This CL updates the parallel job library in the au test harness to be more robust.

au_test_harness/parallel_test_job.py

 # Copyright (c) 2011 The Chromium OS Authors. All rights reserved.
 # Use of this source code is governed by a BSD-style license that can be
 # found in the LICENSE file.
 
 """Module containing methods/classes related to running parallel test jobs."""
 
+import multiprocessing
 import sys
 import threading
 import time
 
 import cros_build_lib as cros_lib
 
-class ParallelJob(threading.Thread):
-  """Small wrapper for threading.  Thread that releases a semaphores on exit."""
+class ParallelJobTimeoutError(Exception):
+  """Thrown when a job ran for longer than expected."""
+  pass
 
-  def __init__(self, starting_semaphore, ending_semaphore, target, args):
+
+class ParallelJob(multiprocessing.Process):
+  """Small wrapper for Process that stores output of a its target method."""

[dgarrett, 2011/04/08 03:56:42]

its -> it's

[sosa, 2011/04/08 18:59:04]

it's == it is :)

On 2011/04/08 03:56:42, dgarrett wrote:

+
+  MAX_TIMEOUT = 1800

[dgarrett, 2011/04/08 03:56:42]

Units?

[sosa, 2011/04/08 18:59:04]

Done.

+  SLEEP_TIMEOUT = 180
+
+  def __init__(self, starting_semaphore, target, args):
     """Initializes an instance of a job.
 
     Args:
       starting_semaphore: Semaphore used by caller to wait on such that
-        there isn't more than a certain number of threads running.  Should
-        be initialized to a value for the number of threads wanting to be run
-        at a time.
-      ending_semaphore:  Semaphore is released every time a job ends.  Should be
-        initialized to 0 before starting first job.  Should be acquired once for
-        each job.  Threading.Thread.join() has a bug where if the run function
-        terminates too quickly join() will hang forever.
+        there isn't more than a certain number of parallel_jobs running.  Should
+        be initialized to a value for the number of parallel_jobs wanting to be
+        run at a time.
       target:  The func to run.
       args:  Args to pass to the fun.
     """
-    threading.Thread.__init__(self, target=target, args=args)
+    super(ParallelJob, self).__init__(target=target, args=args)
     self._target = target
     self._args = args
     self._starting_semaphore = starting_semaphore
-    self._ending_semaphore = ending_semaphore
-    self._output = None
-    self._completed = False
 
   def run(self):
     """Thread override.  Runs the method specified and sets output."""
     try:
-      self._output = self._target(*self._args)
+      self._target(*self._args)
     finally:
-      # Our own clean up.
-      self._Cleanup()
-      self._completed = True
-      # From threading.py to avoid a refcycle.
-      del self._target, self._args
+      self._starting_semaphore.release()
 
-  def GetOutput(self):
-    """Returns the output of the method run."""
-    assert self._completed, 'GetOutput called before thread was run.'
-    return self._output
+  @classmethod
+  def WaitUntilJobsComplete(cls, parallel_jobs):
+    """Waits until all parallel_jobs have completed before returning.
 
-  def _Cleanup(self):
-    """Releases semaphores for a waiting caller."""
-    self._starting_semaphore.release()
-    self._ending_semaphore.release()
+    Given an array of parallel_jobs, returns once all parallel_jobs have
+    completed or a max timeout is reached.
+
+    Raises:
+      ParallelJobTimeoutError:  if max timeout is reached.
+    """
+    def GetCurrentActiveCount():
+      """Returns the (number of active jobs, first active job)."""
+      active_count = 0
+      active_job = None
+      for parallel_job in parallel_jobs:
+        if parallel_job.is_alive():
+          active_count += 1
+          if not active_job:
+            active_job = parallel_job
+
+      return (active_count, parallel_job)
+
+    start_time = time.time()
+    while (time.time() - start_time) < cls.MAX_TIMEOUT:
+      (active_count, active_job) = GetCurrentActiveCount()
+      if active_count == 0:
+        return
+      else:
+        print >> sys.stderr, (
+            'Process Pool Active: Waiting on %d/%d jobs to complete' %
+            (active_count, len(parallel_jobs)))
+        active_job.join(cls.SLEEP_TIMEOUT)
+
+    for parallel_job in parallel_jobs:
+      if parallel_job.is_alive():
+        parallel_job.terminate()
+
+    raise ParallelJobTimeoutError('Exceeded max time of %d seconds to wait for '
+                                  'job completion.' % cls.MAX_TIMEOUT)
 
