Particle System Tests
This week I planned to add support for particles to the system test function and I got most of the way there before running into some snags. I was initially using disk initial conditions to test with, however that only has particles in it, no hydro fields, so I didn’t realize that when run with both hydro and particles Cholla produces two output files per time step instead of just one. I’m working on adding that support, and it should be done, but I ran into some issues with how the particles part of Cholla generates random numbers.
Random Number Generator Hell
Cholla generates some particle properties randomly and this of course makes it non-deterministic and almost impossible to test. To enable us to perform deterministic tests I set up a new input parameter to cholla,
prng_seed, which sets the seed for the pseudo random number generator (PRNG). While I was doing this I improved the quality of the default random seed generation. Previously it just used
std::random_device isn’t required to be actually random by the C++ standard and so isn’t a reliable source of randomness, plus it only provides 32 bits of randomness and our generator,
std::mt19937_64, can take up to 64 bits. I instead now convert the result of
std::random_device to a string, then concatenate the string version of the MPI process ID and the
std::chrono::high_resolution_clock time, then use
std::hash to hash the resultant string. This should provide a robust source of 64 bit randomness across machines, time, and for every MPI rank, even if
std::random_device just returns a constant or
std::chrono::high_resolution_clock returns the same time for multiple process.
At the last minute this week I discovered that there are two different pseudo random number generators used in Cholla. I’ve asked the people involved if I can combine them into one but I haven’t heard back from everyone yet.
I noticed that the GitHub Actions badge in Cholla isn’t showing the proper status. I couldn’t figure out why so I contacted GitHub support but haven’t head back. I think the issue is that all the current builds are triggered by PRs from forks of Cholla. If that’s the issue, or if I don’t hear back from GitHub support, I’ll just set up a weekly or nightly build.
- Found a bug in how particle IDs were assigned that resulted in duplication
- Wrote a python script to concatenate the particle and field data sets to facilitate making fiducial data files.
- Created a particle based system test
- Refactored system tests as parameterized tests and added support for them in