Teleconference 2016-11-09

From VTKM
Jump to: navigation, search

Attendees: Ken Moreland (SNL), Chris Sewell (LANL), Tom Otahal (SNL), Matt Larsen (LLNL), Rob Maynard (Kitware), Brent Lesley (OU)

Chris now has access to a cluster hosted by Nvidia that has PASCAL and NVLINK hardware. This provides him an opportunity to compare performance of streaming vs just using unified memory.

Claire Guibaud from CEA has been visiting LANL. During her stay she has been working on code to compute atom distances. She is now nearing completion.

Tom recently got an account on Ellis, a KNL testbed machine at Sandia. He has been able to build VTK-m and run the tests although there appear to be 2 test failures. There has not been an opportunity to test performance yet. Tom is also having some trouble with avx512 instructions with the Intel compiler.

Matt has been working on Strawman and the related VTK-m work.

Matt also has run into some issues with needing to represent fields with vectors of around 100 components each. This is problematic for VTK-m because the vectors are of fixed length and generally small. The best solution would be to not have component lengths that long. This will like lead to very poor memory performance. Barring that, there are some solutions using Vec-like objects that can adapt the size.

Rob is working on gradient point to cell. To implement this efficiently, he needs a mechanism to have a cell to point map and then get the full cell information on each cell incident on the destination point. To do this he is introducing the concept of a whole cell set ControlSignature tag. He has a branch that implements this. It is working OK but it has slowed down compiling significantly due to the many cases of different cell set types.

Brent is working on a maximal clique algorithm. He has seen some good timing on some large graphs but there are some issues with other graphs. This is still a work in progress.

Ken is working on a branch that changes all of the export macros. This change is likely to cause merge issues with all open merge requests and topic branches, so we want to time its merge during a slow period. Chris has an open issue with the streaming work, but that should be in soon. If everything settles, we might try to merge this in next week.