Multiprocessing.Pool makes Numpy matrix multiplication slower

Question

So, I am playing around with multiprocessing.Pool and Numpy, but it seems I missed some important point. Why is the pool version much slower? I looked at htop and I can see several processes be created, but they all share one of the CPUs adding up to ~100%.

$ cat test_multi.py 
import numpy as np
from timeit import timeit
from multiprocessing import Pool


def mmul(matrix):
    for i in range(100):
        matrix = matrix * matrix
    return matrix

if __name__ == '__main__':
    matrices = []
    for i in range(4):
        matrices.append(np.random.random_integers(100, size=(1000, 1000)))

    pool = Pool(8)
    print timeit(lambda: map(mmul, matrices), number=20)
    print timeit(lambda: pool.map(mmul, matrices), number=20)

$ python test_multi.py 
16.0265390873
19.097837925

[update]

• changed to timeit for benchmarking processes
• init Pool with a number of my cores
• changed computation so that there is more computation and less memory transfer (I hope)

Still no change. pool version is still slower and I can see in htop that only one core is used also several processes are spawned.

[update2]

At the moment I am reading about @Jan-Philip Gehrcke's suggestion to use multiprocessing.Process() and Queue. But in the meantime I would like to know:

1. Why does my example work for tiago? What could be the reason it is not working on my machine1?
2. Is in my example code any copying between the processes? I intended my code to give each thread one matrix of the matrices list.
3. Is my code a bad example, because I use Numpy?

I learned that often one gets better answer, when the others know my end goal so: I have a lot of files, which are atm loaded and processed in a serial fashion. The processing is CPU intense, so I assume much could be gained by parallelization. My aim is it to call the python function that analyses a file in parallel. Furthermore this function is just an interface to C code, I assume, that makes a difference.

1 Ubuntu 12.04, Python 2.7.3, i7 860 @ 2.80 - Please leave a comment if you need more info.

[update3]

Here are the results from Stefano's example code. For some reason there is no speed up. :/

testing with 16 matrices
base  4.27
   1  5.07
   2  4.76
   4  4.71
   8  4.78
  16  4.79
testing with 32 matrices
base  8.82
   1 10.39
   2 10.58
   4 10.73
   8  9.46
  16  9.54
testing with 64 matrices
base 17.38
   1 19.34
   2 19.62
   4 19.59
   8 19.39
  16 19.34

[update 4] answer to Jan-Philip Gehrcke's comment

Sorry that I haven't made myself clearer. As I wrote in Update 2 my main goal is it to parallelize many serial calls of a 3rd party Python library function. This function is an interface to some C code. I was recommended to use Pool, but this didn't work, so I tried something simpler, the shown above example with numpy. But also there I could not achieve a performance improvement, even though it looks for me 'emberassing parallelizable`. So I assume I must have missed something important. This information is what I am looking for with this question and bounty.

[update 5]

Thanks for all your tremendous input. But reading through your answers only creates more questions for me. For that reason I will read about the basics and create new SO questions when I have a clearer understanding of what I don't know.

Answer 1

Regarding the fact that all of your processes are running on the same CPU, see my answer here.

During import, numpy changes the CPU affinity of the parent process, such that when you later use Pool all of the worker processes that it spawns will end up vying for for the same core, rather than using all of the cores available on your machine.

You can call taskset after you import numpy to reset the CPU affinity so that all cores are used:

import numpy as np
import os
from timeit import timeit
from multiprocessing import Pool


def mmul(matrix):
    for i in range(100):
        matrix = matrix * matrix
    return matrix

if __name__ == '__main__':

    matrices = []
    for i in range(4):
        matrices.append(np.random.random_integers(100, size=(1000, 1000)))

    print timeit(lambda: map(mmul, matrices), number=20)

    # after importing numpy, reset the CPU affinity of the parent process so
    # that it will use all cores
    os.system("taskset -p 0xff %d" % os.getpid())

    pool = Pool(8)
    print timeit(lambda: pool.map(mmul, matrices), number=20)

Output:

    $ python tmp.py                                     
    12.4765810966
    pid 29150's current affinity mask: 1
    pid 29150's new affinity mask: ff
    13.4136221409

If you watch CPU useage using top while you run this script, you should see it using all of your cores when it executes the 'parallel' part. As others have pointed out, in your original example the overhead involved in pickling data, process creation etc. probably outweigh any possible benefit from parallelisation.

