Post

HLLD Debugging 2 and Crusher

Posted Updated
By Robert Caddy 2 min read

HLLD Debugging

I found the last few bugs in the double star state and so now the HLLD solver passes all tests when operating in the primary (x) direction! The bugs were mostly typos but one test input parameters had to be redone since the HLLD solver doesn’t handle small numbers (1e-9) identically to Athena’s solver where my test data comes from. The new tests for testing the Y and Z directions in the solver, which are really just the old tests rotated 90 degrees, don’t currently pass but that is likely an indexing error that should be easy to resolve. I’ve been developing this solver with test driven development and it’s been a fantastic success. The tests immediately tells me where issues show up and make it really easy to trace backwards from there to find the culprit. Overall it’s been a great success and I will be using test driven development where I can in the future.

Crusher

The new Frontier testbed, Crusher, became available this week and is the first publid testbed to use Frontier hardware. We had a day long training on the system and I got tests, and cholla in general, running on Crusher with only minimal changes. I did find a “bug” that the margin we had for “equal” small floating point numbers wasn’t large enough. I originally set it to \( 10^{-14} \) since that is an order of magnitude larger than the difference between compiling with the XL vs GCC compilers on NVIDIA hardware. It turns out the difference between XL/GCC on NVIDIA and Clang on AMD is more like \( 10^{-12} \) for system tests so I set the margin on system tests to \( 5 \times 10^{-12} \), elsewhere the margin is still \( 10^{-14} \). I’m not worried about the larger errors because the differences are still only about one ten billionth of a percent.

Other

  • Updated this website to Chirpy v5.0.2
  • Jenkins will be available for CI soon using Pitt CRC hardware so I can hopefully set up automated testing for Cholla soon
  • I had an inspiration to help my dotfiles system deal with the multiple login and compute nodes on various systems so I made some significant changes to them and ported those changes over to the public version

Original HLLD paper: Miyoshi & Kusano 2005

Blog post on the HLLD Algorithm: HLLD Algorithm

MHD, HLLD
Debugging HLLD Crusher Testing Continuous Integration
This post is licensed under CC BY 4.0 by the author.