Recommended parallelization for a2f/phonselfen calculations?
Posted: Fri May 31, 2019 1:49 pm
Hello,
In this post I was wondering what the recommended way to run the a2f/phonselfen calculation in EPW is. I've been using 4 nodes, 36 processors each (144 procs total), and so of course am using -npools 144.
However, the (k, k+q) calculations seem to be going somewhat slowly given this amount of parallelization. Trying --nk 12 --nt 12 doesn't seem to help much. It's fine for grids that end up having only 2000 (k, k+q) pairs, but it appears having 20x20x20 k/q fine grids produces 8000 points. Furthermore, 30x30x30 points (or more) might be needed for convergence. Only about 1000 points are calculated in 20 hours, which seems slow to me. I thought the coarse 10x10x10 phonon grid calculation would take the longest, but it appears this part is!
I can share the input files, etc. It's a simple system, Niobium, but I seem to need fine meshes for convergence nevertheless - the a2f magnitude is too high for example.
In this post I was wondering what the recommended way to run the a2f/phonselfen calculation in EPW is. I've been using 4 nodes, 36 processors each (144 procs total), and so of course am using -npools 144.
However, the (k, k+q) calculations seem to be going somewhat slowly given this amount of parallelization. Trying --nk 12 --nt 12 doesn't seem to help much. It's fine for grids that end up having only 2000 (k, k+q) pairs, but it appears having 20x20x20 k/q fine grids produces 8000 points. Furthermore, 30x30x30 points (or more) might be needed for convergence. Only about 1000 points are calculated in 20 hours, which seems slow to me. I thought the coarse 10x10x10 phonon grid calculation would take the longest, but it appears this part is!
I can share the input files, etc. It's a simple system, Niobium, but I seem to need fine meshes for convergence nevertheless - the a2f magnitude is too high for example.