| commit | 82e5d9d99cb09f0e9bce2885c19d84848d9194cb | |
| author | Berk Hess <hess@kth.se> | |
| Mon, 23 Oct 2017 07:39:32 +0000 (23 09:39 +0200) | ||
| committer | Aleksei Iupinov <a.yupinov@gmail.com> | |
| Mon, 20 Nov 2017 06:45:36 +0000 (20 07:45 +0100) | ||
| tree | 3d638ca9918777bb5641090bc68cb781e1f87536 | treesnapshot (tar.gz zip) |
| parent | 029b967b6e9c162462b2a7dea2f0c25441ac7662 | commitdiff |
Prevent PME tuning excessive grid scaling
We limit the maximum grid scaling to a factor 1.8. This allows
plenty of room for shifting work from PME on CPU to short-range
GPU kernels, but avoids excessive scaling for diminishing return
in performance for a significant increase in power consumption,
communication volume (which may with fluctuating network load not
show up during tuning) as well as limiting load balancing.
Change-Id: I85c02478faa6b67c063b6e1b45a9ac1755b2d81e
We limit the maximum grid scaling to a factor 1.8. This allows
plenty of room for shifting work from PME on CPU to short-range
GPU kernels, but avoids excessive scaling for diminishing return
in performance for a significant increase in power consumption,
communication volume (which may with fluctuating network load not
show up during tuning) as well as limiting load balancing.
Change-Id: I85c02478faa6b67c063b6e1b45a9ac1755b2d81e
| src/gromacs/ewald/pme-load-balancing.cpp | diffblobblamehistory |