large amount of serial code with p-coupling
The "rest" time with pressure coupling - especially with no DD and GPUs it can be 5-6x larger than without pcoupl; at the same time, the difference is only 2x with DD.
I suggest adding some multi-threading if possible.
Reduced cost of Berendsen pressure coupling
The coordinate and box scaling for Berendsen pressure coupling is now
done only every nstpcouple steps. It now also uses OpenMP threading.
Partially solves #1694
#5 Updated by Szilárd Páll over 5 years ago
#8 Updated by Berk Hess about 5 years ago
- Status changed from In Progress to Closed
The two merged changes significantly reduce the serial time with pressure coupling. The calcvir OpenMP parallelization in https://gerrit.gromacs.org/#/c/4272/ probably helps on Intel MIC, but gives mixed results on normal x86 CPUs. So we consider this issue closed.
We should continue to discuss which parts of https://gerrit.gromacs.org/#/c/4272/ should be merged (or modified).