Project

General

Profile

Task #2253

reword or remove the CPU-GPU load imbalance note

Added by Szilárd Páll about 2 years ago. Updated almost 2 years ago.

Status:
Closed
Priority:
Normal
Assignee:
Category:
mdrun
Target version:
Difficulty:
simple
Close

Description

The notes need adjustment for PME-GPU.The imbalance note will become less meaningful with both nonbonded and PME offloaded (we expect most runs to be GPU-bound).

Additionally, the suggestion to tweak the cut-off in the GPU-bound nonbonded-offload only case should be ammended as inexperienced users may not be aware of the implications of doing that.

Associated revisions

Revision f8137a6e (diff)
Added by Berk Hess almost 2 years ago

Reword CPU/GPU imbalance notes

Changes text in CPU/GPU imbalance from "performance loss" to "wasting
resources", since in some cases one can not get higher performance.
Replaced "GPU has less load" by "CPU has more load".
Removed hint to reduce the cut-off, since one often can not do this.
Note that with CUDA all theses notes are never printed, since we no
longer have timings on (by default), unlike with OpenCL.

Fixes #2253

Change-Id: Ib4a9752ad27c1cd2a3cd751a217249694a56d3b7

History

#1 Updated by Berk Hess about 2 years ago

I would suggest to remove it.

#2 Updated by Szilárd Páll about 2 years ago

Berk Hess wrote:

I would suggest to remove it.

For all cases? Especially with only non-bonded offload it's still useful information, is it not?

#3 Updated by Szilárd Páll almost 2 years ago

  • Category set to mdrun
  • Assignee set to Szilárd Páll

#4 Updated by Mark Abraham almost 2 years ago

We should keep it if it is accurate and actionable. If we can't tell at run time that it would be faster to run a different task assignment, then we should keep quiet. Yes, there's a downside (a user missing an optimization opportunity) which competes with the downside that they might just ignore everything we say because we tell them about everything and they understand very little of it to know whether (and how) to act.

There's a large number of things that might possibly lead to a performance improvement if changed. For example, if the FEP kernels dominate performance, we could suggest compiling with icc.

#5 Updated by Mark Abraham almost 2 years ago

Are we going to do anything here?

#6 Updated by Berk Hess almost 2 years ago

With both NB and PME on GPU we only print the imbalance time and no further notes. Should we remove the imbalance print? That is still useful information, even if you can not act on it.

#7 Updated by Gerrit Code Review Bot almost 2 years ago

Gerrit received a related patchset '1' for Issue #2253.
Uploader: Berk Hess ()
Change-Id: gromacs~release-2018~Ib4a9752ad27c1cd2a3cd751a217249694a56d3b7
Gerrit URL: https://gerrit.gromacs.org/7449

#8 Updated by Berk Hess almost 2 years ago

  • Status changed from New to Fix uploaded
  • Assignee changed from Szilárd Páll to Berk Hess

#9 Updated by Mark Abraham almost 2 years ago

  • Status changed from Fix uploaded to Closed

Also available in: Atom PDF