Scaling tests of linear algebra operations

This notebook documents how various linear algebra operations scale on Haswell vs. KNL.

Stephen Bailey
Lawrence Berkeley National Lab
Spring 2017

Setup at NERSC

module load python/3.5-anaconda
cd $SCRATCH/desi/knlcode/knltest/code
python linalg-scaling.py

srun -n 1 -c 24 python linalg-scaling.py
srun -n 1 -c 256 python linalg-scaling.py

#- KNL
export OMP_NUM_THREADS=256
srun -n 1 -c 256 --cpu_bind=cores python linalg-scaling.py

#- HSW
export OMP_NUM_THREADS=64
srun -n 1 -c 64 --cpu_bind=cores python linalg-scaling.py

export OMP_NUM_THREADS=32
srun -n 1 -c 32 --cpu_bind=cores python linalg-scaling.py

To do

repeat tests with IDP 2 instead of Anaconda distribution
add scaling tests for matrix-matrix multiply, matrix-vector multipy, and dot products



In [1]:

    
%pylab inline
import numpy as np
from astropy.table import Table









    



Populating the interactive namespace from numpy and matplotlib



In [2]:

    
mac = Table.read('data/linalg-scaling/mac.txt', format='ascii')
hsw = Table.read('data/linalg-scaling/hsw.txt', format='ascii')
knl = Table.read('data/linalg-scaling/knl.txt', format='ascii')



In [5]:

    
figure(figsize=(12,6))
for i, method in enumerate(['svd', 'chol', 'eigh']):
    subplot(231+i)
    plot(hsw['n'], hsw[method], label='Haswell')
    plot(knl['n'], knl[method], label='KNL')
    # plot(mac['n'], mac[method], label='mac')
    xticks([0,1000,2000,3000,4000])
    if i == 0:
        ylabel('Time per linalg calculation')
        legend(loc='upper left')
    subplot(234+i)
    plot(hsw['n'], hsw[method]/hsw[method]); xlabel('{} size'.format(method)); ylim(0,6)
    plot(knl['n'], knl[method]/hsw[method]); xlabel('{} size'.format(method)); ylim(0,6)
    # plot(mac['n'], mac[method]/hsw[method]); xlabel('{} size'.format(method)); ylim(0,6)

    xticks([0,1000,2000,3000,4000])
    if i == 0:
        ylabel('Time relative to Haswell')



In [7]:

    
figure(figsize=(12,3))
for i, method in enumerate(['svd', 'chol', 'eigh']):
    subplot(131+i)
    plot(hsw['n'], (4000/hsw['n'])*hsw[method], label='Haswell')
    plot(knl['n'], (4000/hsw['n'])*knl[method], label='KNL')
    xlabel('{} size'.format(method))

subplot(131)
ylabel('Time for 4000/size calculations')









    Out[7]:





<matplotlib.text.Text at 0x10beadb50>



In [ ]: