engine is running (which could be a different host than the controller). |
|
160 | 161 | Once the JSON files are copied over, everything should work fine. |
|
161 | 162 | |
|
162 | 163 | Currently, there are two JSON files that the controller creates: |
|
163 | 164 | |
|
164 | 165 | ipcontroller-engine.json |
|
165 | 166 | This JSON file has the information necessary for an engine to connect |
|
166 | 167 | to a controller. |
|
167 | 168 | |
|
168 | 169 | ipcontroller-client.json |
|
169 | 170 | The client's connection information. This may not differ from the engine's, |
|
170 | 171 | but since the controller may listen on different ports for clients and |
|
171 | 172 | engines, it is stored separately. |
|
172 | 173 | |
|
173 | 174 | More details of how these JSON files are used are given below. |
|
174 | 175 | |
|
175 | 176 | A detailed description of the security model and its implementation in IPython |
|
176 | 177 | can be found :ref:`here <parallelsecurity>`. |
|
177 | 178 | |
|
178 | 179 | .. warning:: |
|
179 | 180 | |
|
180 | 181 | Even at its most secure, the Controller listens on ports on localhost, and |
|
181 | 182 | every time you make a tunnel, you open a localhost port on the connecting |
|
182 | 183 | machine that points to the Controller. If localhost on the Controller's |
|
183 | 184 | machine, or the machine of any client or engine, is untrusted, then your |
|
184 | 185 | Controller is insecure. There is no way around this with ZeroMQ. |
|
185 | 186 | |
|
186 | 187 | |
|
187 | 188 | |
|
188 | 189 | Getting Started |
|
189 | 190 | =============== |
|
190 | 191 | |
|
191 | 192 | To use IPython for parallel computing, you need to start one instance of the |
|
192 | 193 | controller and one or more instances of the engine. Initially, it is best to |
|
193 | 194 | simply start a controller and engines on a single host using the |
|
194 | 195 | :command:`ipclusterz` command. To start a controller and 4 engines on your |
|
195 | 196 | localhost, just do:: |
|
196 | 197 | |
|
197 | 198 | $ ipclusterz start -n 4 |
|
198 | 199 | |
|
199 | 200 | More details about starting the IPython controller and engines can be found |
|
200 | 201 | :ref:`here <parallel_process>` |
|
201 | 202 | |
|
202 | 203 | Once you have started the IPython controller and one or more engines, you |
|
203 | 204 | are ready to use the engines to do something useful. To make sure |
|
204 | 205 | everything is working correctly, try the following commands: |
|
205 | 206 | |
|
206 | 207 | .. sourcecode:: ipython |
|
207 | 208 | |
|
208 | 209 | In [1]: from IPython.zmq.parallel import client |
|
209 | 210 | |
|
210 | 211 | In [2]: c = client.Client() |
|
211 | 212 | |
|
212 | 213 | In [4]: c.ids |
|
213 | 214 | Out[4]: set([0, 1, 2, 3]) |
|
214 | 215 | |
|
215 |
In [5]: c.apply(lambda : "Hello, World" |
|
|
216 | In [5]: c[:].apply_sync(lambda : "Hello, World") | |
|
216 | 217 | Out[5]: [ 'Hello, World', 'Hello, World', 'Hello, World', 'Hello, World' ] |
|
217 | 218 | |
|
218 | 219 | |
|
219 | 220 | When a client is created with no arguments, the client tries to find the corresponding |
|
220 | 221 | JSON file in the local `~/.ipython/clusterz_default/security` directory. If it finds it, |
|
221 | 222 | you are set. If you have put the JSON file in a different location or it has a different |
|
222 | 223 | name, create the client like this: |
|
223 | 224 | |
|
224 | 225 | .. sourcecode:: ipython |
|
225 | 226 | |
|
226 | 227 | In [2]: c = client.Client('/path/to/my/ipcontroller-client.json') |
|
227 | 228 | |
|
228 | 229 | Remember, a client needs to be able to see the Hub's ports to connect. So if they are on a |
|
229 | 230 | different machine, you may need to use an ssh server to tunnel access to that machine, |
|
230 | 231 | then you would connect to it with: |
|
231 | 232 | |
|
232 | 233 | .. sourcecode:: ipython |
|
233 | 234 | |
|
234 | 235 | In [2]: c = client.Client(sshserver='myhub.example.com') |
|
235 | 236 | |
|
236 | 237 | Where 'myhub.example.com' is the url or IP address of the machine on |
|
237 | which the Hub process is running. | |
|
238 | which the Hub process is running (or another machine that has direct access to the Hub's ports). | |
|
238 | 239 | |
|
239 | 240 | You are now ready to learn more about the :ref:`Direct |
|
240 | <parallelmultiengine>` and :ref:`LoadBalanced <paralleltask>` interfaces to the | |
|
241 | <parallel_multiengine>` and :ref:`LoadBalanced <parallel_task>` interfaces to the | |
|
241 | 242 | controller. |
|
242 | 243 | |
|
243 | 244 | .. [ZeroMQ] ZeroMQ. http://www.zeromq.org |
@@ -1,845 +1,799 b'' | |||
|
1 | .. _parallelmultiengine: | |
|
1 | .. _parallel_multiengine: | |
|
2 | 2 | |
|
3 | 3 | ========================== |
|
4 | 4 | IPython's Direct interface |
|
5 | 5 | ========================== |
|
6 | 6 | |
|
7 | 7 | The direct, or multiengine, interface represents one possible way of working with a set of |
|
8 | 8 | IPython engines. The basic idea behind the multiengine interface is that the |
|
9 | 9 | capabilities of each engine are directly and explicitly exposed to the user. |
|
10 | 10 | Thus, in the multiengine interface, each engine is given an id that is used to |
|
11 | 11 | identify the engine and give it work to do. This interface is very intuitive |
|
12 |
and is designed with interactive usage in mind, and is th |
|
|
12 | and is designed with interactive usage in mind, and is the best place for | |
|
13 | 13 | new users of IPython to begin. |
|
14 | 14 | |
|
15 | 15 | Starting the IPython controller and engines |
|
16 | 16 | =========================================== |
|
17 | 17 | |
|
18 | 18 | To follow along with this tutorial, you will need to start the IPython |
|
19 | 19 | controller and four IPython engines. The simplest way of doing this is to use |
|
20 | 20 | the :command:`ipclusterz` command:: |
|
21 | 21 | |
|
22 | 22 | $ ipclusterz start -n 4 |
|
23 | 23 | |
|
24 | 24 | For more detailed information about starting the controller and engines, see |
|
25 | 25 | our :ref:`introduction <ip1par>` to using IPython for parallel computing. |
|
26 | 26 | |
|
27 | 27 | Creating a ``Client`` instance |
|
28 | 28 | ============================== |
|
29 | 29 | |
|
30 | 30 | The first step is to import the IPython :mod:`IPython.zmq.parallel.client` |
|
31 | 31 | module and then create a :class:`.Client` instance: |
|
32 | 32 | |
|
33 | 33 | .. sourcecode:: ipython |
|
34 | 34 | |
|
35 | 35 | In [1]: from IPython.zmq.parallel import client |
|
36 | 36 | |
|
37 | 37 | In [2]: rc = client.Client() |
|
38 | 38 | |
|
39 | 39 | This form assumes that the default connection information (stored in |
|
40 | 40 | :file:`ipcontroller-client.json` found in :file:`IPYTHON_DIR/clusterz_default/security`) is |
|
41 | 41 | accurate. If the controller was started on a remote machine, you must copy that connection |
|
42 | 42 | file to the client machine, or enter its contents as arguments to the Client constructor: |
|
43 | 43 | |
|
44 | 44 | .. sourcecode:: ipython |
|
45 | 45 | |
|
46 | 46 | # If you have copied the json connector file from the controller: |
|
47 | 47 | In [2]: rc = client.Client('/path/to/ipcontroller-client.json') |
|
48 | 48 | # or to connect with a specific profile you have set up: |
|
49 | 49 | In [3]: rc = client.Client(profile='mpi') |
|
50 | 50 | |
|
51 | 51 | |
|
52 | 52 | To make sure there are engines connected to the controller, users can get a list |
|
53 | 53 | of engine ids: |
|
54 | 54 | |
|
55 | 55 | .. sourcecode:: ipython |
|
56 | 56 | |
|
57 | 57 | In [3]: rc.ids |
|
58 | 58 | Out[3]: [0, 1, 2, 3] |
|
59 | 59 | |
|
60 | 60 | Here we see that there are four engines ready to do work for us. |
|
61 | 61 | |
|
62 | 62 | For direct execution, we will make use of a :class:`DirectView` object, which can be |
|
63 | 63 | constructed via list-access to the client: |
|
64 | 64 | |
|
65 | 65 | .. sourcecode:: ipython |
|
66 | 66 | |
|
67 | 67 | In [4]: dview = rc[:] # use all engines |
|
68 | 68 | |
|
69 | 69 | .. seealso:: |
|
70 | 70 | |
|
71 | 71 | For more information, see the in-depth explanation of :ref:`Views <parallel_details>`. |
|
72 | 72 | |
|
73 | 73 | |
|
74 | 74 | Quick and easy parallelism |
|
75 | 75 | ========================== |
|
76 | 76 | |
|
77 | 77 | In many cases, you simply want to apply a Python function to a sequence of |
|
78 | 78 | objects, but *in parallel*. The client interface provides a simple way |
|
79 | 79 | of accomplishing this: using the DirectView's :meth:`~DirectView.map` method. |
|
80 | 80 | |
|
81 | 81 | Parallel map |
|
82 | 82 | ------------ |
|
83 | 83 | |
|
84 | 84 | Python's builtin :func:`map` functions allows a function to be applied to a |
|
85 | 85 | sequence element-by-element. This type of code is typically trivial to |
|
86 | 86 | parallelize. In fact, since IPython's interface is all about functions anyway, |
|
87 | 87 | you can just use the builtin :func:`map` with a :class:`RemoteFunction`, or a |
|
88 | 88 | DirectView's :meth:`map` method: |
|
89 | 89 | |
|
90 | 90 | .. sourcecode:: ipython |
|
91 | 91 | |
|
92 | 92 | In [62]: serial_result = map(lambda x:x**10, range(32)) |
