2022.04.08 Meeting Notes

Jump to bottom

Philipp Grete edited this page Apr 8, 2022 · 1 revision

Agenda

Individual/group updates

JR

setting up data structure for buffer pools (to reduce mem alloc for sparse vars)
new class to abstract bound "info"
- original boundary val info (neighbors etc.) does not need to be tied to "variables"
- now separate
- will make this separate, then test in parallel, then remove old

FG

will share link to defense next Thursday
planning some shearing box sims later this year targeting jets, which likely need face-centered CT
will work on CT and face fields around late April
for coordinates, we will setup a separate meeting in due time

PG

Fixed phys_id (tag) bug and implemented separate communicator for swarms (which doesn't work as expected yet. BR is also looking into this)
Got time for texascale days next Monday. Will run scaling tests for the paper. Can also run other tests if people provide input files.

Comm patterns

One comm per var? YES
What's the largest num of comms? ~8k
PG will check if this can be addressed open PR on phys_ids

Parthenon user/dev meeting/workshop

Looks like ~10 external (from LANL) people can be expected
JD will check admin aspect of a meeting in a conference room at LANL
tentative date: first week of Oct

Benchmarking (AJ)

current tests
- testing with Phoebus, 3d blast wave
- 10x slowdown for using 128 ranks with 1 thread/core versus 8 ranks with 16 threads each
- not clear what's going on. Typical behavior observed is the other way round
- will share input file to reproduce on different system
microbenchmarks
- what would be useful pattern?
- more realistic, e.g., static multilevel mesh but varying num of vars and ghost zones