add benchmark for matrix multiplication

hpp/benchmark_mm/bench.py

@@ -0,0 +1,29 @@
+
+import os
+import sys
+import timeit
+
+import numpy as np
+import threadpoolctl as th
+
+
+def multiply_matrix(N, T):
+    x = np.zeros((N, N), dtype="float64")
+    with th.threadpool_limits(limits=T, user_api='blas'):
+        y = x @ x
+
+
+if __name__ == '__main__':
+    processes = sys.argv[1]
+    times = []
+    results = open(f'results_{processes}', 'wt')
+    for n_threads in range(len(os.sched_getaffinity(0))):
+        for matrix_size in [30, 300, 3000]:
+            time = min(timeit.repeat(lambda: multiply_matrix(N=matrix_size, T=n_threads), number=5))/5
+            results.write(f'{n_threads}_{matrix_size}_{time}\n')
+            results.flush()
+    results.close()
+
+
+
+

hpp/benchmark_mm/plot.py

@@ -0,0 +1,30 @@
+import matplotlib.pyplot as plt
+import pandas as pd
+
+data = []
+for i in range(1, 4):
+    with open(f"hpp/benchmark_mm/results_{i}", "r") as f:
+        # each row is: processes, threads, size, time
+        data.extend([f"{i}_{line}".split("_") for line in f.readlines()])
+
+df = pd.DataFrame(data, columns=["processes", "threads", "size", "time"])
+
+fig, ax = plt.subplots(3, 1)
+fig.set_figheight(10)
+fig.set_figwidth(6)
+
+for i, s in enumerate(df["size"].unique()):
+    for p in df["processes"].unique():
+        xi = list(df.loc[(df["processes"] == p) & (df["size"] == s), "threads"].astype(float))
+        yi = list(df.loc[(df["processes"] == p) & (df["size"] == s), "time"].astype(float))
+        ax[i].plot(xi, yi, label=p)
+    ax[i].set_title(f"size={s}")
+
+handles, labels = ax[i].get_legend_handles_labels()
+fig.legend(handles, labels, loc='upper center', title="processes")
+
+fig.text(0.5, 0.04, 'n_threads', ha='center')
+fig.text(0.0, 0.5, 'time', va='center', rotation='vertical')
+plt.savefig("hpp/benchmark_mm/results.png", dpi=300)
+
+

hpp/benchmark_mm/results_1

@@ -0,0 +1,24 @@
+0_30_0.00027460739947855474
+0_300_0.0006913389981491492
+0_3000_0.4103931115998421
+1_30_0.00025021140172611924
+1_300_0.0019851766002830116
+1_3000_1.1397250042005909
+2_30_0.00022756299877073615
+2_300_0.0008770572021603584
+2_3000_0.6591780591988936
+3_30_0.0003005806007422507
+3_300_0.000931622800999321
+3_3000_0.5624800918012625
+4_30_0.00035497880016919223
+4_300_0.0012595637992490083
+4_3000_0.5415630329982377
+5_30_0.0003646780009148642
+5_300_0.0009647636004956439
+5_3000_0.5020334165979875
+6_30_0.0003636729990830645
+6_300_0.0009608207998098806
+6_3000_0.49886566500063056
+7_30_0.00037808019842486826
+7_300_0.001537017000373453
+7_3000_0.5028969450009754

hpp/benchmark_mm/results_2

@@ -0,0 +1,24 @@
+0_30_0.0003037196001969278
+0_300_0.0007952529995236546
+0_3000_0.5004591826000251
+1_30_0.0003733070014277473
+1_300_0.0028704026015475394
+1_3000_1.2028277851990423
+2_30_0.0002753072010818869
+2_300_0.0010893295984715223
+2_3000_0.7504769150022185
+3_30_0.0003262779995566234
+3_300_0.0009243134001735598
+3_3000_0.5812608784006443
+4_30_0.00032918679935391994
+4_300_0.001281012801337056
+4_3000_0.5160886729980121
+5_30_0.0003709879994858056
+5_300_0.0012151438015280292
+5_3000_0.5070040075981523
+6_30_0.0008059760002652183
+6_300_0.0009575574018526823
+6_3000_0.5152194449998205
+7_30_0.0004826569987926632
+7_300_0.0013204805989516899
+7_3000_0.4981165808014339

hpp/benchmark_mm/results_3

@@ -0,0 +1,24 @@
+0_30_0.00032405359961558134
+0_300_0.0013136355992173776
+0_3000_0.49723099039983937
+1_30_0.000379611601238139
+1_300_0.0029519002011511473
+1_3000_1.2366342330002227
+2_30_0.00024287900014314802
+2_300_0.0009623870020732284
+2_3000_0.6633257908019005
+3_30_0.0003105259995209053
+3_300_0.000884313802816905
+3_3000_0.5715091922000284
+4_30_0.0003153233992634341
+4_300_0.0013229422009317204
+4_3000_0.4965488420013571
+5_30_0.00035848779953084886
+5_300_0.000948840199271217
+5_3000_0.4973357428010786
+6_30_0.00038418699987232687
+6_300_0.000954620199627243
+6_3000_0.5065642214001855
+7_30_0.0003947592020267621
+7_300_0.0009453961974941194
+7_3000_0.5630277571995975

hpp/benchmark_mm/submit.sh

@@ -0,0 +1,7 @@
+# the number of python processes to start, P=3 if not set
+P=${1:-3}
+# the size of the matrix, use the default in overcommit.py if not set
+for i in $(seq $P); do 
+    echo -n "Starting process ${i}… "
+    python bench.py $i
+done