Edit: I suspect that part of the reason why the single process seems to be consistently faster is that numpy may have some tricks for speeding up that element-wise matrix multiplication that it cannot use when the jobs are spread across multiple cores.

For example, if I just use ordinary Python lists to compute the Fibonacci sequence, I can get a huge speedup from parallelisation. Likewise, if I do element-wise multiplication in a way that takes no advantage of vectorization, I get a similar speedup for the parallel version:

import numpy as np
import os
from timeit import timeit
from multiprocessing import Pool

def fib(dummy):
    n = [1,1]
    for ii in xrange(100000):
        n.append(n[-1]+n[-2])

def silly_mult(matrix):
    for row in matrix:
        for val in row:
            val * val

if __name__ == '__main__':

    dt = timeit(lambda: map(fib, xrange(10)), number=10)
    print "Fibonacci, non-parallel: %.3f" %dt

    matrices = [np.random.randn(1000,1000) for ii in xrange(10)]
    dt = timeit(lambda: map(silly_mult, matrices), number=10)
    print "Silly matrix multiplication, non-parallel: %.3f" %dt

    # after importing numpy, reset the CPU affinity of the parent process so
    # that it will use all CPUS
    os.system("taskset -p 0xff %d" % os.getpid())

    pool = Pool(8)

    dt = timeit(lambda: pool.map(fib,xrange(10)), number=10)
    print "Fibonacci, parallel: %.3f" %dt

    dt = timeit(lambda: pool.map(silly_mult, matrices), number=10)
    print "Silly matrix multiplication, parallel: %.3f" %dt

Output:

$ python tmp.py
Fibonacci, non-parallel: 32.449
Silly matrix multiplication, non-parallel: 40.084
pid 29528's current affinity mask: 1
pid 29528's new affinity mask: ff
Fibonacci, parallel: 9.462
Silly matrix multiplication, parallel: 12.163

Answer 2

The unpredictable competition between communication overhead and computation speedup is definitely the issue here. What you are observing is perfectly fine. Whether you get a net speed-up depends on many factors and is something that has to be quantified properly (as you did).

So why is multiprocessing so "unexpectedly slow" in your case? multiprocessing's map and map_async functions actually pickle Python objects back and forth through pipes that connect the parent with the child processes. This may take a considerable amount of time. During that time, the child processes have almost nothing to do, which is what to see in htop. Between different systems, there might be a considerable pipe transport performance difference, which is also why for some people your pool code is faster than your single CPU code, although for you it is not (other factors might come into play here, this is just an example in order to explain the effect).

What can you do to make it faster?

1. Don't pickle the input on POSIX-compliant systems.

   If you are on Unix, you can get around the parent->child communication overhead via taking advantage of POSIX' process fork behavior (copy memory on write):

   Create your job input (e.g. a list of large matrices) to work on in the parent process in a globally accessible variable. Then create worker processes by calling multiprocessing.Process() yourself. In the children, grab the job input from the global variable. Simply expressed, this makes the child access the memory of the parent without any communication overhead (*, explanation below). Send the result back to the parent, through e.g. a multiprocessing.Queue. This will save a lot of communication overhead, especially if the output is small compared to the input. This method won't work on e.g. Windows, because multiprocessing.Process() there creates an entirely new Python process that does not inherit the state of the parent.

2. Make use of numpy multithreading. Depending on your actual calculation task, it might happen that involving multiprocessing won't help at all. If you compile numpy yourself and enable OpenMP directives, then operations on larges matrices might become very efficiently multithreaded (and distributed over many CPU cores; the GIL is no limiting factor here) by themselves. Basically, this is the most efficient usage of multiple CPU cores you can get in the context of numpy/scipy.

*The child cannot directly access the parent's memory in general. However, after fork(), parent and child are in an equivalent state. It would be stupid to copy the entire memory of the parent to another place in the RAM. That's why the copy-on-write principle jumps in. As long as the child does not change its memory state, it actually accesses the parent's memory. Only upon modification, the corresponding bits and pieces are copied into the memory space of the child.

Major edit:

Let me add a piece of code that crunches a large amount of input data with multiple worker processes and follows the advice "1. Don't pickle the input on POSIX-compliant systems.". Furthermore, the amount of information transferred back to the worker manager (the parent process) is quite low. The heavy computation part of this example is a single value decomposition. It can make heavy use of OpenMP. I have executed the example multiple times:

• Once with 1, 2, or 4 worker processes and OMP_NUM_THREADS=1, so each worker process creates a maximum load of 100 %. There, the mentioned number-of-workers-compute-time scaling behavior is almost linear and the net speedup factor up corresponds to the number of workers involved.
• Once with 1, 2, or 4 worker processes and OMP_NUM_THREADS=4, so that each process creates a maximum load of 400 % (via spawning 4 OpenMP threads). My machine has 16 real cores, so 4 processes with max 400 % load each will almost get the maximum performance out of the machine. The scaling is not perfectly linear anymore and the speedup factor is not the number of workers involved, but the absolute calculation time becomes significantly reduced compared to OMP_NUM_THREADS=1 and time still decreases significantly with the number of worker processes.
• Once with larger input data, 4 cores, and OMP_NUM_THREADS=4. It results in an average system load of 1253 %.
• Once with same setup as last, but OMP_NUM_THREADS=5. It results in an average system load of 1598 %, which suggests that we got everything from that 16 core machine. However, the actual computation wall time does not improve compared to the latter case.

The code:

import os
import time
import math
import numpy as np
from numpy.linalg import svd as svd
import multiprocessing


# If numpy is compiled for OpenMP, then make sure to control
# the number of OpenMP threads via the OMP_NUM_THREADS environment
# variable before running this benchmark.


MATRIX_SIZE = 1000
MATRIX_COUNT = 16


def rnd_matrix():
    offset = np.random.randint(1,10)
    stretch = 2*np.random.rand()+0.1
    return offset + stretch * np.random.rand(MATRIX_SIZE, MATRIX_SIZE)


print "Creating input matrices in parent process."
# Create input in memory. Children access this input.
INPUT = [rnd_matrix() for _ in xrange(MATRIX_COUNT)]


def worker_function(result_queue, worker_index, chunk_boundary):
    """Work on a certain chunk of the globally defined `INPUT` list.
    """
    result_chunk = []
    for m in INPUT[chunk_boundary[0]:chunk_boundary[1]]:
        # Perform single value decomposition (CPU intense).
        u, s, v = svd(m)
        # Build single numeric value as output.
        output =  int(np.sum(s))
        result_chunk.append(output)
    result_queue.put((worker_index, result_chunk))


def work(n_workers=1):
    def calc_chunksize(l, n):
        """Rudimentary function to calculate the size of chunks for equal 
        distribution of a list `l` among `n` workers.
        """
        return int(math.ceil(len(l)/float(n)))

    # Build boundaries (indices for slicing) for chunks of `INPUT` list.
    chunk_size = calc_chunksize(INPUT, n_workers)
    chunk_boundaries = [
        (i, i+chunk_size) for i in xrange(0, len(INPUT), chunk_size)]

    # When n_workers and input list size are of same order of magnitude,
    # the above method might have created less chunks than workers available. 
    if n_workers != len(chunk_boundaries):
        return None

    result_queue = multiprocessing.Queue()
    # Prepare child processes.
    children = []
    for worker_index in xrange(n_workers):
        children.append(
            multiprocessing.Process(
                target=worker_function,
                args=(
                    result_queue,
                    worker_index,
                    chunk_boundaries[worker_index],
                    )
                )
            )

    # Run child processes.
    for c in children:
        c.start()

    # Create result list of length of `INPUT`. Assign results upon arrival.
    results = [None] * len(INPUT)

    # Wait for all results to arrive.
    for _ in xrange(n_workers):
        worker_index, result_chunk = result_queue.get(block=True)
        chunk_boundary = chunk_boundaries[worker_index]
        # Store the chunk of results just received to the overall result list.
        results[chunk_boundary[0]:chunk_boundary[1]] = result_chunk