|
93 | 93 | |
|
94 | In [63]: dview.block = True | |
|
95 | ||
|
96 | In [66]: parallel_result = dview.map(lambda x: x**10, range(32)) | |
|
94 | In [63]: parallel_result = dview.map_sync(lambda x: x**10, range(32)) | |
|
97 | 95 | |
|
98 | 96 | In [67]: serial_result==parallel_result |
|
99 | 97 | Out[67]: True |
|
100 | 98 | |
|
101 | 99 | |
|
102 | 100 | .. note:: |
|
103 | 101 | |
|
104 | 102 | The :class:`DirectView`'s version of :meth:`map` does |
|
105 | 103 | not do dynamic load balancing. For a load balanced version, use a |
|
106 |
:class:`LoadBalancedView` |
|
|
107 | `balanced=True`. | |
|
104 | :class:`LoadBalancedView`. | |
|
108 | 105 | |
|
109 | 106 | .. seealso:: |
|
110 | 107 | |
|
111 | 108 | :meth:`map` is implemented via :class:`ParallelFunction`. |
|
112 | 109 | |
|
113 | 110 | Remote function decorators |
|
114 | 111 | -------------------------- |
|
115 | 112 | |
|
116 | 113 | Remote functions are just like normal functions, but when they are called, |
|
117 | 114 | they execute on one or more engines, rather than locally. IPython provides |
|
118 | 115 | two decorators: |
|
119 | 116 | |
|
120 | 117 | .. sourcecode:: ipython |
|
121 | 118 | |
|
122 |
In [10]: @ |
|
|
119 | In [10]: @dview.remote(block=True) | |
|
123 | 120 | ...: def getpid(): |
|
124 | 121 | ...: import os |
|
125 | 122 | ...: return os.getpid() |
|
126 | 123 | ...: |
|
127 | 124 | |
|
128 | 125 | In [11]: getpid() |
|
129 | 126 | Out[11]: [12345, 12346, 12347, 12348] |
|
130 | 127 | |
|
131 |
|
|
|
128 | The ``@parallel`` decorator creates parallel functions, that break up an element-wise | |
|
132 | 129 | operations and distribute them, reconstructing the result. |
|
133 | 130 | |
|
134 | 131 | .. sourcecode:: ipython |
|
135 | 132 | |
|
136 | 133 | In [12]: import numpy as np |
|
137 | 134 | |
|
138 | 135 | In [13]: A = np.random.random((64,48)) |
|
139 | 136 | |
|
140 |
In [14]: @ |
|
|
137 | In [14]: @dview.parallel(block=True) | |
|
141 | 138 | ...: def pmul(A,B): |
|
142 | 139 | ...: return A*B |
|
143 | 140 | |
|
144 | 141 | In [15]: C_local = A*A |
|
145 | 142 | |
|
146 |
In [16]: C_remote |
|
|
143 | In [16]: C_remote = pmul(A,A) | |
|
147 | 144 | |
|
148 | 145 | In [17]: (C_local == C_remote).all() |
|
149 | 146 | Out[17]: True |
|
150 | 147 | |
|
151 | 148 | .. seealso:: |
|
152 | 149 | |
|
153 | 150 | See the docstrings for the :func:`parallel` and :func:`remote` decorators for |
|
154 | 151 | options. |
|
155 | 152 | |
|
156 | 153 | Calling Python functions |
|
157 | 154 | ======================== |
|
158 | 155 | |
|
159 | 156 | The most basic type of operation that can be performed on the engines is to |
|
160 | 157 | execute Python code or call Python functions. Executing Python code can be |
|
161 | 158 | done in blocking or non-blocking mode (non-blocking is default) using the |
|
162 | :meth:`execute` method, and calling functions can be done via the | |
|
159 | :meth:`.View.execute` method, and calling functions can be done via the | |
|
163 | 160 | :meth:`.View.apply` method. |
|
164 | 161 | |
|
165 | 162 | apply |
|
166 | 163 | ----- |
|
167 | 164 | |
|
168 | 165 | The main method for doing remote execution (in fact, all methods that |
|
169 |
communicate with the engines are built on top of it), is :meth:` |
|
|
170 | Ideally, :meth:`apply` would have the signature ``apply(f,*args,**kwargs)``, | |
|
171 | which would call ``f(*args,**kwargs)`` remotely. However, since :class:`Clients` | |
|
172 | require some more options, they cannot easily provide this interface. | |
|
173 | Instead, they provide the signature: | |
|
174 | ||
|
175 | .. sourcecode:: python | |
|
166 | communicate with the engines are built on top of it), is :meth:`View.apply`. | |
|
176 | 167 | |
|
177 | c.apply(f, args=None, kwargs=None, bound=True, block=None, targets=None, | |
|
178 | after=None, follow=None, timeout=None) | |
|
168 | We strive to provide the cleanest interface we can, so `apply` has the following | |
|
169 | signature: | |
|
179 | 170 | |
|
180 | Where various behavior is controlled via keyword arguments. This means that in the client, | |
|
181 | you must pass `args` as a tuple, and `kwargs` as a dict. | |
|
171 | .. sourcecode:: python | |
|
182 | 172 | |
|
183 | In order to provide the nicer interface, we have :class:`View` classes, which wrap | |
|
184 | :meth:`Client.apply` by using attributes and extra :meth:`apply_x` methods to determine | |
|
185 | the extra keyword arguments. This means that the views can have the desired pattern: | |
|
173 | view.apply(f, *args, **kwargs) | |
|
186 | 174 | |
|
187 | .. sourcecode:: python | |
|
175 | There are various ways to call functions with IPython, and these flags are set as | |
|
176 | attributes of the View. The ``DirectView`` has just two of these flags: | |
|
188 | 177 | |
|
189 | v.apply(f, *args, **kwargs) | |
|
178 | dv.block : bool | |
|
179 | whether to wait for the result, or return an :class:`AsyncResult` object | |
|
180 | immediately | |
|
181 | dv.track : bool | |
|
182 | whether to instruct pyzmq to track when | |
|
183 | This is primarily useful for non-copying sends of numpy arrays that you plan to | |
|
184 | edit in-place. You need to know when it becomes safe to edit the buffer | |
|
185 | without corrupting the message. | |
|
190 | 186 | |
|
191 | 187 | |
|
192 | For instance, performing index-access on a client creates a | |
|
193 | :class:`.DirectView`. | |
|
188 | Creating a view is simple: index-access on a client creates a :class:`.DirectView`. | |
|
194 | 189 | |
|
195 | 190 | .. sourcecode:: ipython |
|
196 | 191 | |
|
197 | 192 | In [4]: view = rc[1:3] |
|
198 | 193 | Out[4]: <DirectView [1, 2]> |
|
199 | 194 | |
|
200 | 195 | In [5]: view.apply<tab> |
|
201 |
view.apply view.apply_async view.apply_ |
|
|
202 | ||
|
203 | A :class:`DirectView` always uses its `targets` attribute, and it will use its `bound` | |
|
204 | and `block` attributes in its :meth:`apply` method, but the suffixed :meth:`apply_x` | |
|
205 | methods allow specifying `bound` and `block` via the different methods. | |
|
196 | view.apply view.apply_async view.apply_sync view.apply_with_flags | |
|
206 | 197 | |
|
207 | ================== ========== ========== | |
|
208 | method block bound | |
|
209 | ================== ========== ========== | |
|
210 | apply self.block self.bound | |
|
211 | apply_sync True False | |
|
212 | apply_async False False | |
|
213 | apply_sync_bound True True | |
|
214 | apply_async_bound False True | |
|
215 | ================== ========== ========== | |
|
216 | ||
|
217 | For explanation of these values, read on. | |
|
198 | For convenience, you can set block temporarily for a single call with the extra sync/async methods. | |
|
218 | 199 | |
|
219 | 200 | Blocking execution |
|
220 | 201 | ------------------ |
|
221 | 202 | |
|
222 | 203 | In blocking mode, the :class:`.DirectView` object (called ``dview`` in |
|
223 | 204 | these examples) submits the command to the controller, which places the |
|
224 | 205 | command in the engines' queues for execution. The :meth:`apply` call then |
|
225 | 206 | blocks until the engines are done executing the command: |
|
226 | 207 | |
|
227 | 208 | .. sourcecode:: ipython |
|
228 | 209 | |
|
229 | 210 | In [2]: dview = rc[:] # A DirectView of all engines |
|
230 | 211 | In [3]: dview.block=True |
|
231 | 212 | In [4]: dview['a'] = 5 |
|
232 | 213 | |
|
233 | 214 | In [5]: dview['b'] = 10 |
|
234 | 215 | |
|
235 |
In [6]: dview.apply |
|
|
216 | In [6]: dview.apply(lambda x: a+b+x, 27) | |
|
236 | 217 | Out[6]: [42, 42, 42, 42] |
|
237 | 218 | |
|
238 | Python commands can be executed on specific engines by calling execute using the ``targets`` | |
|
239 | keyword argument in :meth:`client.execute`, or creating a :class:`DirectView` instance by | |
|
240 | index-access to the client: | |
|
241 | ||
|
242 | .. sourcecode:: ipython | |
|
243 | ||
|
244 | In [6]: rc.execute('c=a+b', targets=[0,2]) | |
|
245 | ||
|
246 | In [7]: rc[1::2].execute('c=a-b') # shorthand for rc.execute('c=a-b',targets=[1,3]) | |
|
247 | ||
|
248 | In [8]: rc[:]['c'] # shorthand for rc.pull('c',targets='all') | |
|
249 | Out[8]: [15, -5, 15, -5] | |
|
250 | ||
|
251 | .. note:: | |
|
219 | You can also select blocking execution on a call-by-call basis with the :meth:`apply_sync` | |
|
220 | method: | |
|
252 | 221 | |
|
253 | Note that every call to ``rc.<meth>(...,targets=x)`` can be made via | |
|
254 | ``rc[<x>].<meth>(...)``, which constructs a View object. The only place | |
|
255 | where this differs in in :meth:`apply`. The :class:`Client` takes many | |
|
256 | arguments to apply, so it requires `args` and `kwargs` to be passed as | |
|
257 | individual arguments. Extended options such as `bound`,`targets`, and | |
|
258 | `block` are controlled by the attributes of the :class:`View` objects, so | |
|
259 | they can provide the much more convenient | |
|
260 | :meth:`View.apply(f,*args,**kwargs)`, which simply calls | |
|
261 | ``f(*args,**kwargs)`` remotely. | |
|
222 | In [7]: dview.block=False | |
|
262 | 223 | |
|
263 | Bound and unbound execution | |
