Show More
@@ -1,355 +1,355 b'' | |||
|
1 | 1 | # worker.py - master-slave parallelism support |
|
2 | 2 | # |
|
3 | 3 | # Copyright 2013 Facebook, Inc. |
|
4 | 4 | # |
|
5 | 5 | # This software may be used and distributed according to the terms of the |
|
6 | 6 | # GNU General Public License version 2 or any later version. |
|
7 | 7 | |
|
8 | 8 | from __future__ import absolute_import |
|
9 | 9 | |
|
10 | 10 | import errno |
|
11 | 11 | import os |
|
12 | 12 | import signal |
|
13 | 13 | import sys |
|
14 | 14 | import threading |
|
15 | 15 | import time |
|
16 | 16 | |
|
17 | 17 | try: |
|
18 | 18 | import selectors |
|
19 | 19 | selectors.BaseSelector |
|
20 | 20 | except ImportError: |
|
21 | 21 | from .thirdparty import selectors2 as selectors |
|
22 | 22 | |
|
23 | 23 | from .i18n import _ |
|
24 | 24 | from . import ( |
|
25 | 25 | encoding, |
|
26 | 26 | error, |
|
27 | 27 | pycompat, |
|
28 | 28 | scmutil, |
|
29 | 29 | util, |
|
30 | 30 | ) |
|
31 | 31 | |
|
32 | 32 | def countcpus(): |
|
33 | 33 | '''try to count the number of CPUs on the system''' |
|
34 | 34 | |
|
35 | 35 | # posix |
|
36 | 36 | try: |
|
37 | 37 | n = int(os.sysconf(r'SC_NPROCESSORS_ONLN')) |
|
38 | 38 | if n > 0: |
|
39 | 39 | return n |
|
40 | 40 | except (AttributeError, ValueError): |
|
41 | 41 | pass |
|
42 | 42 | |
|
43 | 43 | # windows |
|
44 | 44 | try: |
|
45 | 45 | n = int(encoding.environ['NUMBER_OF_PROCESSORS']) |
|
46 | 46 | if n > 0: |
|
47 | 47 | return n |
|
48 | 48 | except (KeyError, ValueError): |
|
49 | 49 | pass |
|
50 | 50 | |
|
51 | 51 | return 1 |
|
52 | 52 | |
|
53 | 53 | def _numworkers(ui): |
|
54 | 54 | s = ui.config('worker', 'numcpus') |
|
55 | 55 | if s: |
|
56 | 56 | try: |
|
57 | 57 | n = int(s) |
|
58 | 58 | if n >= 1: |
|
59 | 59 | return n |
|
60 | 60 | except ValueError: |
|
61 | 61 | raise error.Abort(_('number of cpus must be an integer')) |
|
62 | 62 | return min(max(countcpus(), 4), 32) |
|
63 | 63 | |
|
64 | 64 | if pycompat.isposix or pycompat.iswindows: |
|
65 | _startupcost = 0.01 | |
|
65 | _STARTUP_COST = 0.01 | |
|
66 | 66 | else: |
|
67 | _startupcost = 1e30 | |
|
67 | _STARTUP_COST = 1e30 | |
|
68 | 68 | |
|
69 | 69 | def worthwhile(ui, costperop, nops): |
|
70 | 70 | '''try to determine whether the benefit of multiple processes can |
|
71 | 71 | outweigh the cost of starting them''' |
|
72 | 72 | linear = costperop * nops |
|
73 | 73 | workers = _numworkers(ui) |
|
74 |
benefit = linear - (_ |
|
|
74 | benefit = linear - (_STARTUP_COST * workers + linear / workers) | |
|
75 | 75 | return benefit >= 0.15 |
|
76 | 76 | |
|
77 | 77 | def worker(ui, costperarg, func, staticargs, args): |
|
78 | 78 | '''run a function, possibly in parallel in multiple worker |
|
79 | 79 | processes. |
|
80 | 80 | |
|
81 | 81 | returns a progress iterator |
|
82 | 82 | |
|
83 | 83 | costperarg - cost of a single task |
|
84 | 84 | |
|
85 | 85 | func - function to run |
|
86 | 86 | |
|
87 | 87 | staticargs - arguments to pass to every invocation of the function |
|
88 | 88 | |
|
