After reboot, I still see the same values of about 22-24 GB/s for triad
on 1.6-2.7 GB data set.
I tried this benchmark on the host node "ftlogin2", corrected to AVX
compilation versus AVX512 compilation:
OMP_NUM_THREADS=8 KMP_AFFINITY=scatter ./stream
Function Rate (MB/s) Avg time Min time Max time
Copy: 41624.1384 0.0464 0.0461 0.0467
Scale: 40246.4026 0.0480 0.0477 0.0485
Add: 39071.5522 0.0741 0.0737 0.0747
Triad: 38904.1905 0.0744 0.0740 0.0751
Store: 42255.5550 0.0229 0.0227 0.0231
This machine has 4 channels of DDR3 800-1600, so the expected bandwidth
is somewhere in between 25.6 - 51.2 GB/s. I have measured 38.9 GB/s,
seems ok.
Vitali
For reference:
STREAM triad has degraded on small data size about 1.6 GB.
I would expect that, in cache mode, the result should be identical, as
the entire data set fits into IPM.
Threads 1 channel 16 GB 6 channels192 GB
4 22798.1543 12401.8232
8 34495.5022 19618.5959
16 61197.7447 21515.0307
32 90566.432622264.3076
64 172575.511725265.8044