    # Join child processes (clean up zombies).
    for c in children:
        c.join()
    return results


def main():
    durations = []
    n_children = [1, 2, 4]
    for n in n_children:
        print "Crunching input with %s child(ren)." % n
        t0 = time.time()
        result = work(n)
        if result is None:
            continue
        duration = time.time() - t0
        print "Result computed by %s child process(es): %s" % (n, result)
        print "Duration: %.2f s" % duration
        durations.append(duration)
    normalized_durations = [durations[0]/d for d in durations]
    for n, normdur in zip(n_children, normalized_durations):
        print "%s-children speedup: %.2f" % (n, normdur)


if __name__ == '__main__':
    main()

The output:

$ export OMP_NUM_THREADS=1
$ /usr/bin/time python test2.py 
Creating input matrices in parent process.
Crunching input with 1 child(ren).
Result computed by 1 child process(es): [5587, 8576, 11566, 12315, 7453, 23245, 6136, 12387, 20634, 10661, 15091, 14090, 11997, 20597, 21991, 7972]
Duration: 16.66 s
Crunching input with 2 child(ren).
Result computed by 2 child process(es): [5587, 8576, 11566, 12315, 7453, 23245, 6136, 12387, 20634, 10661, 15091, 14090, 11997, 20597, 21991, 7972]
Duration: 8.27 s
Crunching input with 4 child(ren).
Result computed by 4 child process(es): [5587, 8576, 11566, 12315, 7453, 23245, 6136, 12387, 20634, 10661, 15091, 14090, 11997, 20597, 21991, 7972]
Duration: 4.37 s
1-children speedup: 1.00
2-children speedup: 2.02
4-children speedup: 3.81
48.75user 1.75system 0:30.00elapsed 168%CPU (0avgtext+0avgdata 1007936maxresident)k
0inputs+8outputs (1major+809308minor)pagefaults 0swaps

$ export OMP_NUM_THREADS=4
$ /usr/bin/time python test2.py 
Creating input matrices in parent process.
Crunching input with 1 child(ren).
Result computed by 1 child process(es): [22735, 5932, 15692, 14129, 6953, 12383, 17178, 14896, 16270, 5591, 4174, 5843, 11740, 17430, 15861, 12137]
Duration: 8.62 s
Crunching input with 2 child(ren).
Result computed by 2 child process(es): [22735, 5932, 15692, 14129, 6953, 12383, 17178, 14896, 16270, 5591, 4174, 5843, 11740, 17430, 15861, 12137]
Duration: 4.92 s
Crunching input with 4 child(ren).
Result computed by 4 child process(es): [22735, 5932, 15692, 14129, 6953, 12383, 17178, 14896, 16270, 5591, 4174, 5843, 11740, 17430, 15861, 12137]
Duration: 2.95 s
1-children speedup: 1.00
2-children speedup: 1.75
4-children speedup: 2.92
106.72user 3.07system 0:17.19elapsed 638%CPU (0avgtext+0avgdata 1022240maxresident)k
0inputs+8outputs (1major+841915minor)pagefaults 0swaps

$ /usr/bin/time python test2.py 
Creating input matrices in parent process.
Crunching input with 4 child(ren).
Result computed by 4 child process(es): [21762, 26806, 10148, 22947, 20900, 8161, 20168, 17439, 23497, 26360, 6789, 11216, 12769, 23022, 26221, 20480, 19140, 13757, 23692, 19541, 24644, 21251, 21000, 21687, 32187, 5639, 23314, 14678, 18289, 12493, 29766, 14987, 12580, 17988, 20853, 4572, 16538, 13284, 18612, 28617, 19017, 23145, 11183, 21018, 10922, 11709, 27895, 8981]
Duration: 12.69 s
4-children speedup: 1.00
174.03user 4.40system 0:14.23elapsed 1253%CPU (0avgtext+0avgdata 2887456maxresident)k
0inputs+8outputs (1major+1211632minor)pagefaults 0swaps

