Attendees: Ken Moreland (SNL), Dave Pugmire (ORNL), Hank Childs (UO), Nick Davis (SNL), David Rogers (LANL), Ollie Lo (LANL), Matt Larsen (LLNL), Sujin Philip (Kitware), James Kress (ORNL), Vicente Bolea (Kitware)
- ECP/VTK-m looks a little underspent.
- Not by much. We look to be pretty on track. (Especially since contracts probably have a sizable delay.)
- Dave might check out the ORNL budget just in case.
- We are being asked by Jim and Mike if we are ready to plug VTK-m into the CI infrastructure on the EAS platforms.
- Details are here (slide 17): https://confluence.exascaleproject.org/display/1ST/2020-10-19+Meeting+Notes+and+Actions
- ECP Annual Meeting: April 12-15
- Submissions for tutorials, breakouts, and other things open (deadline December 15)
- Do we want to organize VTK-m tutorial or breakout? (Usually couple with Alpine.)
- Might have presence at Alpine and/or ADIOS events
- ECP Independent Project Review pushed from October to February
- Means several of our deadlines have been pushed back
- ECP ST Annual Reviews: Delayed to Dec 7-11
- Still "keen" on collecting information about early access systems by end of August
- Let's think about possible highlights
- App engagement
- Success with porting (e.g. run on Aurora)
- Engagement with other ST teams (e.g. using Kokkos)
- Submitted highlight on integration with EFFIS
- Expect to submit another highlight when used in a big run
- Possible highlight slide on particle advection with WarpX.
ECP task updates
Need to just get something up on the Iris/Tulip CI dashboards. Even if it is something simple, we need to show that we are running on these computers.
There will be a hackathon for Iris in a couple of weeks. Some folks from VTK-m might be interested in that.
Matt Larsen has run into an issue with CUDA 11 where it is complaining while compiling DIY that the std:: namespace does not seem to exist. The issue might be related to new CUDA features that support more std objects on devices. As of now, no one has a good solution.
Follow up for benchmarking of Kokkos. Do we have profiling of what is being run during the "slow" filter benchmarks? Sujin is working on changes to the Kokkos device adapter that use Kokkos algorithms like scan and sort for the cases where it is supported. After that we can benchmark again.
Sujin implementing reduce functionality in Kokkos. Kokkos needs to know the identity value for the operation.
Rename master branch?
- Need to wait for some changes to git before the transition (for some automated features of repo).
- Can we for a time have a master that mirrors the new default?
- Don't know, but seems possible. Will have to look into that.