|
264 | --------------------------- | |
|
224 | In [8]: dview.apply_sync(lambda x: a+b+x, 27) | |
|
225 | Out[8]: [42, 42, 42, 42] | |
|
265 | 226 | |
|
266 | The previous example also shows one of the most important things about the IPython | |
|
267 | engines: they have a persistent user namespaces. The :meth:`apply` method can | |
|
268 | be run in either a bound or unbound manner. | |
|
269 | ||
|
270 | When applying a function in a `bound` manner, the first argument to that function | |
|
271 | will be the Engine's namespace, which is a :class:`Namespace` object, a dictionary | |
|
272 | also providing attribute-access to keys. | |
|
273 | ||
|
274 | In all (unbound and bound) execution | |
|
227 | Python commands can be executed as strings on specific engines by using a View's ``execute`` | |
|
228 | method: | |
|
275 | 229 | |
|
276 | 230 | .. sourcecode:: ipython |
|
277 | 231 | |
|
278 | In [9]: dview['b'] = 5 # assign b to 5 everywhere | |
|
279 | ||
|
280 | In [10]: v0 = rc[0] | |
|
232 | In [6]: rc[::2].execute('c=a+b') | |
|
281 | 233 | |
|
282 | # multiply b*2 inplace | |
|
283 | In [12]: v0.apply_sync_bound(lambda ns: ns.b*=2) | |
|
234 | In [7]: rc[1::2].execute('c=a-b') | |
|
284 | 235 | |
|
285 | # b is still available in globals during unbound execution | |
|
286 | In [13]: v0.apply_sync(lambda a: a*b, 3) | |
|
287 | Out[13]: 30 | |
|
236 | In [8]: rc[:]['c'] # shorthand for rc[:].pull('c', block=True) | |
|
237 | Out[8]: [15, -5, 15, -5] | |
|
288 | 238 | |
|
289 | `bound=True` specifies that the engine's namespace is to be passed as the first argument when | |
|
290 | the function is called, and the default `bound=False` specifies that the normal behavior, but | |
|
291 | the engine's namespace will be available as the globals() when the function is called. | |
|
292 | 239 | |
|
293 | 240 | Non-blocking execution |
|
294 | 241 | ---------------------- |
|
295 | 242 | |
|
296 | 243 | In non-blocking mode, :meth:`apply` submits the command to be executed and |
|
297 | 244 | then returns a :class:`AsyncResult` object immediately. The |
|
298 | 245 | :class:`AsyncResult` object gives you a way of getting a result at a later |
|
299 | 246 | time through its :meth:`get` method. |
|
300 | 247 | |
|
301 | 248 | .. Note:: |
|
302 | 249 | |
|
303 | 250 | The :class:`AsyncResult` object provides a superset of the interface in |
|
304 | 251 | :py:class:`multiprocessing.pool.AsyncResult`. See the |
|
305 | 252 | `official Python documentation <http://docs.python.org/library/multiprocessing#multiprocessing.pool.AsyncResult>`_ |
|
306 | 253 | for more. |
|
307 | 254 | |
|
308 | 255 | |
|
309 | 256 | This allows you to quickly submit long running commands without blocking your |
|
310 | 257 | local Python/IPython session: |
|
311 | 258 | |
|
312 | 259 | .. sourcecode:: ipython |
|
313 | 260 | |
|
314 | 261 | # define our function |
|
315 | 262 | In [6]: def wait(t): |
|
316 | 263 | ...: import time |
|
317 | 264 | ...: tic = time.time() |
|
318 | 265 | ...: time.sleep(t) |
|
319 | 266 | ...: return time.time()-tic |
|
320 | 267 | |
|
321 | 268 | # In non-blocking mode |
|
322 | 269 | In [7]: ar = dview.apply_async(wait, 2) |
|
323 | 270 | |
|
324 | 271 | # Now block for the result |
|
325 | 272 | In [8]: ar.get() |
|
326 | 273 | Out[8]: [2.0006198883056641, 1.9997570514678955, 1.9996809959411621, 2.0003249645233154] |
|
327 | 274 | |
|
328 | 275 | # Again in non-blocking mode |
|
329 | 276 | In [9]: ar = dview.apply_async(wait, 10) |
|
330 | 277 | |
|
331 | 278 | # Poll to see if the result is ready |
|
332 | 279 | In [10]: ar.ready() |
|
333 | 280 | Out[10]: False |
|
334 | 281 | |
|
335 | 282 | # ask for the result, but wait a maximum of 1 second: |
|
336 | 283 | In [45]: ar.get(1) |
|
337 | 284 | --------------------------------------------------------------------------- |
|
338 | 285 | TimeoutError Traceback (most recent call last) |
|
339 | 286 | /home/you/<ipython-input-45-7cd858bbb8e0> in <module>() |
|
340 | 287 | ----> 1 ar.get(1) |
|
341 | 288 | |
|
342 | 289 | /path/to/site-packages/IPython/zmq/parallel/asyncresult.pyc in get(self, timeout) |
|
343 | 290 | 62 raise self._exception |
|
344 | 291 | 63 else: |
|
345 | 292 | ---> 64 raise error.TimeoutError("Result not ready.") |
|
346 | 293 | 65 |
|
347 | 294 | 66 def ready(self): |
|
348 | 295 | |
|
349 | 296 | TimeoutError: Result not ready. |
|
350 | 297 | |
|
351 | 298 | .. Note:: |
|
352 | 299 | |
|
353 | 300 | Note the import inside the function. This is a common model, to ensure |
|
354 | that the appropriate modules are imported where the task is run. | |
|
301 | that the appropriate modules are imported where the task is run. You can | |
|
302 | also manually import modules into the engine(s) namespace(s) via | |
|
303 | :meth:`view.execute('import numpy')`. | |
|
355 | 304 | |
|
356 | 305 | Often, it is desirable to wait until a set of :class:`AsyncResult` objects |
|
357 |
are done. For this, there is a the method :meth:` |
|
|
306 | are done. For this, there is a the method :meth:`wait`. This method takes a | |
|
358 | 307 | tuple of :class:`AsyncResult` objects (or `msg_ids` or indices to the client's History), |
|
359 | 308 | and blocks until all of the associated results are ready: |
|
360 | 309 | |
|
361 | 310 | .. sourcecode:: ipython |
|
362 | 311 | |
|
363 |
In [72]: |
|
|
312 | In [72]: dview.block=False | |
|
364 | 313 | |
|
365 | 314 | # A trivial list of AsyncResults objects |
|
366 | 315 | In [73]: pr_list = [dview.apply_async(wait, 3) for i in range(10)] |
|
367 | 316 | |
|
368 | 317 | # Wait until all of them are done |
|
369 |
In [74]: |
|
|
318 | In [74]: dview.wait(pr_list) | |
|
370 | 319 | |
|
371 | 320 | # Then, their results are ready using get() or the `.r` attribute |
|
372 | 321 | In [75]: pr_list[0].get() |
|
373 | 322 | Out[75]: [2.9982571601867676, 2.9982588291168213, 2.9987530708312988, 2.9990990161895752] |
|
374 | 323 | |
|
375 | 324 | |
|
376 | 325 | |
|
377 |
The ``block`` |
|
|
378 |
----------------------- |
|
|
326 | The ``block`` attribute | |
|
327 | ----------------------- | |
|
379 | 328 | |
|
380 |
M |
|
|
329 | Many View methods(excluding :meth:`apply`) accept | |
|
381 | 330 | ``block`` as a keyword argument. As we have seen above, these |
|
382 |
keyword arguments control the blocking mode. The :class:` |
|
|
331 | keyword arguments control the blocking mode. The :class:`View` class also has | |
|
383 | 332 | a :attr:`block` attribute that controls the default behavior when the keyword |
|
384 | 333 | argument is not provided. Thus the following logic is used for :attr:`block`: |
|
385 | 334 | |
|
386 | 335 | * If no keyword argument is provided, the instance attributes are used. |
|
387 | 336 | * Keyword argument, if provided override the instance attributes for |
|
388 | 337 | the duration of a single call. |
|
389 | 338 | |
|
390 | DirectView objects also have a ``bound`` attribute, which is used in the same way. | |
|
391 | ||
|
392 | 339 | The following examples demonstrate how to use the instance attributes: |
|
393 | 340 | |
|
394 | 341 | .. sourcecode:: ipython |
|
395 | 342 | |
|
396 |
In [17]: |
|
|
343 | In [17]: dview.block = False | |
|
397 | 344 | |
|
398 |
In [18]: ar = |
|
|
345 | In [18]: ar = dview.apply(lambda : 10) | |
|
399 | 346 | |
|
400 | 347 | In [19]: ar.get() |
|
401 | Out[19]: [10,10] | |
|
348 | Out[19]: [10, 10, 10, 10] | |
|
402 | 349 | |
|
403 |
In [21]: |
|
|
350 | In [21]: dview.block = True | |
|
404 | 351 | |
|
405 | 352 | # Note targets='all' means all engines |
|
406 |
In [22]: |
|
|
353 | In [22]: dview.apply(lambda : 42) | |
|
407 | 354 | Out[22]: [42, 42, 42, 42] |
|
408 | 355 | |
|
409 |
The :attr:`block` |
|
|
356 | The :attr:`block` and :attr:`targets` instance attributes of the | |
|
410 | 357 | :class:`.DirectView` also determine the behavior of the parallel magic commands. |
|
411 | 358 | |
|
412 | ||
|
413 | 359 | Parallel magic commands |
|
414 | 360 | ----------------------- |
|
415 | 361 | |
|
416 | 362 | .. warning:: |
|
417 | 363 | |
|
418 |
The magics have not been changed to work with the zeromq system. |
|
|
419 | and ``%autopx`` do work, but ``%result`` does not. %px and %autopx *do | |
|
420 | not* print stdin/out. | |
|
364 | The magics have not been changed to work with the zeromq system. The | |
|
365 | magics do work, but *do not* print stdin/out like they used to in IPython.kernel. | |
|
421 | 366 | |
|
422 | 367 | We provide a few IPython magic commands (``%px``, ``%autopx`` and ``%result``) |
|
423 | 368 | that make it more pleasant to execute Python commands on the engines |
|
424 | 369 | interactively. These are simply shortcuts to :meth:`execute` and |
|
425 | 370 | :meth:`get_result` of the :class:`DirectView`. The ``%px`` magic executes a single |
|
426 | 371 | Python command on the engines specified by the :attr:`targets` attribute of the |
|
427 | 372 | :class:`DirectView` instance: |
|
428 | 373 | |
|
429 | 374 | .. sourcecode:: ipython |
|
430 | 375 | |
|
376 | # load the parallel magic extension: | |
|
377 | In [21]: %load_ext parallelmagic | |
|
378 | ||
|
