Basilisk runtime and profiling 🕐

[atharris-work]

Howdy BSK community,

I'm interested in experiences people have improving the runtime of Basilisk sims, particularly if those people have managed to profile complex BSK sims with open-source profilers like Cprofile. Basilisk is definitely known for its fast performance, but I've also found edge cases in which performance can really suffer if you aren't careful. I would love to hear/share some tips and tricks surrounding BSK runtime to help folks eke out as many 'years in a day' as they can 😎.

Some initial thoughts to kick things off:

• Pulling data out of sims, especially via recorders, can take as long as the actual simulation depending on the number of variables recorded and the sim length.
• Some modules seem to have disproportionate runtime impacts, like groundLocation instances.
• Some things I would expect to be expensive seem to have marginal runtime impacts, like increasing the order of gravitational coefficients loaded in useSphericalHarmonicsGravityModel() (to a point...)
• Cprofile + snakeviz is a great open source toolset for understanding the runtime of Basilisk alongside other Python analysis code, but it struggles to provide detailed runtime information beyond sim_model.StepUntilStop.

[Mark2000]

My current stats are around 10-15 years of sim time in 24 hours across 32 cores when doing RL. Some fairly complex bookkeeping going on wrapping the basilisk sim, but ultimately the bsksim running function is 70+% of the time. Can get about twice as fast in simpler environments.