2024.11.14 Meeting Notes

Jump to bottom

Philipp Grete edited this page Nov 14, 2024 · 2 revisions

Individual/group updates

JM

Brandon figured out on Venado that HDF is not working Venado (Cray issue)
- even custom build didn't work (underlying issue with MPI lib)
- BW reports that he successfully build OpenMPI using Cray's libfabric on Frontier -- might be worth a try
- might try OpenPMD backend (but need non-cc fields)

PM

Artemis is open source (for planet formation)
good example for typing system and sparse packs

LR

suffering through coalesced MPI comm PR
- for some edge case there are still minimal differences
- will add doc (for general doc and in case so has to take over)
will try to get solver PR in before leave

BB

Metadata::None output fix got merged

BR

MC transport also open source (in addition to Artemis)
example integrated in Artemis
wrote scripts that runs CI on darwin and reports back to Github
- should be fairly straightforward to adapt to parthenon upstream

FG

working on Artemis performance
upcoming: looking at more detail kernel performance on Venado

AR

fixing merge conflicts in unified par_...
support for registering types with launch bounds translation to index ranges
hydro is ported
upcoming: interfacing with HYPRE

BW

debugging stability issues
- will hopefully be solved by retrying timesteps

PG

submitted proposal for Exascale pioneer call
AthenaPK/Parthenon to be advertised through JUREAP
yt4.4 is out (with Parthenon frontend)
Having issues at large scale on Lumi (but those are machine dependent)
Fixed View of view allocs in 4.4.1

Timesteps <-> Driver

Would be nice to have the capability to redo a timestep (with smaller dt) as backup
What could break?
- restoring params?
Might also want/need to move the calculation of the timestep outside the Step() functions to before Step()

CI tests/limitations

Testing debug build did help in some cases
Probably enough to do on host
Try to build without debug symbols but with debug flag -O0 <- (no -g no -G no -lineinfo)

View of views

What shall we do? SequntialHostInit? Unmanaged inner views? SharedMemSpace outer views rather that host/device copies?
Could we even use mdspan as a container for the outer view? Probably not because we need device pointer on device?
Let's get the fix in (as people depend on it) and then understand afterwards why it's working
Eventually try if unamnaged view also do the trick

PR

https://github.com/parthenon-hpc-lab/parthenon/pull/1175 table for now in favor of flatter bounds

Tentative next meeting 5 Dec