431 | 379 | # Create a DirectView for all targets |
|
432 | 380 | In [22]: dv = rc[:] |
|
433 | 381 | |
|
434 | 382 | # Make this DirectView active for parallel magic commands |
|
435 | 383 | In [23]: dv.activate() |
|
436 | 384 | |
|
437 | 385 | In [24]: dv.block=True |
|
438 | 386 | |
|
439 | 387 | In [25]: import numpy |
|
440 | 388 | |
|
441 | 389 | In [26]: %px import numpy |
|
442 | 390 | Parallel execution on engines: [0, 1, 2, 3] |
|
443 | 391 | |
|
444 | 392 | In [27]: %px a = numpy.random.rand(2,2) |
|
445 | 393 | Parallel execution on engines: [0, 1, 2, 3] |
|
446 | 394 | |
|
447 | 395 | In [28]: %px ev = numpy.linalg.eigvals(a) |
|
448 | 396 | Parallel execution on engines: [0, 1, 2, 3] |
|
449 | 397 | |
|
450 | 398 | In [28]: dv['ev'] |
|
451 | 399 | Out[28]: [ array([ 1.09522024, -0.09645227]), |
|
452 | 400 | array([ 1.21435496, -0.35546712]), |
|
453 | 401 | array([ 0.72180653, 0.07133042]), |
|
454 | 402 | array([ 1.46384341e+00, 1.04353244e-04]) |
|
455 | 403 | ] |
|
456 | 404 | |
|
457 | 405 | The ``%result`` magic gets the most recent result, or takes an argument |
|
458 | 406 | specifying the index of the result to be requested. It is simply a shortcut to the |
|
459 | 407 | :meth:`get_result` method: |
|
460 | 408 | |
|
461 | 409 | .. sourcecode:: ipython |
|
462 | 410 | |
|
463 | 411 | In [29]: dv.apply_async(lambda : ev) |
|
464 | 412 | |
|
465 | 413 | In [30]: %result |
|
466 | 414 | Out[30]: [ [ 1.28167017 0.14197338], |
|
467 | 415 | [-0.14093616 1.27877273], |
|
468 | 416 | [-0.37023573 1.06779409], |
|
469 | 417 | [ 0.83664764 -0.25602658] ] |
|
470 | 418 | |
|
471 | 419 | The ``%autopx`` magic switches to a mode where everything you type is executed |
|
472 | 420 | on the engines given by the :attr:`targets` attribute: |
|
473 | 421 | |
|
474 | 422 | .. sourcecode:: ipython |
|
475 | 423 | |
|
476 | 424 | In [30]: dv.block=False |
|
477 | 425 | |
|
478 | 426 | In [31]: %autopx |
|
479 | 427 | Auto Parallel Enabled |
|
480 | 428 | Type %autopx to disable |
|
481 | 429 | |
|
482 | 430 | In [32]: max_evals = [] |
|
483 | 431 | <IPython.zmq.parallel.asyncresult.AsyncResult object at 0x17b8a70> |
|
484 | 432 | |
|
485 | 433 | In [33]: for i in range(100): |
|
486 | 434 | ....: a = numpy.random.rand(10,10) |
|
487 | 435 | ....: a = a+a.transpose() |
|
488 | 436 | ....: evals = numpy.linalg.eigvals(a) |
|
489 | 437 | ....: max_evals.append(evals[0].real) |
|
490 | 438 | ....: |
|
491 | 439 | ....: |
|
492 | 440 | <IPython.zmq.parallel.asyncresult.AsyncResult object at 0x17af8f0> |
|
493 | 441 | |
|
494 | 442 | In [34]: %autopx |
|
495 | 443 | Auto Parallel Disabled |
|
496 | 444 | |
|
497 | 445 | In [35]: dv.block=True |
|
498 | 446 | |
|
499 | 447 | In [36]: px ans= "Average max eigenvalue is: %f"%(sum(max_evals)/len(max_evals)) |
|
500 | 448 | Parallel execution on engines: [0, 1, 2, 3] |
|
501 | 449 | |
|
502 | 450 | In [37]: dv['ans'] |
|
503 | 451 | Out[37]: [ 'Average max eigenvalue is: 10.1387247332', |
|
504 | 452 | 'Average max eigenvalue is: 10.2076902286', |
|
505 | 453 | 'Average max eigenvalue is: 10.1891484655', |
|
506 | 454 | 'Average max eigenvalue is: 10.1158837784',] |
|
507 | 455 | |
|
508 | 456 | |
|
509 | 457 | Moving Python objects around |
|
510 | 458 | ============================ |
|
511 | 459 | |
|
512 | 460 | In addition to calling functions and executing code on engines, you can |
|
513 | 461 | transfer Python objects to and from your IPython session and the engines. In |
|
514 | 462 | IPython, these operations are called :meth:`push` (sending an object to the |
|
515 | 463 |
engines) and :meth:`pull` (getting an object from the engines). |
|
516 | 464 | |
|
517 | 465 | Basic push and pull |
|
518 | 466 | ------------------- |
|
519 | 467 | |
|
520 | 468 | Here are some examples of how you use :meth:`push` and :meth:`pull`: |
|
521 | 469 | |
|
522 | 470 | .. sourcecode:: ipython |
|
523 | 471 | |
|
524 |
In [38]: |
|
|
472 | In [38]: dview.push(dict(a=1.03234,b=3453)) | |
|
525 | 473 | Out[38]: [None,None,None,None] |
|
526 | 474 | |
|
527 |
In [39]: |
|
|
475 | In [39]: dview.pull('a') | |
|
528 | 476 | Out[39]: [ 1.03234, 1.03234, 1.03234, 1.03234] |
|
529 | 477 | |
|
530 |
In [40]: rc.pull('b' |
|
|
478 | In [40]: rc[0].pull('b') | |
|
531 | 479 | Out[40]: 3453 |
|
532 | 480 | |
|
533 |
In [41]: |
|
|
481 | In [41]: dview.pull(('a','b')) | |
|
534 | 482 | Out[41]: [ [1.03234, 3453], [1.03234, 3453], [1.03234, 3453], [1.03234, 3453] ] |
|
535 | 483 | |
|
536 | # zmq client does not have zip_pull | |
|
537 | In [42]: rc.zip_pull(('a','b')) | |
|
538 | Out[42]: [(1.03234, 1.03234, 1.03234, 1.03234), (3453, 3453, 3453, 3453)] | |
|
539 | ||
|
540 | In [43]: rc.push(dict(c='speed')) | |
|
484 | In [43]: dview.push(dict(c='speed')) | |
|
541 | 485 | Out[43]: [None,None,None,None] |
|
542 | 486 | |
|
543 | 487 | In non-blocking mode :meth:`push` and :meth:`pull` also return |
|
544 | 488 | :class:`AsyncResult` objects: |
|
545 | 489 | |
|
546 | 490 | .. sourcecode:: ipython |
|
547 | 491 | |
|
548 |
In [4 |
|
|
549 | ||
|
550 | In [48]: ar = rc.pull('a') | |
|
492 | In [48]: ar = dview.pull('a', block=False) | |
|
551 | 493 | |
|
552 | 494 | In [49]: ar.get() |
|
553 | 495 | Out[49]: [1.03234, 1.03234, 1.03234, 1.03234] |
|
554 | 496 | |
|
555 | 497 | |
|
556 | 498 | Dictionary interface |
|
557 | 499 | -------------------- |
|
558 | 500 | |
|
559 | 501 | Since a Python namespace is just a :class:`dict`, :class:`DirectView` objects provide |
|
560 | 502 | dictionary-style access by key and methods such as :meth:`get` and |
|
561 | 503 | :meth:`update` for convenience. This make the remote namespaces of the engines |
|
562 | 504 | appear as a local dictionary. Underneath, these methods call :meth:`apply`: |
|
563 | 505 | |
|
564 | 506 | .. sourcecode:: ipython |
|
565 | 507 | |
|
566 | In [50]: dview.block=True | |
|
567 | ||
|
568 | 508 | In [51]: dview['a']=['foo','bar'] |
|
569 | 509 | |
|
570 | 510 | In [52]: dview['a'] |
|
571 | 511 | Out[52]: [ ['foo', 'bar'], ['foo', 'bar'], ['foo', 'bar'], ['foo', 'bar'] ] |
|
572 | 512 | |
|
573 | 513 | Scatter and gather |
|
574 | 514 | ------------------ |
|
575 | 515 | |
|
576 | 516 | Sometimes it is useful to partition a sequence and push the partitions to |
|
577 | 517 | different engines. In MPI language, this is know as scatter/gather and we |
|
578 | 518 | follow that terminology. However, it is important to remember that in |
|
579 | 519 | IPython's :class:`Client` class, :meth:`scatter` is from the |
|
580 | 520 | interactive IPython session to the engines and :meth:`gather` is from the |
|
581 | 521 | engines back to the interactive IPython session. For scatter/gather operations |
|
582 | 522 | between engines, MPI should be used: |
|
583 | 523 | |
|
584 | 524 | .. sourcecode:: ipython |
|
585 | 525 | |
|
586 | 526 | In [58]: dview.scatter('a',range(16)) |
|
587 | 527 | Out[58]: [None,None,None,None] |
|
588 | 528 | |
|
589 | 529 | In [59]: dview['a'] |
|
590 | 530 | Out[59]: [ [0, 1, 2, 3], [4, 5, 6, 7], [8, 9, 10, 11], [12, 13, 14, 15] ] |
|
591 | 531 | |
|
592 | 532 | In [60]: dview.gather('a') |
|
593 | 533 | Out[60]: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15] |
|
594 | 534 | |
|
595 | 535 | Other things to look at |
|
596 | 536 | ======================= |
|
597 | 537 | |
|
598 | 538 | How to do parallel list comprehensions |
|
599 | 539 | -------------------------------------- |
|
600 | 540 | |
|
601 | 541 | In many cases list comprehensions are nicer than using the map function. While |
|
602 | 542 | we don't have fully parallel list comprehensions, it is simple to get the |
|
603 | 543 | basic effect using :meth:`scatter` and :meth:`gather`: |
|
604 | 544 | |
|
605 | 545 | .. sourcecode:: ipython |
|
606 | 546 | |
|
607 | 547 | In [66]: dview.scatter('x',range(64)) |
|
608 | 548 | |
|
609 | In [67]: px y = [i**10 for i in x] | |
|
549 | In [67]: %px y = [i**10 for i in x] | |
|
610 | 550 | Parallel execution on engines: [0, 1, 2, 3] |
|
611 | 551 | Out[67]: |
|
612 | 552 | |
|
613 | 553 | In [68]: y = dview.gather('y') |
|
614 | 554 | |
|
615 | 555 | In [69]: print y |
|
616 | 556 | [0, 1, 1024, 59049, 1048576, 9765625, 60466176, 282475249, 1073741824,...] |
|
617 | 557 | |
|
618 | 558 | Parallel exceptions |
|
619 | 559 | ------------------- |
|
620 | 560 | |
|
621 | 561 | In the multiengine interface, parallel commands can raise Python exceptions, |
|
622 | 562 | just like serial commands. But, it is a little subtle, because a single |
|
623 | 563 | parallel command can actually raise multiple exceptions (one for each engine |
|
624 | 564 | the command was run on). To express this idea, we have a |
|
625 | 565 | :exc:`CompositeError` exception class that will be raised in most cases. The |
|
626 | 566 | :exc:`CompositeError` class is a special type of exception that wraps one or |
|
627 | 567 | more other types of exceptions. Here is how it works: |
|
628 | 568 | |
|
629 | 569 | .. sourcecode:: ipython |
|
630 | 570 | |
|
631 | 571 | In [76]: dview.block=True |
|
632 | 572 | |
|
633 | 573 | In [77]: dview.execute('1/0') |
|
634 | 574 | --------------------------------------------------------------------------- |
