search

hpc-uk / archer-benchmarks

GNU General Public License v3.0
29 stars 20 forks source link

Use memory bandwidth per *node*, not per *core*? #25

Open dorugeber opened 5 years ago

dorugeber commented 5 years ago

I haven't looked through the code, but I copied the bandwidth and OpenSBLI performance numbers into LibreOffice, made a scatter plot out of it, and the R^2 is coming out as 0.49 (=> R = 0.7), which is very different to the near-0 numbers given in the analysis.

image

dorugeber commented 5 years ago

Oh, I think I see. The analysis is using (STREAM) bandwidth per core. That doesn't seem like a sensible measurement: there's no reason for the whole-node performance to correlate with bandwidth-per-core.

Published bandwidth per node gives a meaningful correlation; bandwidth measured by STREAM an even better one.

aturner-epcc commented 5 years ago

@dorugeber Thanks for this comment too. Will rerun this analysis and update the issue with thoughts/conclusions.