89 | 89 | args - arguments to split into chunks, to pass to individual |
|
90 | 90 | workers |
|
91 | 91 | ''' |
|
92 | 92 | enabled = ui.configbool('worker', 'enabled') |
|
93 | 93 | if enabled and worthwhile(ui, costperarg, len(args)): |
|
94 | 94 | return _platformworker(ui, func, staticargs, args) |
|
95 | 95 | return func(*staticargs + (args,)) |
|
96 | 96 | |
|
97 | 97 | def _posixworker(ui, func, staticargs, args): |
|
98 | 98 | workers = _numworkers(ui) |
|
99 | 99 | oldhandler = signal.getsignal(signal.SIGINT) |
|
100 | 100 | signal.signal(signal.SIGINT, signal.SIG_IGN) |
|
101 | 101 | pids, problem = set(), [0] |
|
102 | 102 | def killworkers(): |
|
103 | 103 | # unregister SIGCHLD handler as all children will be killed. This |
|
104 | 104 | # function shouldn't be interrupted by another SIGCHLD; otherwise pids |
|
105 | 105 | # could be updated while iterating, which would cause inconsistency. |
|
106 | 106 | signal.signal(signal.SIGCHLD, oldchldhandler) |
|
107 | 107 | # if one worker bails, there's no good reason to wait for the rest |
|
108 | 108 | for p in pids: |
|
109 | 109 | try: |
|
110 | 110 | os.kill(p, signal.SIGTERM) |
|
111 | 111 | except OSError as err: |
|
112 | 112 | if err.errno != errno.ESRCH: |
|
113 | 113 | raise |
|
114 | 114 | def waitforworkers(blocking=True): |
|
115 | 115 | for pid in pids.copy(): |
|
116 | 116 | p = st = 0 |
|
117 | 117 | while True: |
|
118 | 118 | try: |
|
119 | 119 | p, st = os.waitpid(pid, (0 if blocking else os.WNOHANG)) |
|
120 | 120 | break |
|
121 | 121 | except OSError as e: |
|
122 | 122 | if e.errno == errno.EINTR: |
|
123 | 123 | continue |
|
124 | 124 | elif e.errno == errno.ECHILD: |
|
125 | 125 | # child would already be reaped, but pids yet been |
|
126 | 126 | # updated (maybe interrupted just after waitpid) |
|
127 | 127 | pids.discard(pid) |
|
128 | 128 | break |
|
129 | 129 | else: |
|
130 | 130 | raise |
|
131 | 131 | if not p: |
|
132 | 132 | # skip subsequent steps, because child process should |
|
133 | 133 | # be still running in this case |
|
134 | 134 | continue |
|
135 | 135 | pids.discard(p) |
|
136 | 136 | st = _exitstatus(st) |
|
137 | 137 | if st and not problem[0]: |
|
138 | 138 | problem[0] = st |
|
139 | 139 | def sigchldhandler(signum, frame): |
|
140 | 140 | waitforworkers(blocking=False) |
|
141 | 141 | if problem[0]: |
|
142 | 142 | killworkers() |
|
143 | 143 | oldchldhandler = signal.signal(signal.SIGCHLD, sigchldhandler) |
|
144 | 144 | ui.flush() |
|
145 | 145 | parentpid = os.getpid() |
|
146 | 146 | pipes = [] |
|
147 | 147 | for pargs in partition(args, workers): |
|
148 | 148 | # Every worker gets its own pipe to send results on, so we don't have to |
|
149 | 149 | # implement atomic writes larger than PIPE_BUF. Each forked process has |
|
150 | 150 | # its own pipe's descriptors in the local variables, and the parent |
|
151 | 151 | # process has the full list of pipe descriptors (and it doesn't really |
|
152 | 152 | # care what order they're in). |
|
153 | 153 | rfd, wfd = os.pipe() |
|
154 | 154 | pipes.append((rfd, wfd)) |
|
