Logo Questions Linux Laravel Mysql Ubuntu Git Menu
 

how to keep track of asynchronous results returned from a multiprocessing pool

I am trying to add multiprocessing to some code which features functions that I can not modify. I want to submit these functions as jobs to a multiprocessing pool asynchronously. I am doing something much like the code shown here. However, I am not sure how to keep track of results. How can I know to which applied function a returned result corresponds?

The important points to emphasise are that I can not modify the existing functions (other things rely on them remaining as they are) and that results can be returned in an order different to the order in which the function jobs are applied to the pool.

Thanks for any thoughts on this!

EDIT: Some attempt code is below:

import multiprocessing
from multiprocessing import Pool
import os
import signal
import time
import inspect

def multiply(multiplicand1=0, multiplicand2=0):
    return multiplicand1*multiplicand2

def workFunctionTest(**kwargs):
    time.sleep(3)
    return kwargs

def printHR(object):
    """
    This function prints a specified object in a human readable way.
    """
    # dictionary
    if isinstance(object, dict):
        for key, value in sorted(object.items()):
            print u'{a1}: {a2}'.format(a1=key, a2=value)
    # list or tuple
    elif isinstance(object, list) or isinstance(object, tuple):
        for element in object:
            print element
    # other
    else:
        print object

class Job(object):
    def __init__(
        self,
        workFunction=workFunctionTest,
        workFunctionKeywordArguments={'testString': "hello world"},
        workFunctionTimeout=1,
        naturalLanguageString=None,
        classInstance=None,
        resultGetter=None,
        result=None
        ):
        self.workFunction=workFunction
        self.workFunctionKeywordArguments=workFunctionKeywordArguments
        self.workFunctionTimeout=workFunctionTimeout
        self.naturalLanguageString=naturalLanguageString
        self.classInstance=self.__class__.__name__
        self.resultGetter=resultGetter
        self.result=result
    def description(self):
        descriptionString=""
        for key, value in sorted(vars(self).items()):
            descriptionString+=str("{a1}:{a2} ".format(a1=key, a2=value))
        return descriptionString
    def printout(self):
        """
        This method prints a dictionary of all data attributes.
        """
        printHR(vars(self))

class JobGroup(object):
    """
    This class acts as a container for jobs. The data attribute jobs is a list of job objects.
    """
    def __init__(
        self,
        jobs=None,
        naturalLanguageString="null",
        classInstance=None,
        result=None
        ):
        self.jobs=jobs
        self.naturalLanguageString=naturalLanguageString
        self.classInstance=self.__class__.__name__
        self.result=result
    def description(self):
        descriptionString=""
        for key, value in sorted(vars(self).items()):
            descriptionString+=str("{a1}:{a2} ".format(a1=key, a2=value))
        return descriptionString
    def printout(self):
        """
        This method prints a dictionary of all data attributes.
        """
        printHR(vars(self))

def initialise_processes():
    signal.signal(signal.SIGINT, signal.SIG_IGN)

