This repository has been archived by the owner on Oct 23, 2020. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 0
Agenda Notes Tuesday September 5th 2017 11:00 AM
Michael Duda edited this page Sep 5, 2017
·
6 revisions
Date: 5 September 2017
Time: 11:00 MDT / 19:00 CET
Call-in number: 1-866-740-1260
Access Code: 4978161
- Guest Bill Arndt of LBL, working with Mark, will explain potential threading problem with excessive barriers. See
- Mark: proposal for MPAS-QuickViz repo: A way to share visualization and analysis scripts, but with a low bar for reviews and testing.
- sample: https://github.com/mark-petersen/testQuickViz
- proposed location: https://github.com/MPAS-Dev/MPAS-QuickViz/
- Framework modifications:
Participants: Bill Arndt, Dom, Mark, Matt, Michael
Bill has been looking at threading issues -- MPAS-O is really slow on KNL
- Through testing and profiling, significant portion of time in MPAS-O on KNL were tracked down to halo communications (specifically, OpenMP synchronization issues)
- in code linked in the agenda, we have messages on the order of kB and overhead in loops dominating runtime
- serializing the buffer packing loop significantly sped up halo exchange (~40% faster)
- L.8451 in mpas_dmpar.F works if it becomes an OMP MASTER directive
- maybe we could create some code in the testing core to exercise this halo exchange only?
- predicates on buffer size probably won’t help?
- maybe a minimum chunk size could have similar effect to predicates?
All agreed that this sounds good
- Committing too many multi-MB files could make the repository size unwieldy, in which case we could periodically rebuild the repo starting only several months in the past
- Probably no need to create PRs unless changing someone else's scripts