2021.03.18 Meeting Notes
- Individual/group updates
- Scaling update
- IO (performance and default basename handling)
- Argonne GPU Hackathon (pgrete)
- C++17 (pgrete)
- TODO
- Paper
- Next release
- Review non-WIP PRs
- Added Documentation for merging from forks
- Made some progress on generating performance metrics with Darshan for IO
- Opened up a pr to cleanup parameter inputs class
- Gitlab workshop
- Resolved issues with internal gitlab setup
- Still refactoring gitlab pr based on feedback
Working on other projects
- Working through particle pr feedback, investigating possible mpi communication issues
- Went through particle pr
- Added regression testing for leap frog integrator
- Looking at particle accelerator, particles don't seem to be passing through meshblocks correctly.
- Has been looking into options for keeping kernels simple to avoid performance issues related over subscribing registers.
- Looked at static mesh refinement, looks like parthenon already has static mesh refinement implemented just needs to be documented.
- Looking into adding optional outputs to allow compatibility with Yt.
- Thinking through fix for solver's that Phil is working on.
- p9/volta seems to be crashing due to device memory allocations
- The p9 numbers are not really as great as we would like to see, Phil believes there is still room for an order of magnitude increase in performance.
- Problems, with the load balancing are causing the performance numbers to be unpredictable.
IO performance metrics from RZAnsel
- Not much to report, will probably try to figure out what is going on with the load balancing during the hackathon.
- Phil brought up looking into switching to the c++17 standard, the motivation largely being for using constexpr functionality to fix the performance issues seen in the kokkos kernels.
- Kokkos will be updating later this year to supporting c++17
- It also does not appear that switching to supporting c++17 will break anything downstream
- It is worth pointing out that Sierra and Summit will not be supporting Kokkos with the c++17 features, though there are ways around this.
- Finish profiling hdf5 io performance (Joshua Brown), use weak scaling, look at write rates.
- Open branch for testing compiling with c++17 (Phil)
- Yt compatibility (Forrest)
- Create next release (Phil)
- Will wait to discuss when Josh Dolence is present.
- Still waiting to here about what medium is most appropriate for collaborating on a paper.
- Does not appear to be any reason to wait to create the next release.