speed up AsyncResult._wait_for_outputs (don't waste 0.01 seconds)

Greatly reduces waiting time on ready() call for large calculations.

IPython/parallel/client/asyncresult.py

@@ -1,708 +1,708 b''
 """AsyncResult objects for the client
 
 Authors:
 
 * MinRK
 """
 #-----------------------------------------------------------------------------
 #  Copyright (C) 2010-2011  The IPython Development Team
 #
 #  Distributed under the terms of the BSD License.  The full license is in
 #  the file COPYING, distributed as part of this software.
 #-----------------------------------------------------------------------------
 
 #-----------------------------------------------------------------------------
 # Imports
 #-----------------------------------------------------------------------------
 
 from __future__ import print_function
 
 import sys
 import time
 from datetime import datetime
 
 from zmq import MessageTracker
 
 from IPython.core.display import clear_output, display, display_pretty
 from IPython.external.decorator import decorator
 from IPython.parallel import error
 
 #-----------------------------------------------------------------------------
 # Functions
 #-----------------------------------------------------------------------------
 
 def _total_seconds(td):
     """timedelta.total_seconds was added in 2.7"""
     try:
         # Python >= 2.7
         return td.total_seconds()
     except AttributeError:
         # Python 2.6
         return 1e-6 * (td.microseconds + (td.seconds + td.days * 24 * 3600) * 10**6)
 
 def _raw_text(s):
     display_pretty(s, raw=True)
 
 #-----------------------------------------------------------------------------
 # Classes
 #-----------------------------------------------------------------------------
 
 # global empty tracker that's always done:
 finished_tracker = MessageTracker()
 
 @decorator
 def check_ready(f, self, *args, **kwargs):
     """Call spin() to sync state prior to calling the method."""
     self.wait(0)
     if not self._ready:
         raise error.TimeoutError("result not ready")
     return f(self, *args, **kwargs)
 
 class AsyncResult(object):
     """Class for representing results of non-blocking calls.
 
     Provides the same interface as :py:class:`multiprocessing.pool.AsyncResult`.
     """
 
     msg_ids = None
     _targets = None
     _tracker = None
     _single_result = False
 
     def __init__(self, client, msg_ids, fname='unknown', targets=None, tracker=None):
         if isinstance(msg_ids, basestring):
             # always a list
             msg_ids = [msg_ids]
         if tracker is None:
             # default to always done
             tracker = finished_tracker
         self._client = client
         self.msg_ids = msg_ids
         self._fname=fname
         self._targets = targets
         self._tracker = tracker
         self._ready = False
         self._outputs_ready = False
         self._success = None
         self._metadata = [ self._client.metadata.get(id) for id in self.msg_ids ]
         if len(msg_ids) == 1:
             self._single_result = not isinstance(targets, (list, tuple))
         else:
             self._single_result = False
 
     def __repr__(self):
         if self._ready:
             return "<%s: finished>"%(self.__class__.__name__)
         else:
             return "<%s: %s>"%(self.__class__.__name__,self._fname)
 
 
     def _reconstruct_result(self, res):
         """Reconstruct our result from actual result list (always a list)
 
         Override me in subclasses for turning a list of results
         into the expected form.
         """
         if self._single_result:
             return res[0]
         else:
             return res
 
     def get(self, timeout=-1):
         """Return the result when it arrives.
 
         If `timeout` is not ``None`` and the result does not arrive within
         `timeout` seconds then ``TimeoutError`` is raised. If the
         remote call raised an exception then that exception will be reraised
         by get() inside a `RemoteError`.
         """
         if not self.ready():
             self.wait(timeout)
 
         if self._ready:
             if self._success:
                 return self._result
             else:
                 raise self._exception
         else:
             raise error.TimeoutError("Result not ready.")
 
     def _check_ready(self):
         if not self.ready():
             raise error.TimeoutError("Result not ready.")
     
     def ready(self):
         """Return whether the call has completed."""
         if not self._ready:
             self.wait(0)
         elif not self._outputs_ready:
             self._wait_for_outputs(0)
         
         return self._ready
 
     def wait(self, timeout=-1):
         """Wait until the result is available or until `timeout` seconds pass.
 
