Python Multiprocessing.Pool lazy iteration

Question

I'm wondering about the way that python's Multiprocessing.Pool class works with map, imap, and map_async. My particular problem is that I want to map on an iterator that creates memory-heavy objects, and don't want all these objects to be generated into memory at the same time. I wanted to see if the various map() functions would wring my iterator dry, or intelligently call the next() function only as child processes slowly advanced, so I hacked up some tests as such:

def g():
  for el in xrange(100):
    print el
    yield el

def f(x):
  time.sleep(1)
  return x*x

if __name__ == '__main__':
  pool = Pool(processes=4)              # start 4 worker processes
  go = g()
  g2 = pool.imap(f, go)
  g2.next()

And so on with map, imap, and map_async. This is the most flagrant example however, as simply calling next() a single time on g2 prints out all my elements from my generator g(), whereas if imap were doing this 'lazily' I would expect it to only call go.next() once, and therefore print out only '1'.

Can someone clear up what is happening, and if there is some way to have the process pool 'lazily' evaluate the iterator as needed?

Thanks,

Gabe

Answer 1

Let's look at the end of the program first.

The multiprocessing module uses atexit to call multiprocessing.util._exit_function when your program ends.

If you remove g2.next(), your program ends quickly.

The _exit_function eventually calls Pool._terminate_pool. The main thread changes the state of pool._task_handler._state from RUN to TERMINATE. Meanwhile the pool._task_handler thread is looping in Pool._handle_tasks and bails out when it reaches the condition

            if thread._state:
                debug('task handler found thread._state != RUN')
                break

(See /usr/lib/python2.6/multiprocessing/pool.py)

This is what stops the task handler from fully consuming your generator, g(). If you look in Pool._handle_tasks you'll see

        for i, task in enumerate(taskseq):
            ...
            try:
                put(task)
            except IOError:
                debug('could not put task on queue')
                break

This is the code which consumes your generator. (taskseq is not exactly your generator, but as taskseq is consumed, so is your generator.)

In contrast, when you call g2.next() the main thread calls IMapIterator.next, and waits when it reaches self._cond.wait(timeout).

That the main thread is waiting instead of calling _exit_function is what allows the task handler thread to run normally, which means fully consuming the generator as it puts tasks in the workers' inqueue in the Pool._handle_tasks function.

The bottom line is that all Pool map functions consume the entire iterable that it is given. If you'd like to consume the generator in chunks, you could do this instead:

import multiprocessing as mp
import itertools
import time


def g():
    for el in xrange(50):
        print el
        yield el


def f(x):
    time.sleep(1)
    return x * x

if __name__ == '__main__':
    pool = mp.Pool(processes=4)              # start 4 worker processes
    go = g()
    result = []
    N = 11
    while True:
        g2 = pool.map(f, itertools.islice(go, N))
        if g2:
            result.extend(g2)
            time.sleep(1)
        else:
            break
    print(result)

Answer 2

What you want is implemented in the NuMap package, from the website:

NuMap is a parallel (thread- or process-based, local or remote), buffered, multi-task, itertools.imap or multiprocessing.Pool.imap function replacement. Like imap it evaluates a function on elements of a sequence or iterable, and it does so lazily. Laziness can be adjusted via the “stride” and “buffer” arguments.

Answer 3

I had this problem too and was disappointed to learn that map consumes all its elements. I coded a function which consumes the iterator lazily using the Queue data type in multiprocessing. This is similar to what @unutbu describes in a comment to his answer but as he points out, suffers from having no callback mechanism for re-loading the Queue. The Queue datatype instead exposes a timeout parameter and I've used 100 milliseconds to good effect.

from multiprocessing import Process, Queue, cpu_count
from Queue import Full as QueueFull
from Queue import Empty as QueueEmpty

def worker(recvq, sendq):
    for func, args in iter(recvq.get, None):
        result = func(*args)
        sendq.put(result)

def pool_imap_unordered(function, iterable, procs=cpu_count()):
    # Create queues for sending/receiving items from iterable.

    sendq = Queue(procs)
    recvq = Queue()

    # Start worker processes.

    for rpt in xrange(procs):
        Process(target=worker, args=(sendq, recvq)).start()

    # Iterate iterable and communicate with worker processes.

    send_len = 0
    recv_len = 0
    itr = iter(iterable)

    try:
        value = itr.next()
        while True:
            try:
                sendq.put((function, value), True, 0.1)
                send_len += 1
                value = itr.next()
            except QueueFull:
                while True:
                    try:
                        result = recvq.get(False)
                        recv_len += 1
                        yield result
                    except QueueEmpty:
                        break
    except StopIteration:
        pass

    # Collect all remaining results.

    while recv_len < send_len:
        result = recvq.get()
        recv_len += 1
        yield result

    # Terminate worker processes.

    for rpt in xrange(procs):
        sendq.put(None)

This solution has the advantage of not batching requests to Pool.map. One individual worker can not block others from making progress. YMMV. Note that you may want to use a different object to signal termination for the workers. In the example, I've used None.

Tested on "Python 2.7 (r27:82525, Jul 4 2010, 09:01:59) [MSC v.1500 32 bit (Intel)] on win32"

Answer 4

In this example (see code, please) 2 workers.

Pool work as expected: when worker is free, then to do next iteration.

This code as code in topic, except one thing: argument size = 64 k.

64 k - default socket buffer size.

import itertools
from multiprocessing import Pool
from time import sleep


def f( x ):
    print( "f()" )
    sleep( 3 )
    return x


def get_reader():
    for x in range( 10 ):
        print( "readed: ", x )
        value = " " * 1024 * 64 # 64k
        yield value


if __name__ == '__main__':

    p = Pool( processes=2 )

    data = p.imap( f, get_reader() )

    p.close()
    p.join()

Answer 5

I ran into this issue as well, and came to a different solution than the other answers here so I figured I would share it.

import collections, multiprocessing

def map_prefetch(func, data, lookahead=128, workers=16, timeout=10):
    with multiprocessing.Pool(workers) as pool:
        q = collections.deque()
        for x in data:
            q.append(pool.apply_async(func, (x,)))
            if len(q) >= lookahead:
                yield q.popleft().get(timeout=timeout)
        while len(q):
            yield q.popleft().get(timeout=timeout)

for x in map_prefetch(myfunction, huge_data_iterator):
    # do stuff with x

Basically is uses a queue to send at most lookahead pending requests to the worker pool, enforcing a limit on buffered results. The work starts asap within that limit so it can run in parallel. Also the result remains in order.