$ export OMP_NUM_THREADS=5
$ /usr/bin/time python test2.py 
Creating input matrices in parent process.
Crunching input with 4 child(ren).
Result computed by 4 child process(es): [19528, 17575, 21792, 24303, 6352, 22422, 25338, 18183, 15895, 19644, 20161, 22556, 24657, 30571, 13940, 18891, 10866, 21363, 20585, 15289, 6732, 10851, 11492, 29146, 12611, 15022, 18967, 25171, 10759, 27283, 30413, 14519, 25456, 18934, 28445, 12768, 28152, 24055, 9285, 26834, 27731, 33398, 10172, 22364, 12117, 14967, 18498, 8111]
Duration: 13.08 s
4-children speedup: 1.00
230.16user 5.98system 0:14.77elapsed 1598%CPU (0avgtext+0avgdata 2898640maxresident)k
0inputs+8outputs (1major+1219611minor)pagefaults 0swaps

Answer 3

Your code is correct. I just ran it my system (with 2 cores, hyperthreading) and obtained the following results:

$ python test_multi.py 
30.8623809814
19.3914041519

I looked at the processes and, as expected, the parallel part showing several processes working at near 100%. This must be something in your system or python installation.

Answer 4

By default, Pool only uses n processes, where n is the number of CPUs on your machine. You need to specify how many processes you want it to use, like Pool(5).

See here for more info

Answer 5

Measuring arithmetic throughput is a very difficult task: basically your test case is too simple, and I see many problems.

First you are testing integer arithmetic: is there a special reason? With floating point you get results that are comparable across many different architectures.

Second matrix = matrix*matrix overwrites the input parameter (matrices are passed by ref and not by value), and each sample has to work on different data...

Last tests should be conducted over a wider range of problem size and number of workers, in order to grasp general trends.

So here is my modified test script

import numpy as np
from timeit import timeit
from multiprocessing import Pool

def mmul(matrix):
    mymatrix = matrix.copy()
    for i in range(100):
        mymatrix *= mymatrix
    return mymatrix

if __name__ == '__main__':

    for n in (16, 32, 64):
        matrices = []
        for i in range(n):
            matrices.append(np.random.random_sample(size=(1000, 1000)))

        stmt = 'from __main__ import mmul, matrices'
        print 'testing with', n, 'matrices'
        print 'base',
        print '%5.2f' % timeit('r = map(mmul, matrices)', setup=stmt, number=1)

        stmt = 'from __main__ import mmul, matrices, pool'
        for i in (1, 2, 4, 8, 16):
            pool = Pool(i)
            print "%4d" % i, 
            print '%5.2f' % timeit('r = pool.map(mmul, matrices)', setup=stmt, number=1)
            pool.close()
            pool.join()

and my results:

$ python test_multi.py 
testing with 16 matrices
base  5.77
   1  6.72
   2  3.64
   4  3.41
   8  2.58
  16  2.47
testing with 32 matrices
base 11.69
   1 11.87
   2  9.15
   4  5.48
   8  4.68
  16  3.81
testing with 64 matrices
base 22.36
   1 25.65
   2 15.60
   4 12.20
   8  9.28
  16  9.04

[UPDATE] I run this example at home on a different computer, obtaining a consistent slow-down:

testing with 16 matrices
base  2.42
   1  2.99
   2  2.64
   4  2.80
   8  2.90
  16  2.93
testing with 32 matrices
base  4.77
   1  6.01
   2  5.38
   4  5.76
   8  6.02
  16  6.03
testing with 64 matrices
base  9.92
   1 12.41
   2 10.64
   4 11.03
   8 11.55
  16 11.59

I have to confess that I do not know who is to blame (numpy, python, compiler, kernel)...

Answer 6

Solution

Set the following environment variables before any calculation (you may need to set them before doing import numpy for some earlier versions of numpy):

os.environ["OMP_NUM_THREADS"] = "1"
os.environ["MKL_NUM_THREADS"] = "1"
os.environ["OPENBLAS_NUM_THREADS"] = "1"
os.environ["VECLIB_MAXIMUM_THREADS"] = "1"
os.environ["NUMEXPR_NUM_THREADS"] = "1"

How does it work

The implementation of numpy is already using multithreading with optimization libraries such as OpenMP or MKL or OpenBLAS, etc. That's why we don't see much improvement by implementing multiprocessing ourselves. Even worse, we suffer too many threads. For example, if my machine has 8 CPU cores, when I write single-processing code, numpy may use 8 threads for the calculation. Then I use multiprocessing to start 8 processes, I get 64 threads. This is not beneficial, and context switching between threads and other overhead can cost more time. By setting the above environment variables, we limit the number of threads per process to 1, so we get the most efficient number of total threads.

