2021.04.01 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
- Worked on trying to get performance data for io
- Trying to debug compiler issues with mpich
- Also been looking into hdf5 bug
- Tried to close currently open pr's
- Has mainly been working on sparse update. Dense on block is now implemented but not fully tested.
- Previously, it was assumed that you could take the first block to figure out what fields were on all the blocks, this is no longer true, so a global reduce is now used.
- Has also been working on implementing HDF5 compression seems to be getting some good numbers with lossless compression. However, the example he has been using is a bit trivial e.g. lots of 0's which make compression a bit more simplistic than what probably will be encountered. Factor of 20 reduction in file size.
- Some points worth noting is that hdf5 has some issues with all but the latest versions with parallel io and compression.
No one was present due to spring break etc...
- In previous meeting, noted that we did not see the performance improvements we expected when comparing HSW nodes with power9 this was suggested to be a result of using too trivial of a problem. After using more realistic calculations in AthenaPK noted about 8x - 11x improvement in performance.
- Opened a pr to cleanup old OpenMP code that was actually causing problems.
- Opened a pr to right history file output
- Still working on editing hdf5 output so as to be compatible Yt.
- Added a way to name individual rows.
- The only moderate output added is the cell centers which were not previously output. This will be useful information beyond Yt for more complex coordinate systems other than Cartesian.
No updates
Joshua Brown - Ran into some problems with tests crashing when trying to read hdf5 on Trinitite. Do have some preliminary data from RZAnsel, but not for enough nodes to get anything meaningful.
- Team update is occuring next week Wednesday
- Both Jonas and Phil have accounts
- Need to figure out what exaclty is the roadblock with this, there is some confusion as to how much support exists for c++17 with the xl compilers.
- Part of the problem is only the latest CUDA versions have support for c++17.
- Kokkos will only be supporting c++17 after it's next major release some time later this year. So this has to move forward eventually anyway.
- Joshua - Get write rates on Trinitite for larger node count, increase output to around 30 gb, switch to using static mesh or static amr - also be useful to output mesh block count
Just needed on more approval before the merge was made