|
635 | 575 | CompositeError Traceback (most recent call last) |
|
636 |
/ |
|
|
637 | ----> 1 dview.execute('1/0') | |
|
638 | ||
|
639 | ... | |
|
576 | /home/you/<ipython-input-10-15c2c22dec39> in <module>() | |
|
577 | ----> 1 dview.execute('1/0', block=True) | |
|
578 | ||
|
579 | /path/to/site-packages/IPython/zmq/parallel/view.py in execute(self, code, block) | |
|
580 | 460 default: self.block | |
|
581 | 461 """ | |
|
582 | --> 462 return self.apply_with_flags(util._execute, args=(code,), block=block) | |
|
583 | 463 | |
|
584 | 464 def run(self, filename, block=None): | |
|
585 | ||
|
586 | /home/you/<string> in apply_with_flags(self, f, args, kwargs, block, track) | |
|
587 | ||
|
588 | /path/to/site-packages/IPython/zmq/parallel/view.py in sync_results(f, self, *args, **kwargs) | |
|
589 | 46 def sync_results(f, self, *args, **kwargs): | |
|
590 | 47 """sync relevant results from self.client to our results attribute.""" | |
|
591 | ---> 48 ret = f(self, *args, **kwargs) | |
|
592 | 49 delta = self.outstanding.difference(self.client.outstanding) | |
|
593 | 50 completed = self.outstanding.intersection(delta) | |
|
594 | ||
|
595 | /home/you/<string> in apply_with_flags(self, f, args, kwargs, block, track) | |
|
596 | ||
|
597 | /path/to/site-packages/IPython/zmq/parallel/view.py in save_ids(f, self, *args, **kwargs) | |
|
598 | 35 n_previous = len(self.client.history) | |
|
599 | 36 try: | |
|
600 | ---> 37 ret = f(self, *args, **kwargs) | |
|
601 | 38 finally: | |
|
602 | 39 nmsgs = len(self.client.history) - n_previous | |
|
603 | ||
|
604 | /path/to/site-packages/IPython/zmq/parallel/view.py in apply_with_flags(self, f, args, kwargs, block, track) | |
|
605 | 398 if block: | |
|
606 | 399 try: | |
|
607 | --> 400 return ar.get() | |
|
608 | 401 except KeyboardInterrupt: | |
|
609 | 402 pass | |
|
640 | 610 | |
|
641 | /Users/minrk/dev/ip/mine/IPython/zmq/parallel/client.pyc in apply(self, f, args, kwargs, bound, block, targets, balanced, after, follow, timeout) | |
|
642 |
|
|
|
643 |
|
|
|
644 | -> 1014 return self._apply_direct(f, args, kwargs, **options) | |
|
645 | 1015 | |
|
646 | 1016 def _apply_balanced(self, f, args, kwargs, bound=None, block=None, targets=None, | |
|
647 | ||
|
648 | /Users/minrk/dev/ip/mine/IPython/zmq/parallel/client.pyc in _apply_direct(self, f, args, kwargs, bound, block, targets) | |
|
649 | 1100 if block: | |
|
650 | 1101 try: | |
|
651 | -> 1102 return ar.get() | |
|
652 | 1103 except KeyboardInterrupt: | |
|
653 | 1104 return ar | |
|
654 | ||
|
655 | /Users/minrk/dev/ip/mine/IPython/zmq/parallel/asyncresult.pyc in get(self, timeout) | |
|
656 | 78 return self._result | |
|
657 | 79 else: | |
|
658 | ---> 80 raise self._exception | |
|
659 | 81 else: | |
|
660 | 82 raise error.TimeoutError("Result not ready.") | |
|
611 | /path/to/site-packages/IPython/zmq/parallel/asyncresult.pyc in get(self, timeout) | |
|
612 | 87 return self._result | |
|
613 | 88 else: | |
|
614 | ---> 89 raise self._exception | |
|
615 | 90 else: | |
|
616 | 91 raise error.TimeoutError("Result not ready.") | |
|
661 | 617 | |
|
662 | 618 | CompositeError: one or more exceptions from call to method: _execute |
|
663 | 619 | [0:apply]: ZeroDivisionError: integer division or modulo by zero |
|
664 | 620 | [1:apply]: ZeroDivisionError: integer division or modulo by zero |
|
665 | 621 | [2:apply]: ZeroDivisionError: integer division or modulo by zero |
|
666 | 622 | [3:apply]: ZeroDivisionError: integer division or modulo by zero |
|
667 | 623 | |
|
624 | ||
|
668 | 625 | Notice how the error message printed when :exc:`CompositeError` is raised has |
|
669 | 626 | information about the individual exceptions that were raised on each engine. |
|
670 | 627 | If you want, you can even raise one of these original exceptions: |
|
671 | 628 | |
|
672 | 629 | .. sourcecode:: ipython |
|
673 | 630 | |
|
674 | 631 | In [80]: try: |
|
675 |
....: |
|
|
632 | ....: dview.execute('1/0') | |
|
676 | 633 | ....: except client.CompositeError, e: |
|
677 | 634 | ....: e.raise_exception() |
|
678 | 635 | ....: |
|
679 | 636 | ....: |
|
680 | 637 | --------------------------------------------------------------------------- |
|
681 | 638 | ZeroDivisionError Traceback (most recent call last) |
|
682 | 639 | |
|
683 | 640 | /ipython1-client-r3021/docs/examples/<ipython console> in <module>() |
|
684 | 641 | |
|
685 | 642 | /ipython1-client-r3021/ipython1/kernel/error.pyc in raise_exception(self, excid) |
|
686 | 643 | 156 raise IndexError("an exception with index %i does not exist"%excid) |
|
687 | 644 | 157 else: |
|
688 | 645 | --> 158 raise et, ev, etb |
|
689 | 646 | 159 |
|
690 | 647 | 160 def collect_exceptions(rlist, method): |
|
691 | 648 | |
|
692 | 649 | ZeroDivisionError: integer division or modulo by zero |
|
693 | 650 | |
|
694 | 651 | If you are working in IPython, you can simple type ``%debug`` after one of |
|
695 | 652 | these :exc:`CompositeError` exceptions is raised, and inspect the exception |
|
696 | 653 | instance: |
|
697 | 654 | |
|
698 | 655 | .. sourcecode:: ipython |
|
699 | 656 | |
|
700 |
In [81]: |
|
|
657 | In [81]: dview.execute('1/0') | |
|
701 | 658 | --------------------------------------------------------------------------- |
|
702 | 659 | CompositeError Traceback (most recent call last) |
|
703 |
/ |
|
|
704 |
----> 1 |
|
|
705 | ||
|
706 | /Users/minrk/<string> in execute(self, code, targets, block) | |
|
707 | ||
|
708 | /Users/minrk/dev/ip/mine/IPython/zmq/parallel/client.pyc in defaultblock(f, self, *args, **kwargs) | |
|
709 | 88 self.block = block | |
|
710 | 89 try: | |
|
711 | ---> 90 ret = f(self, *args, **kwargs) | |
|
712 | 91 finally: | |
|
713 | 92 self.block = saveblock | |
|
714 | ||
|
715 | /Users/minrk/dev/ip/mine/IPython/zmq/parallel/client.pyc in execute(self, code, targets, block) | |
|
716 | 855 default: self.block | |
|
717 | 856 """ | |
|
718 | --> 857 result = self.apply(_execute, (code,), targets=targets, block=block, bound=True, balanced=False) | |
|
719 | 858 if not block: | |
|
720 | 859 return result | |
|
721 | ||
|
722 |
/ |
|
|
723 | ||
|
724 |
/ |
|
|
725 | 88 self.block = block | |
|
726 |
|
|
|
727 |
---> |
|
|
728 |
|
|
|
729 | 92 self.block = saveblock | |
|
730 | ||
|
731 |
/ |
|
|
732 | 1012 raise ValueError(msg) | |
|
733 |
|
|
|
734 | -> 1014 return self._apply_direct(f, args, kwargs, **options) | |
|
735 | 1015 | |
|
736 | 1016 def _apply_balanced(self, f, args, kwargs, bound=None, block=None, targets=None, | |
|
737 | ||
|
738 | /Users/minrk/dev/ip/mine/IPython/zmq/parallel/client.pyc in _apply_direct(self, f, args, kwargs, bound, block, targets) | |
|
739 | 1100 if block: | |
|
740 | 1101 try: | |
|
741 | -> 1102 return ar.get() | |
|
742 | 1103 except KeyboardInterrupt: | |
|
743 | 1104 return ar | |
|
660 | /home/you/<ipython-input-10-15c2c22dec39> in <module>() | |
|
661 | ----> 1 dview.execute('1/0', block=True) | |
|
662 | ||
|
663 | /path/to/site-packages/IPython/zmq/parallel/view.py in execute(self, code, block) | |
|
664 | 460 default: self.block | |
|
665 | 461 """ | |
|
666 | --> 462 return self.apply_with_flags(util._execute, args=(code,), block=block) | |
|
667 | 463 | |
|
668 | 464 def run(self, filename, block=None): | |
|
669 | ||
|
670 | /home/you/<string> in apply_with_flags(self, f, args, kwargs, block, track) | |
|
671 | ||
|
672 | /path/to/site-packages/IPython/zmq/parallel/view.py in sync_results(f, self, *args, **kwargs) | |
|
673 | 46 def sync_results(f, self, *args, **kwargs): | |
|
674 | 47 """sync relevant results from self.client to our results attribute.""" | |
|
675 | ---> 48 ret = f(self, *args, **kwargs) | |
|
676 | 49 delta = self.outstanding.difference(self.client.outstanding) | |
|
677 | 50 completed = self.outstanding.intersection(delta) | |
|
678 | ||
|
679 | /home/you/<string> in apply_with_flags(self, f, args, kwargs, block, track) | |
|
680 | ||
|
681 | /path/to/site-packages/IPython/zmq/parallel/view.py in save_ids(f, self, *args, **kwargs) | |
|
682 | 35 n_previous = len(self.client.history) | |
|
683 | 36 try: | |
|
684 | ---> 37 ret = f(self, *args, **kwargs) | |
|
685 | 38 finally: | |
|
686 | 39 nmsgs = len(self.client.history) - n_previous | |
|
687 | ||
|
688 | /path/to/site-packages/IPython/zmq/parallel/view.py in apply_with_flags(self, f, args, kwargs, block, track) | |
|
689 | 398 if block: | |
|
690 | 399 try: | |
|
691 | --> 400 return ar.get() | |
|
692 | 401 except KeyboardInterrupt: | |
|
693 | 402 pass | |
|
744 | 694 | |
|
745 |
/ |
|
|
746 |
|
|
|
747 |
|
|
|
748 |
---> 8 |
|
|
749 |
|
|
|
750 |
|
|
|
695 | /path/to/site-packages/IPython/zmq/parallel/asyncresult.pyc in get(self, timeout) | |
|
696 | 87 return self._result | |
|
697 | 88 else: | |
|
698 | ---> 89 raise self._exception | |
|
699 | 90 else: | |
|
700 | 91 raise error.TimeoutError("Result not ready.") | |
|
751 | 701 | |
|
752 | 702 | CompositeError: one or more exceptions from call to method: _execute |
|
753 | 703 | [0:apply]: ZeroDivisionError: integer division or modulo by zero |
|
754 | 704 | [1:apply]: ZeroDivisionError: integer division or modulo by zero |
