Open aturner-epcc opened 6 years ago
I think at least some of this is related to compiler options on Thomas. I've done some experimentation using the compiler options for CSD3-Skylake on Thomas and other UCL systems and they dramatically improve the performance.
I've put Makefile.ucl files in the repo with these options for OpenSBLI and updated the Thomas build instructions.
I'll commit back results as they come through the scheduler.
The Thomas results for OpenSBLI do not seem to follow the performance trends we expect to see. Why is this? Do we see an improvement for Tesseract in the same way for Peta4-Skylake with the additional memory channels?