pyLossless on HPC (Digital Research Alliance of Canada )

[scott-huberty]

I have EEG files that will run through pylossless in < 30 minutes each on my local computer (16 GB desktop), where for example, FASTICA on average finishes in 3 minutes.

These same files on the DRAC HPC (I am using Graham), run just as fast through the pipeline methods UP to ICA. But ICA is a total bottleneck.

On the HPC (Graham), on average my files get to the first ICA (FASTICA) step within a few minutes of pylossless starting. FASTICA won't finish by the time the SLURM job times out (2-hour job, 16 GB memory - which should be way over kill...).

@Andesha I know you battled with this with your files too. I remember you describing that mne / sklearn use jobllib in the backend, which doesn't recognize the available CPUs etc in the HPC environment.

It would be great to be able to run pylossless on the HPC clusters. Do you have any suggestions?

[scott-huberty]

In summary: FastICA takes 3 minutes on my computer. On the HPC, FastICA won't finish, even after running for hours (record job length so far: 8 hours...).

[christian-oreilly]

Could it be due to how the packages are built (e.g., NumPy)? Some numerical routines can be compiled in different ways (e.g., BLAS and LAPACK) with a significant impact on processing time if I remember well...

[Andesha]

I’m away for a bit but you’re exactly right. This was part of what I ran into. Hopefully Monday I can hand you my recipe for getting things installed properly and it’ll run no problem.

…

On Sat, May 27, 2023 at 8:08 AM christian-oreilly ***@***.***> wrote: Could it be due to how the packages are built (e.g., NumPy)? Some numerical routines can be compiled in different ways (e.g., BLAS and LAPACK) with a significant impact on processing time if I remember well... — Reply to this email directly, view it on GitHub <#126 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACN2HO7AF33ZSOXZ6G6YJVTXIHVFDANCNFSM6AAAAAAYPLQPTQ> . You are receiving this because you were mentioned.Message ID: ***@***.***>