         This method always returns None.
         """
         if self._ready:
             self._wait_for_outputs(timeout)
             return
         self._ready = self._client.wait(self.msg_ids, timeout)
         if self._ready:
             try:
                 results = map(self._client.results.get, self.msg_ids)
                 self._result = results
                 if self._single_result:
                     r = results[0]
                     if isinstance(r, Exception):
                         raise r
                 else:
                     results = error.collect_exceptions(results, self._fname)
                 self._result = self._reconstruct_result(results)
             except Exception as e:
                 self._exception = e
                 self._success = False
             else:
                 self._success = True
             finally:
                 if timeout is None or timeout < 0:
                     # cutoff infinite wait at 10s
                     timeout = 10
                 self._wait_for_outputs(timeout)
 
 
     def successful(self):
         """Return whether the call completed without raising an exception.
 
         Will raise ``AssertionError`` if the result is not ready.
         """
         assert self.ready()
         return self._success
 
     #----------------------------------------------------------------
     # Extra methods not in mp.pool.AsyncResult
     #----------------------------------------------------------------
 
     def get_dict(self, timeout=-1):
         """Get the results as a dict, keyed by engine_id.
 
         timeout behavior is described in `get()`.
         """
 
         results = self.get(timeout)
         engine_ids = [ md['engine_id'] for md in self._metadata ]
         bycount = sorted(engine_ids, key=lambda k: engine_ids.count(k))
         maxcount = bycount.count(bycount[-1])
         if maxcount > 1:
             raise ValueError("Cannot build dict, %i jobs ran on engine #%i"%(
                     maxcount, bycount[-1]))
 
         return dict(zip(engine_ids,results))
 
     @property
     def result(self):
         """result property wrapper for `get(timeout=-1)`."""
         return self.get()
 
     # abbreviated alias:
     r = result
 
     @property
     def metadata(self):
         """property for accessing execution metadata."""
         if self._single_result:
             return self._metadata[0]
         else:
             return self._metadata
 
     @property
     def result_dict(self):
         """result property as a dict."""
         return self.get_dict()
 
     def __dict__(self):
         return self.get_dict(0)
 
     def abort(self):
         """abort my tasks."""
         assert not self.ready(), "Can't abort, I am already done!"
         return self._client.abort(self.msg_ids, targets=self._targets, block=True)
 
     @property
     def sent(self):
         """check whether my messages have been sent."""
         return self._tracker.done
 
     def wait_for_send(self, timeout=-1):
         """wait for pyzmq send to complete.
 
         This is necessary when sending arrays that you intend to edit in-place.
         `timeout` is in seconds, and will raise TimeoutError if it is reached
         before the send completes.
         """
         return self._tracker.wait(timeout)
 
     #-------------------------------------
     # dict-access
     #-------------------------------------
 
     def __getitem__(self, key):
         """getitem returns result value(s) if keyed by int/slice, or metadata if key is str.
         """
         if isinstance(key, int):
             self._check_ready()
             return error.collect_exceptions([self._result[key]], self._fname)[0]
         elif isinstance(key, slice):
             self._check_ready()
             return error.collect_exceptions(self._result[key], self._fname)
         elif isinstance(key, basestring):
             # metadata proxy *does not* require that results are done
             self.wait(0)
             values = [ md[key] for md in self._metadata ]
             if self._single_result:
                 return values[0]
             else:
                 return values
         else:
             raise TypeError("Invalid key type %r, must be 'int','slice', or 'str'"%type(key))
 
     def __getattr__(self, key):
         """getattr maps to getitem for convenient attr access to metadata."""
         try:
             return self.__getitem__(key)
         except (error.TimeoutError, KeyError):
             raise AttributeError("%r object has no attribute %r"%(
                     self.__class__.__name__, key))
 
     # asynchronous iterator:
     def __iter__(self):
         if self._single_result:
             raise TypeError("AsyncResults with a single result are not iterable.")
         try:
             rlist = self.get(0)
         except error.TimeoutError:
             # wait for each result individually
             for msg_id in self.msg_ids:
                 ar = AsyncResult(self._client, msg_id, self._fname)
                 yield ar.get()
         else:
             # already done
             for r in rlist:
                 yield r
     
     def __len__(self):
         return len(self.msg_ids)
     