def execute(
        jobObject=None,
        numberOfProcesses=multiprocessing.cpu_count()
        ):
        # Determine the current function name.
    functionName=str(inspect.stack()[0][3])
    def collateResults(result):
        """
        This is a process pool callback function which collates a list of results returned.
        """
        # Determine the caller function name.
        functionName=str(inspect.stack()[1][3])
        print("{a1}: result: {a2}".format(a1=functionName, a2=result))
        results.append(result)
    def getResults(job):
        # Determine the current function name.
        functionName=str(inspect.stack()[0][3])
        while True:
            try:
                result=job.resultGetter.get(job.workFunctionTimeout)
                break
            except multiprocessing.TimeoutError:
                print("{a1}: subprocess timeout for job".format(a1=functionName, a2=job.description()))
        #job.result=result
        return result
    # Create a process pool.
    pool1 = multiprocessing.Pool(numberOfProcesses, initialise_processes)
    print("{a1}: pool {a2} of {a3} processes created".format(a1=functionName, a2=str(pool1), a3=str(numberOfProcesses)))
    # Unpack the input job object and submit it to the process pool.
    print("{a1}: unpacking and applying job object {a2} to pool...".format(a1=functionName, a2=jobObject))
    if isinstance(jobObject, Job):
        # If the input job object is a job, apply it to the pool with its associated timeout specification.
        # Return a list of results.
        job=jobObject
        print("{a1}: job submitted to pool: {a2}".format(a1=functionName, a2=job.description()))
        # Apply the job to the pool, saving the object pool.ApplyResult to the job object.
        job.resultGetter=pool1.apply_async(
                func=job.workFunction,
                kwds=job.workFunctionKeywordArguments
        )
        # Get results.
        # Acquire the job result with respect to the specified job timeout and apply this result to the job data attribute result.
        print("{a1}: getting results for job...".format(a1=functionName))
        job.result=getResults(job)
        print("{a1}: job completed: {a2}".format(a1=functionName, a2=job.description()))
        print("{a1}: job result: {a2}".format(a1=functionName, a2=job.result))
        # Return the job result from execute.
        return job.result
        pool1.terminate()
        pool1.join()
    elif isinstance(jobObject, JobGroup):
        # If the input job object is a job group, cycle through each job and apply it to the pool with its associated timeout specification.
        for job in jobObject.jobs:
            print("{a1}: job submitted to pool: {a2}".format(a1=functionName, a2=job.description()))
            # Apply the job to the pool, saving the object pool.ApplyResult to the job object.
            job.resultGetter=pool1.apply_async(
                    func=job.workFunction,
                    kwds=job.workFunctionKeywordArguments
            )
        # Get results.
        # Cycle through each job and and append the result for the job to a list of results.
        results=[]
        for job in jobObject.jobs:
            # Acquire the job result with respect to the specified job timeout and apply this result to the job data attribute result.
            print("{a1}: getting results for job...".format(a1=functionName))
            job.result=getResults(job)
            print("{a1}: job completed: {a2}".format(a1=functionName, a2=job.description()))
            #print("{a1}: job result: {a2}".format(a1=functionName, a2=job.result))
            # Collate the results.
            results.append(job.result)
        # Apply the list of results to the job group data attribute results.
        jobObject.results=results
        print("{a1}: job group results: {a2}".format(a1=functionName, a2=jobObject.results))
        # Return the job result list from execute.
        return jobObject.results
        pool1.terminate()
        pool1.join()
    else:
        # invalid input object
        print("{a1}: invalid job object {a2}".format(a1=functionName, a2=jobObject))

def main():
    print('-'*80)
    print("MULTIPROCESSING SYSTEM DEMONSTRATION\n")

    # Create a job.
    print("# creating a job...\n")
    job1=Job(
            workFunction=workFunctionTest,
            workFunctionKeywordArguments={'testString': "hello world"},
            workFunctionTimeout=4
    )
    print("- printout of new job object:")
    job1.printout()
    print("\n- printout of new job object in logging format:")
    print job1.description()

    # Create another job.
    print("\n# creating another job...\n")
    job2=Job(
            workFunction=multiply,
            workFunctionKeywordArguments={'multiplicand1': 2, 'multiplicand2': 3},
            workFunctionTimeout=6
    )
    print("- printout of new job object:")
    job2.printout()
    print("\n- printout of new job object in logging format:")
    print job2.description()

    # Create a JobGroup object.
    print("\n# creating a job group (of jobs 1 and 2)...\n")
    jobGroup1=JobGroup(
            jobs=[job1, job2],
    )
    print("- printout of new job group object:")
    jobGroup1.printout()
    print("\n- printout of new job group object in logging format:")
    print jobGroup1.description()

    # Submit the job group.
    print("\nready to submit job group")
    response=raw_input("\nPress Enter to continue...\n")
    execute(jobGroup1)

    response=raw_input("\nNote the results printed above. Press Enter to continue the demonstration.\n")

    # Demonstrate timeout.
    print("\n # creating a new job in order to demonstrate timeout functionality...\n")
    job3=Job(
            workFunction=workFunctionTest,
            workFunctionKeywordArguments={'testString': "hello world"},
            workFunctionTimeout=1
    )
    print("- printout of new job object:")
    job3.printout()
    print("\n- printout of new job object in logging format:")
    print job3.description()
    print("\nNote the timeout specification of only 1 second.")