|
755 | 705 | [2:apply]: ZeroDivisionError: integer division or modulo by zero |
|
756 | 706 | [3:apply]: ZeroDivisionError: integer division or modulo by zero |
|
757 | 707 | |
|
758 | 708 | In [82]: %debug |
|
759 | 709 | > /Users/minrk/dev/ip/mine/IPython/zmq/parallel/asyncresult.py(80)get() |
|
760 | 710 | 79 else: |
|
761 | 711 | ---> 80 raise self._exception |
|
762 | 712 | 81 else: |
|
763 | 713 | |
|
764 | 714 | |
|
765 | 715 | # With the debugger running, e is the exceptions instance. We can tab complete |
|
766 | 716 | # on it and see the extra methods that are available. |
|
767 | 717 | ipdb> e. |
|
768 | 718 | e.__class__ e.__getitem__ e.__new__ e.__setstate__ e.args |
|
769 | 719 | e.__delattr__ e.__getslice__ e.__reduce__ e.__str__ e.elist |
|
770 | 720 | e.__dict__ e.__hash__ e.__reduce_ex__ e.__weakref__ e.message |
|
771 | 721 | e.__doc__ e.__init__ e.__repr__ e._get_engine_str e.print_tracebacks |
|
772 | 722 | e.__getattribute__ e.__module__ e.__setattr__ e._get_traceback e.raise_exception |
|
773 | 723 | ipdb> e.print_tracebacks() |
|
774 | 724 | [0:apply]: |
|
775 | 725 | Traceback (most recent call last): |
|
776 | 726 | File "/Users/minrk/dev/ip/mine/IPython/zmq/parallel/streamkernel.py", line 332, in apply_request |
|
777 | 727 | exec code in working, working |
|
778 | 728 | File "<string>", line 1, in <module> |
|
779 | 729 | File "/Users/minrk/dev/ip/mine/IPython/zmq/parallel/client.py", line 69, in _execute |
|
780 | 730 | exec code in globals() |
|
781 | 731 | File "<string>", line 1, in <module> |
|
782 | 732 | ZeroDivisionError: integer division or modulo by zero |
|
783 | 733 | |
|
784 | 734 | |
|
785 | 735 | [1:apply]: |
|
786 | 736 | Traceback (most recent call last): |
|
787 | 737 | File "/Users/minrk/dev/ip/mine/IPython/zmq/parallel/streamkernel.py", line 332, in apply_request |
|
788 | 738 | exec code in working, working |
|
789 | 739 | File "<string>", line 1, in <module> |
|
790 | 740 | File "/Users/minrk/dev/ip/mine/IPython/zmq/parallel/client.py", line 69, in _execute |
|
791 | 741 | exec code in globals() |
|
792 | 742 | File "<string>", line 1, in <module> |
|
793 | 743 | ZeroDivisionError: integer division or modulo by zero |
|
794 | 744 | |
|
795 | 745 | |
|
796 | 746 | [2:apply]: |
|
797 | 747 | Traceback (most recent call last): |
|
798 | 748 | File "/Users/minrk/dev/ip/mine/IPython/zmq/parallel/streamkernel.py", line 332, in apply_request |
|
799 | 749 | exec code in working, working |
|
800 | 750 | File "<string>", line 1, in <module> |
|
801 | 751 | File "/Users/minrk/dev/ip/mine/IPython/zmq/parallel/client.py", line 69, in _execute |
|
802 | 752 | exec code in globals() |
|
803 | 753 | File "<string>", line 1, in <module> |
|
804 | 754 | ZeroDivisionError: integer division or modulo by zero |
|
805 | 755 | |
|
806 | 756 | |
|
807 | 757 | [3:apply]: |
|
808 | 758 | Traceback (most recent call last): |
|
809 | 759 | File "/Users/minrk/dev/ip/mine/IPython/zmq/parallel/streamkernel.py", line 332, in apply_request |
|
810 | 760 | exec code in working, working |
|
811 | 761 | File "<string>", line 1, in <module> |
|
812 | 762 | File "/Users/minrk/dev/ip/mine/IPython/zmq/parallel/client.py", line 69, in _execute |
|
813 | 763 | exec code in globals() |
|
814 | 764 | File "<string>", line 1, in <module> |
|
815 | 765 | ZeroDivisionError: integer division or modulo by zero |
|
816 | 766 | |
|
817 | 767 | |
|
768 | .. note:: | |
|
769 | ||
|
770 | TODO: The above tracebacks are not up to date | |
|
771 | ||
|
818 | 772 | |
|
819 | 773 | All of this same error handling magic even works in non-blocking mode: |
|
820 | 774 | |
|
821 | 775 | .. sourcecode:: ipython |
|
822 | 776 | |
|
823 |
In [83]: |
|
|
777 | In [83]: dview.block=False | |
|
824 | 778 | |
|
825 |
In [84]: ar = |
|
|
779 | In [84]: ar = dview.execute('1/0') | |
|
826 | 780 | |
|
827 | 781 | In [85]: ar.get() |
|
828 | 782 | --------------------------------------------------------------------------- |
|
829 | 783 | CompositeError Traceback (most recent call last) |
|
830 | 784 | /Users/minrk/<ipython-input-3-8531eb3d26fb> in <module>() |
|
831 | 785 | ----> 1 ar.get() |
|
832 | 786 | |
|
833 | 787 | /Users/minrk/dev/ip/mine/IPython/zmq/parallel/asyncresult.pyc in get(self, timeout) |
|
834 | 788 | 78 return self._result |
|
835 | 789 | 79 else: |
|
836 | 790 | ---> 80 raise self._exception |
|
837 | 791 | 81 else: |
|
838 | 792 | 82 raise error.TimeoutError("Result not ready.") |
|
839 | 793 | |
|
840 | 794 | CompositeError: one or more exceptions from call to method: _execute |
|
841 | 795 | [0:apply]: ZeroDivisionError: integer division or modulo by zero |
|
842 | 796 | [1:apply]: ZeroDivisionError: integer division or modulo by zero |
|
843 | 797 | [2:apply]: ZeroDivisionError: integer division or modulo by zero |
|
844 | 798 | [3:apply]: ZeroDivisionError: integer division or modulo by zero |
|
845 | 799 |
@@ -1,412 +1,419 b'' | |||
|
1 | .. _paralleltask: | |
|
1 | .. _parallel_task: | |
|
2 | 2 | |
|
3 | 3 | ========================== |
|
4 | 4 | The IPython task interface |
|
5 | 5 | ========================== |
|
6 | 6 | |
|
7 | 7 | The task interface to the cluster presents the engines as a fault tolerant, |
|
8 | 8 | dynamic load-balanced system of workers. Unlike the multiengine interface, in |
|
9 | 9 | the task interface the user have no direct access to individual engines. By |
|
10 | 10 | allowing the IPython scheduler to assign work, this interface is simultaneously |
|
11 | 11 | simpler and more powerful. |
|
12 | 12 | |
|
13 | 13 | Best of all, the user can use both of these interfaces running at the same time |
|
14 | 14 | to take advantage of their respective strengths. When the user can break up |
|
15 | 15 | the user's work into segments that do not depend on previous execution, the |
|
16 | 16 | task interface is ideal. But it also has more power and flexibility, allowing |
|
17 | 17 | the user to guide the distribution of jobs, without having to assign tasks to |
|
18 | 18 | engines explicitly. |
|
19 | 19 | |
|
20 | 20 | Starting the IPython controller and engines |
|
21 | 21 | =========================================== |
|
22 | 22 | |
|
23 | 23 | To follow along with this tutorial, you will need to start the IPython |
|
24 | 24 | controller and four IPython engines. The simplest way of doing this is to use |
|
25 | 25 | the :command:`ipclusterz` command:: |
|
26 | 26 | |
|
27 | 27 | $ ipclusterz start -n 4 |
|
28 | 28 | |
|
29 | 29 | For more detailed information about starting the controller and engines, see |
|
30 | 30 | our :ref:`introduction <ip1par>` to using IPython for parallel computing. |
|
31 | 31 | |
|
32 | 32 | Creating a ``Client`` instance |
|
33 | 33 | ============================== |
|
34 | 34 | |
|
35 | 35 | The first step is to import the IPython :mod:`IPython.zmq.parallel.client` |
|
36 | 36 | module and then create a :class:`.Client` instance, and we will also be using |
|
37 | 37 | a :class:`LoadBalancedView`, here called `lview`: |
|
38 | 38 | |
|
39 | 39 | .. sourcecode:: ipython |
|
40 | 40 | |
|
41 | 41 | In [1]: from IPython.zmq.parallel import client |
|
42 | 42 | |
|
43 | 43 | In [2]: rc = client.Client() |
|
44 | 44 | |
|
45 | 45 | |
|
46 | 46 | This form assumes that the controller was started on localhost with default |
|
47 | 47 | configuration. If not, the location of the controller must be given as an |
|
48 | 48 | argument to the constructor: |
|
49 | 49 | |
|
50 | 50 | .. sourcecode:: ipython |
|
51 | 51 | |
|
52 | 52 | # for a visible LAN controller listening on an external port: |
|
53 | 53 | In [2]: rc = client.Client('tcp://192.168.1.16:10101') |
|
54 | 54 | # or to connect with a specific profile you have set up: |
|
55 | 55 | In [3]: rc = client.Client(profile='mpi') |
|
56 | 56 | |
|
57 |
For load-balanced execution, we will make use of a :class:`LoadBalancedView` object, which can |
|
|
57 | For load-balanced execution, we will make use of a :class:`LoadBalancedView` object, which can | |
|
58 | be constructed via the client's :meth:`load_balanced_view` method: | |
|
58 | 59 | |
|
59 | 60 | .. sourcecode:: ipython |
|
60 | 61 | |
|
61 | In [4]: lview = rc.view() # default load-balanced view | |
|
62 | In [4]: lview = rc.load_balanced_view() # default load-balanced view | |
|
62 | 63 | |
|
63 | 64 | .. seealso:: |
|
64 | 65 | |
|
65 | 66 | For more information, see the in-depth explanation of :ref:`Views <parallel_details>`. |
|
66 | 67 | |
|
67 | 68 | |
|
68 | 69 | Quick and easy parallelism |
|
69 | 70 | ========================== |
|
70 | 71 | |
|
71 | 72 | In many cases, you simply want to apply a Python function to a sequence of |
|
72 | 73 | objects, but *in parallel*. Like the multiengine interface, these can be |
|
73 | 74 | implemented via the task interface. The exact same tools can perform these |
|
74 | 75 | actions in load-balanced ways as well as multiplexed ways: a parallel version |
|
75 | 76 | of :func:`map` and :func:`@parallel` function decorator. If one specifies the |
|
76 | 77 | argument `balanced=True`, then they are dynamically load balanced. Thus, if the |
|
