2024.08.22 Meeting Notes

Agenda

• Individual/group updates
• Next release
• time offset >dt in time integrators
• par dispatch
• Review non-WIP PRs

Group updates

• Roberts: Little thinking about how to make fused loops. WIP PR. Uses indexers to clean things up a bit.
  • Some questions of performance but it will probably work
  • Also some concerns about additional complexity
  • But much more flexible
• Miller
  • PIC code building on parthenon
  • Some tutorial materials on parthenon. JMM will share with pgrete later.
• Prather:
  • Some small 2 line changes.
  • KHARMA on Venado. Excellent performance on 1 GPU but bad scaling.
• Barker:
  • Phoebus runs on mutiple GPUS with grace-hopper
  • pgrete will share instructions.
  • pbrady: Cuda Unified memory produced substantial performance boost
• Brady:
  • Time offsets discussion
• Ryan:
  • Performance problems in particles infrastructure improved. Factor of 4-5x performance improvement.
• Grete:
  • Production sims with particles. Will follow up with performance measurements.
  • Buffer pool memory usage reduction.
    • maybe need to add some overridable capability when doing sparsity
  • Team policy/comm buffer filling. Experimenting with small block limit.

Next Release

We should do one. Will wait for swarms PR and/or open PRs that are close. Buffer pooling + swrms. Will delay loop abstraction and openpmd.

Will push par reduce inner PR in right now too.

Time integrators

Mathematica derivation to be added to repo. Mathematicians to be reached out to.

Par dispatch

Delayed until next time