Paper Writing & Scaling Test

MHD Methods Paper

This week I finished the first draft of the MHD methods paper and did some revisions. It’s all coming together and I hope to submit in the next 4-8 weeks.

_ctSlope Template and Test

I tried out making all the compile time literal arguments in the _ctSlope function into template arguments. Turns out the compiler was already doing that and it made no difference. I also added a test for that function in PR #320.

Better Scaling Tests

I found a few issues with how I was do scaling tests, namely using the wrong timer and outputting data. After fixing those issues in my various scripts I started another round of scaling tests for both PLMC and PPMC; I’m still waiting for them to get through the queue. The changes I made for this are in PR #321.

After some more investigation it does appear that the lowered performance in large runs is due to a small number of poorly performing GPUs. The average timing for the integrator is almost identical to the minimum time but much less than the maximum, indicative of a small number of slow GPUs.

Cells Per GPU Tests

I wrote some code to check the GPU memory usage for Cholla. It turns out that we we’re underutilizing the GPUs on Frontier and can fit about 5.6 times more cells on each GPU, potentially enabling much larger simulations. PR #322.


  • Rebase various branches and PRs
  • Get PR #309 ready for review
This post is licensed under CC BY 4.0 by the author.