77 | 78 | execution time per item varies significantly, you should use the versions in |
|
78 | 79 | the task interface. |
|
79 | 80 | |
|
80 | 81 | Parallel map |
|
81 | 82 | ------------ |
|
82 | 83 | |
|
83 | 84 | To load-balance :meth:`map`,simply use a LoadBalancedView: |
|
84 | 85 | |
|
85 | 86 | .. sourcecode:: ipython |
|
86 | 87 | |
|
87 | 88 | In [62]: lview.block = True |
|
88 | 89 | |
|
89 | 90 | In [63]: serial_result = map(lambda x:x**10, range(32)) |
|
90 | 91 | |
|
91 | 92 | In [64]: parallel_result = lview.map(lambda x:x**10, range(32)) |
|
92 | 93 | |
|
93 | 94 | In [65]: serial_result==parallel_result |
|
94 | 95 | Out[65]: True |
|
95 | 96 | |
|
96 | 97 | Parallel function decorator |
|
97 | 98 | --------------------------- |
|
98 | 99 | |
|
99 | 100 | Parallel functions are just like normal function, but they can be called on |
|
100 | 101 | sequences and *in parallel*. The multiengine interface provides a decorator |
|
101 | 102 | that turns any Python function into a parallel function: |
|
102 | 103 | |
|
103 | 104 | .. sourcecode:: ipython |
|
104 | 105 | |
|
105 | 106 | In [10]: @lview.parallel() |
|
106 | 107 | ....: def f(x): |
|
107 | 108 | ....: return 10.0*x**4 |
|
108 | 109 | ....: |
|
109 | 110 | |
|
110 | 111 | In [11]: f.map(range(32)) # this is done in parallel |
|
111 | 112 | Out[11]: [0.0,10.0,160.0,...] |
|
112 | 113 | |
|
114 | .. _parallel_dependencies: | |
|
115 | ||
|
113 | 116 | Dependencies |
|
114 | 117 | ============ |
|
115 | 118 | |
|
116 | 119 | Often, pure atomic load-balancing is too primitive for your work. In these cases, you |
|
117 | 120 | may want to associate some kind of `Dependency` that describes when, where, or whether |
|
118 | 121 | a task can be run. In IPython, we provide two types of dependencies: |
|
119 | 122 | `Functional Dependencies`_ and `Graph Dependencies`_ |
|
120 | 123 | |
|
121 | 124 | .. note:: |
|
122 | 125 | |
|
123 | 126 | It is important to note that the pure ZeroMQ scheduler does not support dependencies, |
|
124 | 127 | and you will see errors or warnings if you try to use dependencies with the pure |
|
125 | 128 | scheduler. |
|
126 | 129 | |
|
127 | 130 | Functional Dependencies |
|
128 | 131 | ----------------------- |
|
129 | 132 | |
|
130 | 133 | Functional dependencies are used to determine whether a given engine is capable of running |
|
131 | 134 | a particular task. This is implemented via a special :class:`Exception` class, |
|
132 | 135 | :class:`UnmetDependency`, found in `IPython.zmq.parallel.error`. Its use is very simple: |
|
133 | 136 | if a task fails with an UnmetDependency exception, then the scheduler, instead of relaying |
|
134 | 137 | the error up to the client like any other error, catches the error, and submits the task |
|
135 | 138 | to a different engine. This will repeat indefinitely, and a task will never be submitted |
|
136 | 139 | to a given engine a second time. |
|
137 | 140 | |
|
138 | 141 | You can manually raise the :class:`UnmetDependency` yourself, but IPython has provided |
|
139 | 142 | some decorators for facilitating this behavior. |
|
140 | 143 | |
|
141 | 144 | There are two decorators and a class used for functional dependencies: |
|
142 | 145 | |
|
143 | 146 | .. sourcecode:: ipython |
|
144 | 147 | |
|
145 | 148 | In [9]: from IPython.zmq.parallel.dependency import depend, require, dependent |
|
146 | 149 | |
|
147 | 150 | @require |
|
148 | 151 | ******** |
|
149 | 152 | |
|
150 | 153 | The simplest sort of dependency is requiring that a Python module is available. The |
|
151 | 154 | ``@require`` decorator lets you define a function that will only run on engines where names |
|
152 | 155 | you specify are importable: |
|
153 | 156 | |
|
154 | 157 | .. sourcecode:: ipython |
|
155 | 158 | |
|
156 | 159 | In [10]: @require('numpy', 'zmq') |
|
157 | 160 | ...: def myfunc(): |
|
158 | 161 | ...: import numpy,zmq |
|
159 | 162 | ...: return dostuff() |
|
160 | 163 | |
|
161 | 164 | Now, any time you apply :func:`myfunc`, the task will only run on a machine that has |
|
162 | 165 | numpy and pyzmq available. |
|
163 | 166 | |
|
164 | 167 | @depend |
|
165 | 168 | ******* |
|
166 | 169 | |
|
167 | 170 | The ``@depend`` decorator lets you decorate any function with any *other* function to |
|
168 | 171 | evaluate the dependency. The dependency function will be called at the start of the task, |
|
169 | 172 | and if it returns ``False``, then the dependency will be considered unmet, and the task |
|
170 | 173 | will be assigned to another engine. If the dependency returns *anything other than |
|
171 | 174 | ``False``*, the rest of the task will continue. |
|
172 | 175 | |
|
173 | 176 | .. sourcecode:: ipython |
|
174 | 177 | |
|
175 | 178 | In [10]: def platform_specific(plat): |
|
176 | 179 | ...: import sys |
|
177 | 180 | ...: return sys.platform == plat |
|
178 | 181 | |
|
179 | 182 | In [11]: @depend(platform_specific, 'darwin') |
|
180 | 183 | ...: def mactask(): |
|
181 | 184 | ...: do_mac_stuff() |
|
182 | 185 | |
|
183 | 186 | In [12]: @depend(platform_specific, 'nt') |
|
184 | 187 | ...: def wintask(): |
|
185 | 188 | ...: do_windows_stuff() |
|
186 | 189 | |
|
187 | 190 | In this case, any time you apply ``mytask``, it will only run on an OSX machine. |
|
188 | 191 | ``@depend`` is just like ``apply``, in that it has a ``@depend(f,*args,**kwargs)`` |
|
189 | 192 | signature. |
|
190 | 193 | |
|
191 | 194 | dependents |
|
192 | 195 | ********** |
|
193 | 196 | |
|
194 | 197 | You don't have to use the decorators on your tasks, if for instance you may want |
|
195 | 198 | to run tasks with a single function but varying dependencies, you can directly construct |
|
196 | 199 | the :class:`dependent` object that the decorators use: |
|
197 | 200 | |
|
198 | 201 | .. sourcecode::ipython |
|
199 | 202 | |
|
200 | 203 | In [13]: def mytask(*args): |
|
201 | 204 | ...: dostuff() |
|
202 | 205 | |
|
203 | 206 | In [14]: mactask = dependent(mytask, platform_specific, 'darwin') |
|
204 | 207 | # this is the same as decorating the declaration of mytask with @depend |
|
205 | 208 | # but you can do it again: |
|
206 | 209 | |
|
207 | 210 | In [15]: wintask = dependent(mytask, platform_specific, 'nt') |
|
208 | 211 | |
|
209 | 212 | # in general: |
|
210 | 213 | In [16]: t = dependent(f, g, *dargs, **dkwargs) |
|
211 | 214 | |
|
212 | 215 | # is equivalent to: |
|
213 | 216 | In [17]: @depend(g, *dargs, **dkwargs) |
|
214 | 217 | ...: def t(a,b,c): |
|
215 | 218 | ...: # contents of f |
|
216 | 219 | |
|
217 | 220 | Graph Dependencies |
|
218 | 221 | ------------------ |
|
219 | 222 | |
|
220 | 223 | Sometimes you want to restrict the time and/or location to run a given task as a function |
|
221 | 224 | of the time and/or location of other tasks. This is implemented via a subclass of |
|
222 | 225 | :class:`set`, called a :class:`Dependency`. A Dependency is just a set of `msg_ids` |
|
223 | 226 | corresponding to tasks, and a few attributes to guide how to decide when the Dependency |
|
224 | 227 | has been met. |
|
225 | 228 | |
|
226 | 229 | The switches we provide for interpreting whether a given dependency set has been met: |
|
227 | 230 | |
|
228 | 231 | any|all |
|
229 | 232 | Whether the dependency is considered met if *any* of the dependencies are done, or |
|
230 | 233 | only after *all* of them have finished. This is set by a Dependency's :attr:`all` |
|
231 | 234 | boolean attribute, which defaults to ``True``. |
|
232 | 235 | |
|
233 | success_only | |
|
234 |
Whether to consider |
|
|
235 | Sometimes you want to run a task after another, but only if that task succeeded. In | |
|
236 | this case, ``success_only`` should be ``True``. However sometimes you may not care | |
|
237 | whether the task succeeds, and always want the second task to run, in which case | |
|
238 | you should use `success_only=False`. The default behavior is to only use successes. | |
|
236 | success [default: True] | |
|
237 | Whether to consider tasks that succeeded as fulfilling dependencies. | |
|
238 | ||
|
239 | failure [default : False] | |
|
240 | Whether to consider tasks that failed as fulfilling dependencies. | |
|
241 | using `failure=True,success=False` is useful for setting up cleanup tasks, to be run | |
|
242 | only when tasks have failed. | |
|
243 | ||
|
244 | Sometimes you want to run a task after another, but only if that task succeeded. In this case, | |
|
245 | ``success`` should be ``True`` and ``failure`` should be ``False``. However sometimes you may | |
|
246 | not care whether the task succeeds, and always want the second task to run, in which case you | |
|
247 | should use `success=failure=True`. The default behavior is to only use successes. | |
|
239 | 248 | |
|
240 | 249 | There are other switches for interpretation that are made at the *task* level. These are |
|
241 | 250 | specified via keyword arguments to the client's :meth:`apply` method. |
|
242 | 251 | |
|
243 | 252 | after,follow |
|
244 | 253 | You may want to run a task *after* a given set of dependencies have been run and/or |
|
