2022.07.14 Meeting Notes

Agenda

JM

revisiting prolongation in one due to performance issues with using only a single team in a kernel
solution will also apply to restriction to reduce code duplication
enrolling custom prolongation/restriction operators will be beneficial for face centered field and high-order methods

LR

working on efficient buffer allocation for sparse variables across ranks and signaling
currently there's a performance issue with using IProbe to determine size incoming message to determine if sparse variables need to be allocated
tested three different ways, can switch between optimal memory size and optimal comm
finished adding state to ParArray (and make them even more generic), e.g., to allow for a per-variable sparse threshold
open new PR to with new sparse packing machinery including introduce a type to the packing
- bug #668 still not fixed

GS

working on trying to get Parthenon/Phoebus/Riot into RFP for next gen system
need to determine good benchmarking case
- potentially both Parthenon and downstream code separately

FG

has a mockup test problem for loop advection test for MHD (plain field loop advection)
using it to determine the changes in interfaces required (e.g., many interfaces assume cell centered fields such as packing)
idea: schedule mini hackathon to talk this through
will open PR with current ideas and afterwards send mail to find date for hackathon

PG

AJ

investigating why some ranks take more time during loadbalance than others
split step into multiple regions: flux correction comm, bound comm, load balance
load imbalance (different numbers of blocks per rank) explains good chuck of different timing but not all
one more source is that not all ranks/blocks take part in flux correction
will present more in-depth data during next data and share upfront on Matrix

it currently takes a "long" time for PRs to go through
target: automate more or the review process, specifically include downstream codes in CI for interface and performance regression testing
PG will tackle the latter after reviewing sparse comm PR
also, small "administrative" PR can go through without waiting for separate downstream code developer review

Added more people to list of people to invite.
- JM will contact Charles and Ben
- PG will contact people at MSU and Princeton
- AJ will check with people at CMU
Foreign national needs about two month lead time for a visit. Everyone else about one.

Next meeting in two weeks 28. Jul