search

SciML / StochasticDiffEq.jl

Solvers for stochastic differential equations which connect with the scientific machine learning (SciML) ecosystem
Other
248 stars 66 forks source link

Move to `juliaecosystem` #443

Closed staticfloat closed 2 years ago

staticfloat commented 2 years ago

We don't have GPU runners on the sandbox runners yet, but we may eventually get them. Until then, we can continue to use juliagpu.

ChrisRackauckas commented 2 years ago

Failed. This got a lot slower?

staticfloat commented 2 years ago

Hmmm, not sure what could be the reason, other than that this is a shared system, but looking into it, it doesn't seem to be overloaded. Are these benchmarks supposed to be highly parallel? It's possible that previously they were given all 128 processors to slam, and now they're only being given 16?

ChrisRackauckas commented 2 years ago

Yeah, those benchmarks are rather large for a very good reason. Can we not get the machine back? We got it for this purpose, so we should reprovision so we can do this again. 😅

staticfloat commented 2 years ago

@ChrisRackauckas actually, looking at the test history here, it looks like this is not the issue at all. The OOPWeakConvergence and IIPWeakConvergence tests have been failing for a long time, even when they were running on an exclusive machine.

staticfloat commented 2 years ago

And watching htop while the tests are running, it appears that things are running single-threaded. Is that expected?

ChrisRackauckas commented 2 years ago

Oh @frankschae this needs to get checked still.

frankschae commented 2 years ago

hmm, I'll have a look -- I expected the tests to be multi-threaded, because they should use EnsembleThreads(), see https://github.com/SciML/DiffEqDevTools.jl/blob/37fde034a03b8dcf440613a81df1483a06ea25f9/src/convergence.jl#L43 .