Assertion Error when using multiprocessing in Python 3.4

Question

I'm pretty new to Python and completely new to parallel processing.

I've been writing code to analyze punctate image data (think PALM lite) and trying to speed up my analysis code using the multiprocessing module.

For small data sets I see a pretty decent speed-up up to four cores. For large datasets I start getting an AssertionError. I tried to make a boiled down example which produces the same error, see below:

import numpy as np
import multiprocessing as mp
import os

class TestClass(object):
    def __init__(self, data):
        super().__init__()
        self.data = data

    def top_level_function(self, nproc = 1):

        if nproc > os.cpu_count():
            nproc = os.cpu_count()

        if nproc == 1:
            sums = [self._sub_function() for i in range(10)]
        elif 1 < nproc:
            print('multiprocessing engaged with {} cores'.format(nproc))
            with mp.Pool(nproc) as p:
                sums = [p.apply_async(self._sub_function) for i in range(10)]
                sums = [pp.get() for pp in sums]

        self.sums = sums

        return sums

    def _sub_function(self):
        return self.data.sum(0)


if __name__ == "__main__":
    t = TestClass(np.zeros((126,512,512)))
    ans = t.top_level_function()
    print(len(ans))
    ans = t.top_level_function(4)
    print(len(ans))

    t = TestClass(np.zeros((126,2048,2048)))
    ans = t.top_level_function()
    print(len(ans))
    ans = t.top_level_function(4)
    print(len(ans))

which outputs:

10
multiprocessing engaged with 4 cores
10
10
multiprocessing engaged with 4 cores
Process SpawnPoolWorker-6:
Traceback (most recent call last):
  File "C:\Anaconda3\lib\multiprocessing\process.py", line 254, in _bootstrap
    self.run()
  File "C:\Anaconda3\lib\multiprocessing\process.py", line 93, in run
    self._target(*self._args, **self._kwargs)
  File "C:\Anaconda3\lib\multiprocessing\pool.py", line 108, in worker
    task = get()
  File "C:\Anaconda3\lib\multiprocessing\queues.py", line 355, in get
    res = self._reader.recv_bytes()
  File "C:\Anaconda3\lib\multiprocessing\connection.py", line 216, in recv_bytes
    buf = self._recv_bytes(maxlength)
  File "C:\Anaconda3\lib\multiprocessing\connection.py", line 318, in _recv_bytes
    return self._get_more_data(ov, maxsize)
  File "C:\Anaconda3\lib\multiprocessing\connection.py", line 337, in _get_more_data
    assert left > 0
AssertionError
Process SpawnPoolWorker-8:
Traceback (most recent call last):
  File "C:\Anaconda3\lib\multiprocessing\process.py", line 254, in _bootstrap
    self.run()
  File "C:\Anaconda3\lib\multiprocessing\process.py", line 93, in run
    self._target(*self._args, **self._kwargs)
  File "C:\Anaconda3\lib\multiprocessing\pool.py", line 108, in worker
    task = get()
  File "C:\Anaconda3\lib\multiprocessing\queues.py", line 355, in get
    res = self._reader.recv_bytes()
  File "C:\Anaconda3\lib\multiprocessing\connection.py", line 216, in recv_bytes
    buf = self._recv_bytes(maxlength)
  File "C:\Anaconda3\lib\multiprocessing\connection.py", line 318, in _recv_bytes
    return self._get_more_data(ov, maxsize)
  File "C:\Anaconda3\lib\multiprocessing\connection.py", line 337, in _get_more_data
    assert left > 0
AssertionError
Traceback (most recent call last):
  File "test.py", line 41, in <module>
    ans = t.top_level_function(4)
  File "test.py", line 21, in top_level_function
    sums = [pp.get() for pp in sums]
  File "test.py", line 21, in <listcomp>
    sums = [pp.get() for pp in sums]
  File "C:\Anaconda3\lib\multiprocessing\pool.py", line 599, in get
    raise self._value
  File "C:\Anaconda3\lib\multiprocessing\pool.py", line 383, in _handle_tasks
    put(task)
  File "C:\Anaconda3\lib\multiprocessing\connection.py", line 206, in send
    self._send_bytes(ForkingPickler.dumps(obj))
  File "C:\Anaconda3\lib\multiprocessing\connection.py", line 280, in _send_bytes
    ov, err = _winapi.WriteFile(self._handle, buf, overlapped=True)
OSError: [WinError 87] The parameter is incorrect

So the first example runs fine, but the later example (larger set of data) crashes.

I'm pretty lost about where this error is coming from and how to fix it. Any help would be greatly appreciated.

Answer 1

When you do

sums = [p.apply_async(self._sub_function) for i in range(10)]

what happens is that self._sub_function will be pickled 10 times and sent to a worker process to be processed. To pickle an instance method, the whole instance ( including the data attribute) has to be pickled. A quick check shows that np.zeros((126,2048,2048)) when pickled requires 4227858596 bytes, and you're sending 10 times that, to 10 different processes.

You're getting a error during _send_bytes , which means the transfer to the worker process was interrupted, my guess would be because you're hitting your memory limit.

You should probably rethink your design, multiprocessing usually works best if each worker can work on part of the problem without needing access to the whole data.