     #-------------------------------------
     # Sugar methods and attributes
     #-------------------------------------
     
     def timedelta(self, start, end, start_key=min, end_key=max):
         """compute the difference between two sets of timestamps
         
         The default behavior is to use the earliest of the first
         and the latest of the second list, but this can be changed
         by passing a different
         
         Parameters
         ----------
         
         start : one or more datetime objects (e.g. ar.submitted)
         end : one or more datetime objects (e.g. ar.received)
         start_key : callable
             Function to call on `start` to extract the relevant
             entry [defalt: min]
         end_key : callable
             Function to call on `end` to extract the relevant
             entry [default: max]
         
         Returns
         -------
         
         dt : float
             The time elapsed (in seconds) between the two selected timestamps.
         """
         if not isinstance(start, datetime):
             # handle single_result AsyncResults, where ar.stamp is single object,
             # not a list
             start = start_key(start)
         if not isinstance(end, datetime):
             # handle single_result AsyncResults, where ar.stamp is single object,
             # not a list
             end = end_key(end)
         return _total_seconds(end - start)
         
     @property
     def progress(self):
         """the number of tasks which have been completed at this point.
         
         Fractional progress would be given by 1.0 * ar.progress / len(ar)
         """
         self.wait(0)
         return len(self) - len(set(self.msg_ids).intersection(self._client.outstanding))
     
     @property
     def elapsed(self):
         """elapsed time since initial submission"""
         if self.ready():
             return self.wall_time
         
         now = submitted = datetime.now()
         for msg_id in self.msg_ids:
             if msg_id in self._client.metadata:
                 stamp = self._client.metadata[msg_id]['submitted']
                 if stamp and stamp < submitted:
                     submitted = stamp
         return _total_seconds(now-submitted)
     
     @property
     @check_ready
     def serial_time(self):
         """serial computation time of a parallel calculation
         
         Computed as the sum of (completed-started) of each task
         """
         t = 0
         for md in self._metadata:
             t += _total_seconds(md['completed'] - md['started'])
         return t
     
     @property
     @check_ready
     def wall_time(self):
         """actual computation time of a parallel calculation
         
         Computed as the time between the latest `received` stamp
         and the earliest `submitted`.
         
         Only reliable if Client was spinning/waiting when the task finished, because
         the `received` timestamp is created when a result is pulled off of the zmq queue,
         which happens as a result of `client.spin()`.
         
         For similar comparison of other timestamp pairs, check out AsyncResult.timedelta.
         
         """
         return self.timedelta(self.submitted, self.received)
     
     def wait_interactive(self, interval=1., timeout=-1):
         """interactive wait, printing progress at regular intervals"""
         if timeout is None:
             timeout = -1
         N = len(self)
         tic = time.time()
         while not self.ready() and (timeout < 0 or time.time() - tic <= timeout):
             self.wait(interval)
             clear_output()
             print("%4i/%i tasks finished after %4i s" % (self.progress, N, self.elapsed), end="")
             sys.stdout.flush()
         print()
         print("done")
     
     def _republish_displaypub(self, content, eid):
         """republish individual displaypub content dicts"""
         try:
             ip = get_ipython()
         except NameError:
             # displaypub is meaningless outside IPython
             return
         md = content['metadata'] or {}
         md['engine'] = eid
         ip.display_pub.publish(content['source'], content['data'], md)
     
     def _display_stream(self, text, prefix='', file=None):
         if not text:
             # nothing to display
             return
         if file is None:
             file = sys.stdout
         end = '' if text.endswith('\n') else '\n'
         
         multiline = text.count('\n') > int(text.endswith('\n'))
         if prefix and multiline and not text.startswith('\n'):
             prefix = prefix + '\n'
         print("%s%s" % (prefix, text), file=file, end=end)
         
     
     def _display_single_result(self):
         self._display_stream(self.stdout)
         self._display_stream(self.stderr, file=sys.stderr)
         
         try:
             get_ipython()
         except NameError:
             # displaypub is meaningless outside IPython
             return
         
         for output in self.outputs:
             self._republish_displaypub(output, self.engine_id)
         
         if self.pyout is not None:
             display(self.get())
     
     def _wait_for_outputs(self, timeout=-1):
         """wait for the 'status=idle' message that indicates we have all outputs
         """
         if self._outputs_ready or not self._success:
             # don't wait on errors
             return
         