Code Example

from timeit import timeit
from multiprocessing import Pool
import sys
import os

import numpy as np

def matmul(_):
    matrix = np.ones(shape=(1000, 1000))
    _ = np.matmul(matrix, matrix)

def mixed(_):
    matrix = np.ones(shape=(1000, 1000))
    _ = np.matmul(matrix, matrix)

    s = 0
    for i in range(1000000):
        s += i

if __name__ == '__main__':
    if sys.argv[1] == "--set-num-threads":
        os.environ["OMP_NUM_THREADS"] = "1"
        os.environ["MKL_NUM_THREADS"] = "1"
        os.environ["OPENBLAS_NUM_THREADS"] = "1"
        os.environ["VECLIB_MAXIMUM_THREADS"] = "1"
        os.environ["NUMEXPR_NUM_THREADS"] = "1"

    if sys.argv[2] == "matmul":
        f = matmul
    elif sys.argv[2] == "mixed":
        f = mixed

    print("Serial:")
    print(timeit(lambda: list(map(f, [0] * 8)), number=20))

    with Pool(8) as pool:
        print("Multiprocessing:")
        print(timeit(lambda: pool.map(f, [0] * 8), number=20))

I tested the code on an AWS p3.2xlarge instance which has 8 vCPUs (which doesn't necessarily mean 8 cores):

$ python test_multi.py --no-set-num-threads matmul
Serial:
3.3447616740000115
Multiprocessing:
3.5941055110000093

$ python test_multi.py --set-num-threads matmul
Serial:
9.464500446000102
Multiprocessing:
2.570238267999912

Before setting those environment variables, the serial version and multiprocessing version didn't make much difference, all about 3 seconds, often the multiprocessing version was slower, just like what is demonstrated by the OP. After setting the number of threads, we see the serial version took 9.46 seconds, becoming much slower! This is proof that numpy is utilizing multithreading even when a single process is used. The multiprocessing version took 2.57 seconds, improved a bit, this may be because cross-thread data transferring time was saved in my implementation.

This example didn't show much power of multiprocessing since numpy is already using parallelizing. Multiprocessing is most beneficial when normal Python intensive CPU calculation is mixed with numpy operations. For example

$ python test_multi.py --no-set-num-threads mixed
Serial:
12.380275611000116
Multiprocessing:
8.190792100999943

$ python test_multi.py --set-num-threads mixed
Serial:
18.512066430999994
Multiprocessing:
4.8058130150000125

Here multiprocessing with the number of threads set to 1 is the fastest.

Remark: this also works for some other CPU computation libraries such as PyTorch.

Answer 7

Since you mention that you have a lot of files, I would suggest the following solution;

• Make a list of filenames.
• Write a function that loads and processes a single file named as the input parameter.
• Use Pool.map() to apply the function to the list of files.

Since every instance now loads its own file, the only data passed around are filenames, not (potentially large) numpy arrays.

Answer 8

I also noticed that when I ran numpy matrix multiplication inside of a Pool.map() function, it ran much slower on certain machines. My goal was to parallelize my work using Pool.map(), and run a process on each core of my machine. When things were running fast, the numpy matrix multiplication was only a small part of the overall work performed in parallel. When I looked at the CPU usage of the processes, I could see that each process could use e.g. 400+% CPU on the machines where it ran slow, but always <=100% on the machines where it ran fast. For me, the solution was to stop numpy from multithreading. It turns out that numpy was set up to multithread on exactly the machines where my Pool.map() was running slow. Evidently, if you are already parallelizing using Pool.map(), then having numpy also parallelize just creates interference. I just called export MKL_NUM_THREADS=1 before running my Python code and it worked fast everywhere.