Parallel

[UgoBruzadin]

Hi Yun hui,

I've recently acquired a new computer and tested running cudaica in parallel. Although when compared individually cudaica runs faster than binica, cudaica really slows down when I run it in parallel (say, 12 files at a time), to the point that the CPU can perform 3 or 4x faster than the GPU.

Do you have any suggestions on how to improve performance? I know cudaica isn't using 100% of my GPU's power, so I imagine there is some headroom and I was wondering what are your thoughts. I remember reading somewhere that there was a way to modify the parameters that cudaica uses on the GPU but I can't find it. Ideally one would use the combo of the GPU and CPU to achieve 100% use of power, but I can't imagine that would be easy (right now I have it randomly chosing GPU or CPU giving CPU preference 2/3s of the time).

Side note: I'm running cudaica.exe you made on an RTX3070 and it runs is flawlessly.

[CloudyDory]

Hi, sorry for the late reply. Currently I have never tried to run CUDAICA in parallel, because I have seen a StackExchange Post that says running multiple CUDA applications at the same time is not actually parallel.

"At the very same slice of time, only kernels from a single CUDA context may be executed on a GPU. This may cause a GPU underutilization if kernels do not occupy the entire GPU resources (memory + compute), and some of the resources may be left unused.".

There may be ways to fix this problem, but currently I really don't have time to learn and implement.

Best,

Yunhui

[winndsd]

The author of EEGLAB recommend a plugin named RELICA, which uses parallel pool and GPU at the same time (no CUDA).
It seems that ICA runs faster with PT than normal 'runica' algorim, using 6 cores, 80% CPU and GPU.
However, cudaica runs faster than both of them, while using 30-40% CPU and 30-60% GPU CUDA in my computer.
I believe there's still some potentials when combine PT with 'cudaica'.
Anyone can further optimize it?

RELICA can be downloaded from ‘https://github.com/sccn/relica’.