         # cast None to -1 for infinite timeout
         if timeout is None:
             timeout = -1
         
         tic = time.time()
-        self._client._flush_iopub(self._client._iopub_socket)
-        self._outputs_ready = all(md['outputs_ready'] for md in self._metadata)
-        while not self._outputs_ready:
-            time.sleep(0.01)
+        while True:
             self._client._flush_iopub(self._client._iopub_socket)
-            self._outputs_ready = all(md['outputs_ready'] for md in self._metadata)
-            if timeout >= 0 and time.time() > tic + timeout:
+            self._outputs_ready = all(md['outputs_ready']
+                                      for md in self._metadata)
+            if self._outputs_ready or \
+               (timeout >= 0 and time.time() > tic + timeout):
                 break
+            time.sleep(0.01)
     
     @check_ready
     def display_outputs(self, groupby="type"):
         """republish the outputs of the computation
         
         Parameters
         ----------
         
         groupby : str [default: type]
             if 'type':
                 Group outputs by type (show all stdout, then all stderr, etc.):
                 
                 [stdout:1] foo
                 [stdout:2] foo
                 [stderr:1] bar
                 [stderr:2] bar
             if 'engine':
                 Display outputs for each engine before moving on to the next:
                 
                 [stdout:1] foo
                 [stderr:1] bar
                 [stdout:2] foo
                 [stderr:2] bar
                 
             if 'order':
                 Like 'type', but further collate individual displaypub
                 outputs.  This is meant for cases of each command producing
                 several plots, and you would like to see all of the first
                 plots together, then all of the second plots, and so on.
         """
         if self._single_result:
             self._display_single_result()
             return
         
         stdouts = self.stdout
         stderrs = self.stderr
         pyouts  = self.pyout
         output_lists = self.outputs
         results = self.get()
         
         targets = self.engine_id
         
         if groupby == "engine":
             for eid,stdout,stderr,outputs,r,pyout in zip(
                     targets, stdouts, stderrs, output_lists, results, pyouts
                 ):
                 self._display_stream(stdout, '[stdout:%i] ' % eid)
                 self._display_stream(stderr, '[stderr:%i] ' % eid, file=sys.stderr)
                 
                 try:
                     get_ipython()
                 except NameError:
                     # displaypub is meaningless outside IPython
                     return 
                 
                 if outputs or pyout is not None:
                     _raw_text('[output:%i]' % eid)
                 
                 for output in outputs:
                     self._republish_displaypub(output, eid)
                 
                 if pyout is not None:
                     display(r)
         
         elif groupby in ('type', 'order'):
             # republish stdout:
             for eid,stdout in zip(targets, stdouts):
                 self._display_stream(stdout, '[stdout:%i] ' % eid)
         
             # republish stderr:
             for eid,stderr in zip(targets, stderrs):
                 self._display_stream(stderr, '[stderr:%i] ' % eid, file=sys.stderr)
         
             try:
                 get_ipython()
             except NameError:
                 # displaypub is meaningless outside IPython
                 return
             
             if groupby == 'order':
                 output_dict = dict((eid, outputs) for eid,outputs in zip(targets, output_lists))
                 N = max(len(outputs) for outputs in output_lists)
                 for i in range(N):
                     for eid in targets:
                         outputs = output_dict[eid]
                         if len(outputs) >= N:
                             _raw_text('[output:%i]' % eid)
                             self._republish_displaypub(outputs[i], eid)
             else:
                 # republish displaypub output
                 for eid,outputs in zip(targets, output_lists):
                     if outputs:
                         _raw_text('[output:%i]' % eid)
                     for output in outputs:
                         self._republish_displaypub(output, eid)
         
             # finally, add pyout:
             for eid,r,pyout in zip(targets, results, pyouts):
                 if pyout is not None:
                     display(r)
         
         else:
             raise ValueError("groupby must be one of 'type', 'engine', 'collate', not %r" % groupby)
         
         
 
 
 class AsyncMapResult(AsyncResult):
     """Class for representing results of non-blocking gathers.
 
     This will properly reconstruct the gather.
     
     This class is iterable at any time, and will wait on results as they come.
     
