Friday, May 23, 2014

Code parallelization with joblib

Recently I've been working on the parallelization of some Python code and I discovered Joblib. It is a library that supports pipelining and offers a good support for parallelization. In this post we will implement a (very naive) paraller matrix by matrix multiplication algorithm to show the parallelization capabilities of this library.
from joblib import Parallel, delayed

def parallel_dot(A,B,n_jobs=2):
     Computes A x B using more CPUs.
     This works only when the number 
     of rows of A and the n_jobs are even.
    parallelizer = Parallel(n_jobs=n_jobs)
    # this iterator returns the functions to execute for each task
    tasks_iterator = ( delayed(,B) 
                      for A_block in np.split(A,n_jobs) )
    result = parallelizer( tasks_iterator )
    # merging the output of the jobs
    return np.vstack(result)
This function spreads the computation across more processes. The strategy applied to distribute the data is very simple. Each process has the full matrix B and a contiguous block of rows of A, so it can compute a block of rows A*B. In the end, the result of each process is stacked to build final matrix.

Let's compare the parallel version of the algorithm with the sequential one:
A = np.random.randint(0,high=10,size=(1000,1000))
B = np.random.randint(0,high=10,size=(1000,1000))
%time _ =,B)
CPU times: user 13.2 s, sys: 36 ms, total: 13.2 s
Wall time: 13.4 s
%time _ = parallel_dot(A,B,n_jobs=2)
CPU times: user 92 ms, sys: 76 ms, total: 168 ms
Wall time: 8.49 s
Wow, we had a speedup of 1.6X, not bad for a so naive algorithm. It's important to notice that the arguments passed as input to the Parallel call are serialized and reallocated in the memory of each worker process. Which means that the last time that parallel_dot have been called, the matrix B have been entirely replicated two times in memory. To avoid this problem, we can dump the matrices on the filesystem and pass a reference to the worker to open them as memory map.
import tempfile
import os
from joblib import load, dump

# saving A and B to a local file for memmapping
temp_folder = tempfile.mkdtemp()
filenameA = os.path.join(temp_folder, 'A.mmap')
dump(A, filenameA)
filenameB = os.path.join(temp_folder, 'B.mmap')
dump(A, filenameB)
Now, when parallel_dot(A_memmap,B_memmap,n_jobs=2) is called, both the processes created will use only a reference to the matrix B..


  1. why if you set B to:
    B = np.random.randint(0,high=10,size=(1000,1))
    is using parallel slower

    1. Because the number of operations required is too small to achieve a speedup using parallelization. In this case, the creation of the processes and the splitting of the data probably require more time than the matrix by matrix multiplication itself.