2021.02.18 Meeting Notes

Agenda

• Individual/group updates
• Discuss if we should keep the ran2 code
  • Decided no, use Kokkos random number generator instead, place responsibility on downstream codes
• Scaling
  • Results waiting for running on Sierra - Galen
  • Performance on large mesh blocks is very promising - Phil
• MPI tag error
  • Conclusion of discussion, this should be a solved problem, not yet a priority,
• Review non-WIP PRs

Individual/group Updates

LANL CS

Jonas

• Spent time building on Cray system, only thing remaining is the ability to read HDF5 files
• Getting updated on Andrew's work on sparse
• Still has questions on what the right policy is for sparse

Josh Brown

• PR for performance metrics is ready for review
• Will be looking into getting performance regression setup on cpu's as well, and for AMD systems

Galen

• Sierra everything is built, and tests run, no results yet

LANL Physics

Josh Dolence

• Been working on getting a physics code open sourced that is not RIOT

Jonah

• Been working on getting access to RZ systems and been troublshooting

AthenaPK

Phil

• Reviewed Ben's pr on device locator
• Working on an implementing of Kelvin-Helmholtz solver in AthenaPK so can compare with RIOT
• Discovered a free error, is able to reproduce on power9 with volta and with intel systems

Jim

• Has been working on a test code, looking at making reconstruction more robust using WENO-Z.
• Has MHD working in pack in one, now looking at making it more performant.
• Will be looking to see if supersonic is more robust with WENO-Z

General Notes

Jimm - made recommendation to be careful when using inline functions in Kokkos kernels, can incur performance penalties because of how the registers are handled.

• Suggestion was made to use templates
• Try Ben's implementation of vtable
• Or is there a way to share registers between branches