155 | 155 | # make sure we use os._exit in all worker code paths. otherwise the |
|
156 | 156 | # worker may do some clean-ups which could cause surprises like |
|
157 | 157 | # deadlock. see sshpeer.cleanup for example. |
|
158 | 158 | # override error handling *before* fork. this is necessary because |
|
159 | 159 | # exception (signal) may arrive after fork, before "pid =" assignment |
|
160 | 160 | # completes, and other exception handler (dispatch.py) can lead to |
|
161 | 161 | # unexpected code path without os._exit. |
|
162 | 162 | ret = -1 |
|
163 | 163 | try: |
|
164 | 164 | pid = os.fork() |
|
165 | 165 | if pid == 0: |
|
166 | 166 | signal.signal(signal.SIGINT, oldhandler) |
|
167 | 167 | signal.signal(signal.SIGCHLD, oldchldhandler) |
|
168 | 168 | |
|
169 | 169 | def workerfunc(): |
|
170 | 170 | for r, w in pipes[:-1]: |
|
171 | 171 | os.close(r) |
|
172 | 172 | os.close(w) |
|
173 | 173 | os.close(rfd) |
|
174 | 174 | for result in func(*(staticargs + (pargs,))): |
|
175 | 175 | os.write(wfd, util.pickle.dumps(result)) |
|
176 | 176 | return 0 |
|
177 | 177 | |
|
178 | 178 | ret = scmutil.callcatch(ui, workerfunc) |
|
179 | 179 | except: # parent re-raises, child never returns |
|
180 | 180 | if os.getpid() == parentpid: |
|
181 | 181 | raise |
|
182 | 182 | exctype = sys.exc_info()[0] |
|
183 | 183 | force = not issubclass(exctype, KeyboardInterrupt) |
|
184 | 184 | ui.traceback(force=force) |
|
185 | 185 | finally: |
|
186 | 186 | if os.getpid() != parentpid: |
|
187 | 187 | try: |
|
188 | 188 | ui.flush() |
|
189 | 189 | except: # never returns, no re-raises |
|
190 | 190 | pass |
|
191 | 191 | finally: |
|
192 | 192 | os._exit(ret & 255) |
|
193 | 193 | pids.add(pid) |
|
194 | 194 | selector = selectors.DefaultSelector() |
|
195 | 195 | for rfd, wfd in pipes: |
|
196 | 196 | os.close(wfd) |
|
197 | 197 | selector.register(os.fdopen(rfd, r'rb', 0), selectors.EVENT_READ) |
|
198 | 198 | def cleanup(): |
|
199 | 199 | signal.signal(signal.SIGINT, oldhandler) |
|
200 | 200 | waitforworkers() |
|
201 | 201 | signal.signal(signal.SIGCHLD, oldchldhandler) |
|
202 | 202 | status = problem[0] |
|
203 | 203 | if status: |
|
204 | 204 | if status < 0: |
|
205 | 205 | os.kill(os.getpid(), -status) |
|
206 | 206 | sys.exit(status) |
|
207 | 207 | try: |
|
208 | 208 | openpipes = len(pipes) |
|
209 | 209 | while openpipes > 0: |
|
210 | 210 | for key, events in selector.select(): |
|
211 | 211 | try: |
|
212 | 212 | yield util.pickle.load(key.fileobj) |
|
213 | 213 | except EOFError: |
|
214 | 214 | selector.unregister(key.fileobj) |
|
215 | 215 | key.fileobj.close() |
|
216 | 216 | openpipes -= 1 |
|
217 | 217 | except IOError as e: |
|
218 | 218 | if e.errno == errno.EINTR: |
|
219 | 219 | continue |
|
220 | 220 | raise |
|
221 | 221 | except: # re-raises |
|
222 | 222 | killworkers() |
|
223 | 223 | cleanup() |
|
224 | 224 | raise |
|
225 | 225 | cleanup() |
|
226 | 226 | |
|
227 | 227 | def _posixexitstatus(code): |
|
228 | 228 | '''convert a posix exit status into the same form returned by |
|
229 | 229 | os.spawnv |
|