     If ordered=False, then the first results to arrive will come first, otherwise
     results will be yielded in the order they were submitted.
     
     """
 
     def __init__(self, client, msg_ids, mapObject, fname='', ordered=True):
         AsyncResult.__init__(self, client, msg_ids, fname=fname)
         self._mapObject = mapObject
         self._single_result = False
         self.ordered = ordered
 
     def _reconstruct_result(self, res):
         """Perform the gather on the actual results."""
         return self._mapObject.joinPartitions(res)
 
     # asynchronous iterator:
     def __iter__(self):
         it = self._ordered_iter if self.ordered else self._unordered_iter
         for r in it():
             yield r
 
     # asynchronous ordered iterator:
     def _ordered_iter(self):
         """iterator for results *as they arrive*, preserving submission order."""
         try:
             rlist = self.get(0)
         except error.TimeoutError:
             # wait for each result individually
             for msg_id in self.msg_ids:
                 ar = AsyncResult(self._client, msg_id, self._fname)
                 rlist = ar.get()
                 try:
                     for r in rlist:
                         yield r
                 except TypeError:
                     # flattened, not a list
                     # this could get broken by flattened data that returns iterables
                     # but most calls to map do not expose the `flatten` argument
                     yield rlist
         else:
             # already done
             for r in rlist:
                 yield r
 
     # asynchronous unordered iterator:
     def _unordered_iter(self):
         """iterator for results *as they arrive*, on FCFS basis, ignoring submission order."""
         try:
             rlist = self.get(0)
         except error.TimeoutError:
             pending = set(self.msg_ids)
             while pending:
                 try:
                     self._client.wait(pending, 1e-3)
                 except error.TimeoutError:
                     # ignore timeout error, because that only means
                     # *some* jobs are outstanding
                     pass
                 # update ready set with those no longer outstanding:
                 ready = pending.difference(self._client.outstanding)
                 # update pending to exclude those that are finished
                 pending = pending.difference(ready)
                 while ready:
                     msg_id = ready.pop()
                     ar = AsyncResult(self._client, msg_id, self._fname)
                     rlist = ar.get()
                     try:
                         for r in rlist:
                             yield r
                     except TypeError:
                         # flattened, not a list
                         # this could get broken by flattened data that returns iterables
                         # but most calls to map do not expose the `flatten` argument
                         yield rlist
         else:
             # already done
             for r in rlist:
                 yield r
 
 
 class AsyncHubResult(AsyncResult):
     """Class to wrap pending results that must be requested from the Hub.
 
     Note that waiting/polling on these objects requires polling the Hubover the network,
     so use `AsyncHubResult.wait()` sparingly.
     """
 
     def _wait_for_outputs(self, timeout=-1):
         """no-op, because HubResults are never incomplete"""
         self._outputs_ready = True
     
     def wait(self, timeout=-1):
         """wait for result to complete."""
         start = time.time()
         if self._ready:
             return
         local_ids = filter(lambda msg_id: msg_id in self._client.outstanding, self.msg_ids)
         local_ready = self._client.wait(local_ids, timeout)
         if local_ready:
             remote_ids = filter(lambda msg_id: msg_id not in self._client.results, self.msg_ids)
             if not remote_ids:
                 self._ready = True
             else:
                 rdict = self._client.result_status(remote_ids, status_only=False)
                 pending = rdict['pending']
                 while pending and (timeout < 0 or time.time() < start+timeout):
                     rdict = self._client.result_status(remote_ids, status_only=False)
                     pending = rdict['pending']
                     if pending:
                         time.sleep(0.1)
                 if not pending:
                     self._ready = True
         if self._ready:
             try:
                 results = map(self._client.results.get, self.msg_ids)
                 self._result = results
                 if self._single_result:
                     r = results[0]
                     if isinstance(r, Exception):
                         raise r
                 else:
                     results = error.collect_exceptions(results, self._fname)
                 self._result = self._reconstruct_result(results)
             except Exception as e:
                 self._exception = e
                 self._success = False
             else:
                 self._success = True
             finally:
                 self._metadata = map(self._client.metadata.get, self.msg_ids)
 
 __all__ = ['AsyncResult', 'AsyncMapResult', 'AsyncHubResult']