status of fft5d_threads and fftw3_threads functionality
There are some stubs in CMakeLists.txt files and fft5d.c for some kind of FFTW threading functionality, but at least some part of it seems to have been discarded. Is retaining this useful? Should something be cleaned up?
Roland, if it's very easy to make this work again, I could benchmark it to see if it helps with GPUs. If it doesn't we should get rid of it.
As a large part of our user-base will be running with OpenMP-only parallelization (on most desktops/workstations and with single-GPU runs), if the multithreaded FFTW has any relevant performance advantage, it would be good to enable this feature.
https://gerrit.gromacs.org/#/c/1895/ is a patch to enable multi-threaded FFTW. Since it doesn't show significant speed-up we abandoned it. If someone wants to revisit it in the future this would be a good starting point. For now all remaining code can be removed if that makes it cleaner.
PS: I didn't benchmark with GPU. But given that I looked at the PME time by itself it shouldn't matter. I only benchmarked on a two socket 6 core Sandy Bridge. Should we benchmark any other hardware or is this sufficient?
OK, sounds like this stuff should be history. I'll remove it as part of my FFT upgrading. No need to make a branch or anything - one point of version control is that you can find old stuff with things like
git log -G'FFTW_THREADS'
Hrrrm I can't reassign it to me :-( Berk, please do.
- Assignee changed from Berk Hess to Mark Abraham
Mark: I added you to the manager list
I have some stuff sitting around in gerrit drafts await resolution of some other issues.
- Target version changed from 4.6 to future
- Target version changed from future to 4.6.x
- Project changed from Source code reorganization to GROMACS
- Category set to mdrun
- Status changed from New to Accepted
- Target version changed from 4.6.x to 5.x
Looks like this is still pending. Not likely to happen for 4.6 or 5.0.
- Target version deleted (
Also available in: Atom