230 | 230 | |
|
231 | 231 | returns None if the process was stopped instead of exiting''' |
|
232 | 232 | if os.WIFEXITED(code): |
|
233 | 233 | return os.WEXITSTATUS(code) |
|
234 | 234 | elif os.WIFSIGNALED(code): |
|
235 | 235 | return -os.WTERMSIG(code) |
|
236 | 236 | |
|
237 | 237 | def _windowsworker(ui, func, staticargs, args): |
|
238 | 238 | class Worker(threading.Thread): |
|
239 | 239 | def __init__(self, taskqueue, resultqueue, func, staticargs, |
|
240 | 240 | group=None, target=None, name=None, verbose=None): |
|
241 | 241 | threading.Thread.__init__(self, group=group, target=target, |
|
242 | 242 | name=name, verbose=verbose) |
|
243 | 243 | self._taskqueue = taskqueue |
|
244 | 244 | self._resultqueue = resultqueue |
|
245 | 245 | self._func = func |
|
246 | 246 | self._staticargs = staticargs |
|
247 | 247 | self._interrupted = False |
|
248 | 248 | self.daemon = True |
|
249 | 249 | self.exception = None |
|
250 | 250 | |
|
251 | 251 | def interrupt(self): |
|
252 | 252 | self._interrupted = True |
|
253 | 253 | |
|
254 | 254 | def run(self): |
|
255 | 255 | try: |
|
256 | 256 | while not self._taskqueue.empty(): |
|
257 | 257 | try: |
|
258 | 258 | args = self._taskqueue.get_nowait() |
|
259 | 259 | for res in self._func(*self._staticargs + (args,)): |
|
260 | 260 | self._resultqueue.put(res) |
|
261 | 261 | # threading doesn't provide a native way to |
|
262 | 262 | # interrupt execution. handle it manually at every |
|
263 | 263 | # iteration. |
|
264 | 264 | if self._interrupted: |
|
265 | 265 | return |
|
266 | 266 | except pycompat.queue.Empty: |
|
267 | 267 | break |
|
268 | 268 | except Exception as e: |
|
269 | 269 | # store the exception such that the main thread can resurface |
|
270 | 270 | # it as if the func was running without workers. |
|
271 | 271 | self.exception = e |
|
272 | 272 | raise |
|
273 | 273 | |
|
274 | 274 | threads = [] |
|
275 | 275 | def trykillworkers(): |
|
276 | 276 | # Allow up to 1 second to clean worker threads nicely |
|
277 | 277 | cleanupend = time.time() + 1 |
|
278 | 278 | for t in threads: |
|
279 | 279 | t.interrupt() |
|
280 | 280 | for t in threads: |
|
281 | 281 | remainingtime = cleanupend - time.time() |
|
282 | 282 | t.join(remainingtime) |
|
283 | 283 | if t.is_alive(): |
|
284 | 284 | # pass over the workers joining failure. it is more |
|
285 | 285 | # important to surface the inital exception than the |
|
286 | 286 | # fact that one of workers may be processing a large |
|
287 | 287 | # task and does not get to handle the interruption. |
|
288 | 288 | ui.warn(_("failed to kill worker threads while " |
|
289 | 289 | "handling an exception\n")) |
|
290 | 290 | return |
|
291 | 291 | |
|
292 | 292 | workers = _numworkers(ui) |
|
293 | 293 | resultqueue = pycompat.queue.Queue() |
|
294 | 294 | taskqueue = pycompat.queue.Queue() |
|
295 | 295 | # partition work to more pieces than workers to minimize the chance |
|
296 | 296 | # of uneven distribution of large tasks between the workers |
|
297 | 297 | for pargs in partition(args, workers * 20): |
|
