How to use IPython.parallel map() with generators

问题:

I am trying to use IPython.parallel map. The inputs to the function I wish to parallelize are generators. Because of size/memory it is not possible for me to convert the generators to lists. See code below:

from itertools import product
from IPython.parallel import Client

c = Client()
v = c[:]
c.ids

def stringcount(longstring, substrings):
    scount = [longstring.count(s) for s in substrings]
    return scount

substrings = product('abc', repeat=2)
longstring = product('abc', repeat=3)

# This is what I want to do in parallel
# I should be 'for longs in longstring' I use range() because it can get long.
for num in range(10): 
    longs = longstring.next()
    subs = substrings.next()
    print(subs, longs)
    count = stringcount(longs, subs)
    print(count)

# This does not work, and I understand why.
# I don't know how to fix it while keeping longstring and substrings as
# generators  
v.map(stringcount, longstring, substrings)

for r in v:
    print(r.get())

回答1:

You can't use View.map with a generator without walking through the entire generator first. But you can write your own custom function to submit batches of tasks from a generator and wait for them incrementally. I don't have a more interesting example, but I can illustrate with a terrible implementation of a prime search.

Start with our token 'data generator':

from math import sqrt

def generate_possible_factors(N):
    """generator for iterating through possible factors for N

    yields 2, every odd integer <= sqrt(N)
    """
    if N <= 3:
        return
    yield 2
    f = 3
    last = int(sqrt(N))
    while f <= last:
        yield f
        f += 2

This just generates a sequence of integers to use when testing if a number is prime.

Now our trivial function that we will use as a task with IPython.parallel

def is_factor(f, N):
    """is f a factor of N?"""
    return (N % f) == 0

and a complete implementation of prime check using the generator and our factor function:

def dumb_prime(N):
    """dumb implementation of is N prime?"""
    for f in generate_possible_factors(N):
        if is_factor(f, N):
            return False
    return True

A parallel version that only submits a limited number of tasks at a time:

def parallel_dumb_prime(N, v, max_outstanding=10, dt=0.1):
    """dumb_prime where each factor is checked remotely

    Up to `max_outstanding` factors will be checked in parallel.

    Submission will halt as soon as we know that N is not prime.
    """
    tasks = set()
    # factors is a generator
    factors = generate_possible_factors(N)
    while True:
        try:
            # submit a batch of tasks, with a maximum of `max_outstanding`
            for i in range(max_outstanding-len(tasks)):
                f = factors.next()
                tasks.add(v.apply_async(is_factor, f, N))
        except StopIteration:
            # no more factors to test, stop submitting
            break
        # get the tasks that are done
        ready = set(task for task in tasks if task.ready())
        while not ready:
            # wait a little bit for some tasks to finish
            v.wait(tasks, timeout=dt)
            ready = set(task for task in tasks if task.ready())

        for t in ready:
            # get the result - if True, N is not prime, we are done
            if t.get():
                return False
        # update tasks to only those that are still pending,
        # and submit the next batch
        tasks.difference_update(ready)
    # check the last few outstanding tasks
    for task in tasks:
        if t.get():
            return False
    # checked all candidates, none are factors, so N is prime
    return True

This submits a limited number of tasks at a time, and as soon as we know that N is not prime, we stop consuming the generator.

To use this function:

from IPython import parallel

rc = parallel.Client()
view = rc.load_balanced_view()

for N in range(900,1000):
    if parallel_dumb_prime(N, view, 10):
        print N

A more complete illustration in a notebook.

回答2:

I took a slightly different approach to your problem that may be useful to others. Below, I attempted to mimic the behavior of the multiprocessing.pool.Pool.imap method by wrapping IPython.parallel.map. This required me to re-write your functions slightly.

import IPython
from itertools import product


def stringcount((longstring, substrings)):
    scount = [longstring.count(s) for s in substrings]
    return (longstring, substrings, scount)

def gen_pairs(long_string, sub_strings):
    for l in long_string:
        s = sub_strings.next()
        yield (l, s)

def imap(function, generator, view, preprocessor=iter, chunksize=256):
    num_cores = len(view.client.ids)
    queue = []
    for i, n in enumerate(preprocessor(generator)):
        queue.append(n)
        if not i % (chunksize * num_cores):
            for result in view.map(function, queue):
                yield result
            queue = []
    for result in view.map(function, queue):
        yield result


client = IPython.parallel.Client()
lbview = client.load_balanced_view()

longstring = product('abc', repeat=3)
substrings = product('abc', repeat=2)

for result in imap(stringcount, gen_pairs(longstring, substrings), lbview):
    print result

The output I'm seeing is on this Notebook: http://nbviewer.ipython.org/gist/driscoll/b8de4bf980de1ad890de