    # Submit the job.
    print("\nready to submit job")
    response=raw_input("\nPress Enter to continue...\n")
    execute(job3)

    response=raw_input("\nNote the recognition of timeouts printed above. This concludes the demonstration.")
    print('-'*80)

if __name__ == '__main__':
    main()

EDIT: This question has been placed [on hold] for the following stated reason:

"Questions asking for code must demonstrate a minimal understanding of the problem being solved. Include attempted solutions, why they didn't work, and the expected results. See also: Stack Overflow question checklist"

This question is not requesting code; it is requesting thoughts, general guidance. A minimal understanding of the problem under consideration is demonstrated (note the correct use of the terms "multiprocessing", "pool" and "asynchronously" and note the reference to prior code). Regarding attempted solutions, I acknowledge that attempted efforts at solutions would have been beneficial. I have added such code now. I hope that I have addressed the concerns raised that lead to the [on hold] status.

like image 210
d3pd Avatar asked Dec 13 '13 23:12

d3pd


People also ask

What does Apply_async return?

The apply_async() function returns an AsyncResult, whereas the apply() function returns the result of the target function. The apply_async() function can execute callback functions when the task is complete, whereas the apply() function cannot execute callback functions.

How do processes pools work in multiprocessing?

Pool is generally used for heterogeneous tasks, whereas multiprocessing. Process is generally used for homogeneous tasks. The Pool is designed to execute heterogeneous tasks, that is tasks that do not resemble each other. For example, each task submitted to the process pool may be a different target function.

When would you use a multiprocessing pool?

Use the multiprocessing pool if your tasks are independent. This means that each task is not dependent on other tasks that could execute at the same time. It also may mean tasks that are not dependent on any data other than data provided via function arguments to the task.


1 Answers

Without seeing actual code, I can only answer in generalities. But there are two general solutions.

First, instead of using a callback and ignoring the AsyncResults, store them in some kind of collection. Then you can just use that collection. For example, if you want to be able to look up the results for a function using that function as a key, just create a dict keyed with the functions:

def in_parallel(funcs):
    results = {}
    pool = mp.Pool()
    for func in funcs:
        results[func] = pool.apply_async(func)
    pool.close()
    pool.join()
    return {func: result.get() for func, result in results.items()}

Alternatively, you can change the callback function to store the results in your collection by key. For example:

def in_parallel(funcs):
    results = {}
    pool = mp.Pool()
    for func in funcs:
        def callback(result, func=func):
            results[func] = result
        pool.apply_async(func, callback=callback)
    pool.close()
    pool.join()
    return results

I'm using the function itself as a key. But you want to use the index instead, that's just as easy. Any value you have, you can use as a key.


Meanwhile, the example you linked is really just calling the same function on a bunch of arguments, waiting for all of them to finish, and leaving the results in some iterable in arbitrary order. That's exactly what imap_unordered does, but a lot more simply. You could replace the whole complicated thing from the linked code with this:

pool = mp.Pool()
results = list(pool.imap_unordered(foo_pool, range(10)))
pool.close()
pool.join()

And then, if you want the results in their original order instead of in arbitrary order, you can just switch to imap or map instead. So:

pool = mp.Pool()
results = pool.map(foo_pool, range(10))
pool.close()
pool.join()

If you need something similar but too complicated to fit into the map paradigm, concurrent.futures will probably make your life easier than multiprocessing. If you're on Python 2.x, you will have to install the backport. But then you can do things that are much harder to do with AsyncResults or callbacks (or map), like composing a whole bunch of futures into one big future. See the examples in the linked docs.


One last note:

The important points to emphasise are that I can not modify the existing functions…

If you can't modify a function, you can always wrap it. For example, let's say I have a function that returns the square of a number, but I'm trying to build a dict mapping numbers to their squares asynchronously, so I need to have the original number as part of the result as well. That's easy:

def number_and_square(x):
    return x, square(x)

And now, I can just apply_async(number_and_square) instead of just square, and get the results I want.

I didn't do that in the examples above because in the first case I stored the key into the collection from the calling side, and in the second place I bound it into the callback function. But binding it into a wrapper around the function is just as easy as either of these, and can be appropriate when neither of these is.

like image 51
abarnert Avatar answered Nov 15 '22 19:11

abarnert