298 | 298 | taskqueue.put(pargs) |
|
299 | 299 | for _i in range(workers): |
|
300 | 300 | t = Worker(taskqueue, resultqueue, func, staticargs) |
|
301 | 301 | threads.append(t) |
|
302 | 302 | t.start() |
|
303 | 303 | try: |
|
304 | 304 | while len(threads) > 0: |
|
305 | 305 | while not resultqueue.empty(): |
|
306 | 306 | yield resultqueue.get() |
|
307 | 307 | threads[0].join(0.05) |
|
308 | 308 | finishedthreads = [_t for _t in threads if not _t.is_alive()] |
|
309 | 309 | for t in finishedthreads: |
|
310 | 310 | if t.exception is not None: |
|
311 | 311 | raise t.exception |
|
312 | 312 | threads.remove(t) |
|
313 | 313 | except (Exception, KeyboardInterrupt): # re-raises |
|
314 | 314 | trykillworkers() |
|
315 | 315 | raise |
|
316 | 316 | while not resultqueue.empty(): |
|
317 | 317 | yield resultqueue.get() |
|
318 | 318 | |
|
319 | 319 | if pycompat.iswindows: |
|
320 | 320 | _platformworker = _windowsworker |
|
321 | 321 | else: |
|
322 | 322 | _platformworker = _posixworker |
|
323 | 323 | _exitstatus = _posixexitstatus |
|
324 | 324 | |
|
325 | 325 | def partition(lst, nslices): |
|
326 | 326 | '''partition a list into N slices of roughly equal size |
|
327 | 327 | |
|
328 | 328 | The current strategy takes every Nth element from the input. If |
|
329 | 329 | we ever write workers that need to preserve grouping in input |
|
330 | 330 | we should consider allowing callers to specify a partition strategy. |
|
331 | 331 | |
|
332 | 332 | mpm is not a fan of this partitioning strategy when files are involved. |
|
333 | 333 | In his words: |
|
334 | 334 | |
|
335 | 335 | Single-threaded Mercurial makes a point of creating and visiting |
|
336 | 336 | files in a fixed order (alphabetical). When creating files in order, |
|
337 | 337 | a typical filesystem is likely to allocate them on nearby regions on |
|
338 | 338 | disk. Thus, when revisiting in the same order, locality is maximized |
|
339 | 339 | and various forms of OS and disk-level caching and read-ahead get a |
|
340 | 340 | chance to work. |
|
341 | 341 | |
|
342 | 342 | This effect can be quite significant on spinning disks. I discovered it |
|
343 | 343 | circa Mercurial v0.4 when revlogs were named by hashes of filenames. |
|
344 | 344 | Tarring a repo and copying it to another disk effectively randomized |
|
345 | 345 | the revlog ordering on disk by sorting the revlogs by hash and suddenly |
|
346 | 346 | performance of my kernel checkout benchmark dropped by ~10x because the |
|
347 | 347 | "working set" of sectors visited no longer fit in the drive's cache and |
|
348 | 348 | the workload switched from streaming to random I/O. |
|
349 | 349 | |
|
350 | 350 | What we should really be doing is have workers read filenames from a |
|
351 | 351 | ordered queue. This preserves locality and also keeps any worker from |
|
352 | 352 | getting more than one file out of balance. |
|
353 | 353 | ''' |
|
354 | 354 | for i in range(nslices): |
|
355 | 355 | yield lst[i::nslices] |
General Comments 0
You need to be logged in to leave comments.
Login now