Hello,
I’m a guy from computer science and want to test OpenMC and then improve its performance (if possible) on big supercomputers.
Currently, I don’t know which inputs to use, which are meaningful for the OpenMC community.
Could someone point me to some medium to large inputs?
Thank you!
Hi Junchao,
A good model to start with would be the Hoogenboom-Martin benchmark. We’ve done a lot of our analyses using that model, although now more work is shifting to the BEAVRS benchmark as it a lot more realistic. I’ll see if we can get you a model of BEAVRS to play with as well.
When you say “improve its performance”, are you talking specifically about parallel performance or just performance in general?
Best regards,
Paul
Hi, Paul,
I want to improve its parallel performance. I read a few papers from Andrew R. Siegel on XSBench, which is a mini-app of OpenMC. From my understanding of these papers, OpenMC is embarrassingly parallel and the bottleneck is on-node performance. More specifically, it is due to random accesses to the cross section data. I find Andrew’s energy banding algorithm is interesting and want to play with it (BTW, why is it not adopted by OpenMC if its performance is better?). Also, it seems that on-node memory size is a constraint for MC. I read your papers on using RMA in MC. I want to further investigate the performance implication and to see whether we can reach the same performance as if we were using on-node memory. I hope these research problems are real concerns instead of fake problems in the OpenMC community.
Thank you.
Hi Junchao,
The short answer as to why energy banding is not implemented in OpenMC is because it would require considerable reorganization of the data structures and algorithms. My personal view is that we should have a better understanding of on-node contention issues before undertaking a significant restructuring of the code; I think the proxy- and mini-apps have done a great job at starting to give us insights, and I’m hoping that further studies will continue to improve our knowledge so that we have a clearer path forward. These problems are absolutely real concerns for the future of MC, not just academic exercises. I’m interested to hear where you go with your work; please let me know if there’s anything more I can do to assist.
Best regards,
Paul