   def __str__(self):
     return '%s(%s)' % (self._target, self._args)
 
 
 def RunParallelJobs(number_of_simultaneous_jobs, jobs, jobs_args,
                     print_status):
   """Runs set number of specified jobs in parallel.
 
   Args:
-    number_of_simultaneous_jobs:  Max number of threads to be run in parallel.
+    number_of_simultaneous_jobs:  Max number of parallel_jobs to be run in
+      parallel.
     jobs:  Array of methods to run.
     jobs_args:  Array of args associated with method calls.
     print_status: True if you'd like this to print out .'s as it runs jobs.
   Returns:
-    Returns an array of results corresponding to each thread.
+    Returns an array of results corresponding to each parallel_job.
   """
   def _TwoTupleize(x, y):
     return (x, y)
 
-  threads = []
-  job_start_semaphore = threading.Semaphore(number_of_simultaneous_jobs)
-  join_semaphore = threading.Semaphore(0)
+  def ProcessOutputWrapper(func, args, output):
+    """Simple function wrapper that puts the output of a function in a queue."""
+    output.put(func(*args))
+
   assert len(jobs) == len(jobs_args), 'Length of args array is wrong.'
-
-  # Create the parallel jobs.
-  for job, args in map(_TwoTupleize, jobs, jobs_args):
-    thread = ParallelJob(job_start_semaphore, join_semaphore, target=job,
-                         args=args)
-    threads.append(thread)
-
   # Cache sudo access.
   cros_lib.RunCommand(['sudo', 'echo', 'Caching sudo credentials'],
                       print_cmd=False, redirect_stdout=True,
                       redirect_stderr=True)
 
+  parallel_jobs = []
+  output_array = []
+
+  # Semaphore used to create a Process Pool.
+  job_start_semaphore = threading.Semaphore(number_of_simultaneous_jobs)

[dgarrett, 2011/04/08 03:56:42]

Is this semaphore really safe across processes instead of threads?

[sosa, 2011/04/08 18:59:04]

I think you're right, switching to multiprocessing.Semaphore instead and tested
with -j 2 and saw all 4 tests still get run.

On 2011/04/08 03:56:42, dgarrett wrote:

+
+  # Create the parallel jobs.
+  for job, args in map(_TwoTupleize, jobs, jobs_args):

[dgarrett, 2011/04/08 03:56:42]

_TwoTupleize could be replaced in-line with "lambda x, y: (x, y)" but that's
just a different style, not sure if it's better.

[sosa, 2011/04/08 18:59:04]

Done.

+    output = multiprocessing.Queue()
+    parallel_job = ParallelJob(job_start_semaphore,
+                               target=ProcessOutputWrapper,
+                               args=(job, args, output))
+    parallel_jobs.append(parallel_job)
+    output_array.append(output)
+
   # We use a semaphore to ensure we don't run more jobs than required.
-  # After each thread finishes, it releases (increments semaphore).
-  # Acquire blocks of num jobs reached and continues when a thread finishes.
-  for next_thread in threads:
+  # After each parallel_job finishes, it releases (increments semaphore).
+  for next_parallel_job in parallel_jobs:
     job_start_semaphore.acquire(blocking=True)
-    next_thread.start()
+    next_parallel_job.start()
 
-  # Wait on the rest of the threads to finish.
-  for thread in threads:
-    while not join_semaphore.acquire(blocking=False):
-      time.sleep(5)
-      if print_status:
-        print >> sys.stderr, '.',
-
-  return [thread.GetOutput() for thread in threads]
+  ParallelJob.WaitUntilJobsComplete(parallel_jobs)
+  return [output.get() for output in output_array]