2020.10.07 Meeting Notes
- Individual/group updates
- AthenaPK scaling using cached MeshBlockPacks and single kernel buffer send/set functions:
- AthenaPK scaling using cached MeshBlockPacks and single kernel buffer send/set functions:
- Discuss Regression Test Failures (Joshua and Andrew)
- Public headers vs. private headers
- Dense on Block - @gshipman
- Interface design (Container versus MeshBlockPacks) -- @pgrete
- Discuss large-scale testing for AMR at smaller mesh block sizes -- @gshipman
- Review non-WIP PRs
Mostly pull request reviews. @Joshua has been trying to wrap up public headers change.
Sriram still working on restart.
Introduced an abstraction for parallel reduce par_reduce.
Ben is getting close to have particle framework changes merged in. On large static, uniform mesh, performance looks good. Still working on GPU performance.
Intel compiler errors cropping up.
Scaling of mesh block packs. Overhead of 256^3 down to 16^3 is now 6x on GPU (compared to baseline of 4x on CPU). Factor of 100 improvement over 3 months ago. This is on a full 2nd-order hydro problem.
Forrest noticed there were a lot of limitations in Jim's code due to register usage. It's an experimental code.
Feedback from Jim:
- Multi-value reduction - basically do a reduction on N variables and produce N results
- Possibly provide a custom reduction operation or something
- Output "tabs" files