perf-matrix-mul.py

import time
import argparse
import multiprocessing
import random

import numpy as np

M = 250
# size ~ 0.5MB
X = np.random.randn(M, M)
# to allow CPU with 6 cores to end tasks at once
TASK_N = 180
# size ~ 90MB
MATRICES = 180


def small_mat_test(i):
    for i in range(TASK_N):
        X.dot(X)


def small_mat(args):
    """
    Testing small matrices that should fit in CPU cache.
    """
    pool = multiprocessing.Pool()

    start_time = time.time()
    pool.map(small_mat_test, list(range(args.n)))
    elapsed_time = time.time() - start_time

    return elapsed_time


def list_mat(args):
    """
    Testing list of bigger matrices that probably not fit in CPU cache.
    """
    arrays = []
    for i in range(MATRICES):
        arrays.append(np.random.randn(M, M))

    def list_mat_test(i):
        idx1 = list(range(MATRICES))
        idx2 = list(range(MATRICES))
        random.shuffle(idx1)
        random.shuffle(idx2)
        for i in range(MATRICES):
            arrays[idx1[i]].dot(arrays[idx2[i]])

    pool = multiprocessing.Pool()

    start_time = time.time()
    pool.map(small_mat_test, list(range(args.n)))
    elapsed_time = time.time() - start_time

    return elapsed_time


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument('--suite', choices=('small_mat', 'list_mat'), default='small_mat')
    parser.add_argument('--n', help='Number of iterations', default=3*16)
    args = parser.parse_args()

    if args.suite == 'small_mat':
        elapsed_time = small_mat(args)
        print('[Small Matrix Test] Elapsed: {:0.3f} seconds Matrix: {}x{} Multiplications: {} MulPerSec: {:.2f}'.format(
            elapsed_time, M, M, args.n * TASK_N, (args.n * TASK_N)/elapsed_time))
    elif args.suite == 'list_mat':
        elapsed_time = list_mat(args)
        print('[List Matrix Test] Elapsed: {:0.3f} seconds Matrix: {}x{} x {} Multiplications: {} MulPerSec: {:.2f}'.format(
            elapsed_time, M, M, MATRICES, args.n * MATRICES, (args.n * MATRICES) / elapsed_time))
    else:
        raise ValueError('Nobody expects invalid parameter!')

if __name__ == '__main__':
    main()

My various results, more or less correlate with https://www.cpubenchmark.net/

AMD Phenom(tm) II X6 1090T (6/6 cores/threads @ 3.2GHz)
    Python 3.4.3 (default, Nov 17 2016, 01:08:31)
    [Small Matrix Test] Elapsed: 6.224 seconds Matrix: 250x250 Multiplications: 8640 MulPerSec: 1388.27
    [List Matrix Test] Elapsed: 6.537 seconds Matrix: 250x250 x 180 Multiplications: 8640 MulPerSec: 1321.78
        
    Python 3.6.1 |Anaconda 4.4.0 (64-bit)
    [Small Matrix Test] Elapsed: 7.070 seconds Matrix: 250x250 Multiplications: 8640 MulPerSec: 1222.03
    [List Matrix Test] Elapsed: 7.141 seconds Matrix: 250x250 x 180 Multiplications: 8640 MulPerSec: 1209.96
    
Intel(R) Celeron(R) CPU  J1900 (4/4 cores/threads @ 1.99GHz)
    Python 3.5.2 (default, Nov 17 2016, 17:05:23)
    [Small Matrix Test] Elapsed: 37.147 seconds Matrix: 250x250 Multiplications: 8640 MulPerSec: 232.59
    [List Matrix Test] Elapsed: 37.301 seconds Matrix: 250x250 x 180 Multiplications: 8640 MulPerSec: 231.63
    
Intel(R) Core(TM) i3-5005U CPU (2/4 cores/threads @ 2.00GHz)
    Python 3.5.2 (default, Nov 17 2016, 17:05:23)
    [Small Matrix Test] Elapsed: 21.233 seconds Matrix: 250x250 Multiplications: 8640 MulPerSec: 406.90
    [List Matrix Test] Elapsed: 21.590 seconds Matrix: 250x250 x 180 Multiplications: 8640 MulPerSec: 400.18