245 | 254 | run it *where* another set of dependencies are met. To support this, every task has an |
|
246 | 255 | `after` dependency to restrict time, and a `follow` dependency to restrict |
|
247 | 256 | destination. |
|
248 | 257 | |
|
249 | 258 | timeout |
|
250 | 259 | You may also want to set a time-limit for how long the scheduler should wait before a |
|
251 | 260 | task's dependencies are met. This is done via a `timeout`, which defaults to 0, which |
|
252 | 261 | indicates that the task should never timeout. If the timeout is reached, and the |
|
253 | 262 | scheduler still hasn't been able to assign the task to an engine, the task will fail |
|
254 | 263 | with a :class:`DependencyTimeout`. |
|
255 | 264 | |
|
256 | 265 | .. note:: |
|
257 | 266 | |
|
258 | 267 | Dependencies only work within the task scheduler. You cannot instruct a load-balanced |
|
259 | 268 | task to run after a job submitted via the MUX interface. |
|
260 | 269 | |
|
261 |
The simplest form of Dependencies is with `all=True,success |
|
|
270 | The simplest form of Dependencies is with `all=True,success=True,failure=False`. In these cases, | |
|
262 | 271 | you can skip using Dependency objects, and just pass msg_ids or AsyncResult objects as the |
|
263 | 272 | `follow` and `after` keywords to :meth:`client.apply`: |
|
264 | 273 | |
|
265 | 274 | .. sourcecode:: ipython |
|
266 | 275 | |
|
267 | 276 | In [14]: client.block=False |
|
268 | 277 | |
|
269 |
In [15]: ar = |
|
|
278 | In [15]: ar = lview.apply(f, args, kwargs) | |
|
270 | 279 | |
|
271 |
In [16]: ar2 = |
|
|
280 | In [16]: ar2 = lview.apply(f2) | |
|
272 | 281 | |
|
273 |
In [17]: ar3 = |
|
|
282 | In [17]: ar3 = lview.apply_with_flags(f3, after=[ar,ar2]) | |
|
274 | 283 | |
|
275 |
In [17]: ar4 = |
|
|
284 | In [17]: ar4 = lview.apply_with_flags(f3, follow=[ar], timeout=2.5) | |
|
276 | 285 | |
|
277 | 286 | |
|
278 | 287 | .. seealso:: |
|
279 | 288 | |
|
280 | 289 | Some parallel workloads can be described as a `Directed Acyclic Graph |
|
281 | 290 | <http://en.wikipedia.org/wiki/Directed_acyclic_graph>`_, or DAG. See :ref:`DAG |
|
282 | 291 | Dependencies <dag_dependencies>` for an example demonstrating how to use map a NetworkX DAG |
|
283 | 292 | onto task dependencies. |
|
284 | 293 | |
|
285 | 294 | |
|
286 | 295 | |
|
287 | 296 | Impossible Dependencies |
|
288 | 297 | *********************** |
|
289 | 298 | |
|
290 | 299 | The schedulers do perform some analysis on graph dependencies to determine whether they |
|
291 | 300 | are not possible to be met. If the scheduler does discover that a dependency cannot be |
|
292 | 301 | met, then the task will fail with an :class:`ImpossibleDependency` error. This way, if the |
|
293 | 302 | scheduler realized that a task can never be run, it won't sit indefinitely in the |
|
294 | 303 | scheduler clogging the pipeline. |
|
295 | 304 | |
|
296 | 305 | The basic cases that are checked: |
|
297 | 306 | |
|
298 | 307 | * depending on nonexistent messages |
|
299 | 308 | * `follow` dependencies were run on more than one machine and `all=True` |
|
300 |
* any dependencies failed and `all=True,success |
|
|
301 |
* all dependencies failed and `all=False,success |
|
|
309 | * any dependencies failed and `all=True,success=True,failures=False` | |
|
310 | * all dependencies failed and `all=False,success=True,failure=False` | |
|
302 | 311 | |
|
303 | 312 | .. warning:: |
|
304 | 313 | |
|
305 | 314 | This analysis has not been proven to be rigorous, so it is likely possible for tasks |
|
306 | 315 | to become impossible to run in obscure situations, so a timeout may be a good choice. |
|
307 | 316 | |
|
308 | 317 | .. _parallel_schedulers: |
|
309 | 318 | |
|
310 | 319 | Schedulers |
|
311 | 320 | ========== |
|
312 | 321 | |
|
313 | 322 | There are a variety of valid ways to determine where jobs should be assigned in a |
|
314 | 323 | load-balancing situation. In IPython, we support several standard schemes, and |
|
315 | 324 | even make it easy to define your own. The scheme can be selected via the ``--scheme`` |
|
316 | 325 | argument to :command:`ipcontrollerz`, or in the :attr:`HubFactory.scheme` attribute |
|
317 | 326 | of a controller config object. |
|
318 | 327 | |
|
319 | 328 | The built-in routing schemes: |
|
320 | 329 | |
|
321 | 330 | To select one of these schemes, simply do:: |
|
322 | 331 | |
|
323 | 332 | $ ipcontrollerz --scheme <schemename> |
|
324 | 333 | for instance: |
|
325 | 334 | $ ipcontrollerz --scheme lru |
|
326 | 335 | |
|
327 | 336 | lru: Least Recently Used |
|
328 | 337 | |
|
329 | 338 | Always assign work to the least-recently-used engine. A close relative of |
|
330 | 339 | round-robin, it will be fair with respect to the number of tasks, agnostic |
|
331 | 340 | with respect to runtime of each task. |
|
332 | 341 | |
|
333 | 342 | plainrandom: Plain Random |
|
334 | 343 | |
|
335 | 344 | Randomly picks an engine on which to run. |
|
336 | 345 | |
|
337 | 346 | twobin: Two-Bin Random |
|
338 | 347 | |
|
339 | 348 | **Requires numpy** |
|
340 | 349 | |
|
341 | 350 | Pick two engines at random, and use the LRU of the two. This is known to be better |
|
342 | 351 | than plain random in many cases, but requires a small amount of computation. |
|
343 | 352 | |
|
344 | 353 | leastload: Least Load |
|
345 | 354 | |
|
346 | 355 | **This is the default scheme** |
|
347 | 356 | |
|
348 | 357 | Always assign tasks to the engine with the fewest outstanding tasks (LRU breaks tie). |
|
349 | 358 | |
|
350 | 359 | weighted: Weighted Two-Bin Random |
|
351 | 360 | |
|
352 | 361 | **Requires numpy** |
|
353 | 362 | |
|
354 | 363 | Pick two engines at random using the number of outstanding tasks as inverse weights, |
|
355 | 364 | and use the one with the lower load. |
|
356 | 365 | |
|
357 | 366 | |
|
358 | 367 | Pure ZMQ Scheduler |
|
359 | 368 | ------------------ |
|
360 | 369 | |
|
361 | 370 | For maximum throughput, the 'pure' scheme is not Python at all, but a C-level |
|
362 | 371 | :class:`MonitoredQueue` from PyZMQ, which uses a ZeroMQ ``XREQ`` socket to perform all |
|
363 | 372 | load-balancing. This scheduler does not support any of the advanced features of the Python |
|
364 | 373 | :class:`.Scheduler`. |
|
365 | 374 | |
|
366 | 375 | Disabled features when using the ZMQ Scheduler: |
|
367 | 376 | |
|
368 | 377 | * Engine unregistration |
|
369 | 378 | Task farming will be disabled if an engine unregisters. |
|
370 | 379 | Further, if an engine is unregistered during computation, the scheduler may not recover. |
|
371 | 380 | * Dependencies |
|
372 | 381 | Since there is no Python logic inside the Scheduler, routing decisions cannot be made |
|
373 | 382 | based on message content. |
|
374 | 383 | * Early destination notification |
|
375 | 384 | The Python schedulers know which engine gets which task, and notify the Hub. This |
|
376 | 385 | allows graceful handling of Engines coming and going. There is no way to know |
|
377 | 386 | where ZeroMQ messages have gone, so there is no way to know what tasks are on which |
|
378 | 387 | engine until they *finish*. This makes recovery from engine shutdown very difficult. |
|
379 | 388 | |
|
380 | 389 | |
|
381 | 390 | .. note:: |
|
382 | 391 | |
|
383 | 392 | TODO: performance comparisons |
|
384 | 393 | |
|
385 | 394 | |
|
386 | 395 | More details |
|
387 | 396 | ============ |
|
388 | 397 | |
|
389 |
The :class:` |
|
|
398 | The :class:`LoadBalancedView` has many more powerful features that allow quite a bit | |
|
390 | 399 | of flexibility in how tasks are defined and run. The next places to look are |
|
391 | 400 | in the following classes: |
|
392 | 401 | |
|
393 |
* :class:`IPython.zmq.parallel. |
|
|
402 | * :class:`IPython.zmq.parallel.view.LoadBalancedView` | |
|
394 | 403 | * :class:`IPython.zmq.parallel.client.AsyncResult` |
|
395 |
* :meth:`IPython.zmq.parallel. |
|
|
404 | * :meth:`IPython.zmq.parallel.view.LoadBalancedView.apply` | |
|
396 | 405 | * :mod:`IPython.zmq.parallel.dependency` |
|
397 | 406 | |
|
398 | 407 | The following is an overview of how to use these classes together: |
|
399 | 408 | |
|
400 |
1. Create a :class:`Client` |
|
|
409 | 1. Create a :class:`Client` and :class:`LoadBalancedView` | |
|
401 | 410 | 2. Define some functions to be run as tasks |
|
402 | 411 | 3. Submit your tasks to using the :meth:`apply` method of your |
|
403 | :class:`Client` instance, specifying `balanced=True`. This signals | |
|
404 | the :class:`Client` to entrust the Scheduler with assigning tasks to engines. | |
|
405 | 4. Use :meth:`Client.get_results` to get the results of the | |
|
412 | :class:`LoadBalancedView` instance. | |
|
413 | 4. Use :meth:`Client.get_result` to get the results of the | |
|
406 | 414 | tasks, or use the :meth:`AsyncResult.get` method of the results to wait |
|
407 | 415 | for and then receive the results. |
|
408 | 416 | |
|
409 | ||
|
410 | 417 | .. seealso:: |
|
411 | 418 | |
|
412 | 419 | A demo of :ref:`DAG Dependencies <dag_dependencies>` with NetworkX and IPython. |
|
1 | NO CONTENT: file was removed |
|
1 | NO CONTENT: file was removed |
General Comments 0
You need to be logged in to leave comments.
Login now