[gmx-developers] Timing

hessb at mpip-mainz.mpg.de hessb at mpip-mainz.mpg.de
Mon Jan 19 08:53:37 CET 2009

> Hi,
> Is there a good reason to summing the time on all cores, when print the
> timing (function wallcycle_sum)? In case there is a load imbalance the
> maximum norm (multiplied with the number of cores) gives a much more
> reasonable time, because the other cores will wait adding practially to
> the
> time. In case their is no load imbalance changing the norm gives the same
> result.

In case there is communication that blocks all nodes at the end
of the code that is timed, you are right.
But is some cases there is communication at all
or there is local communication which causes certain imbalances
for one part to level out with another part.
The write_traj case is a clear case where a lot of time will
disappear into the "rest" timing.

> I noticed that in the case where a significant amount of time is spent in
> write_traj. Then the printed time is still very small, because only one
> core
> (rank=0) is spending time so that the sum is not much affected (for many
> cores).
> BTW: what is GMX_CYCLE_ALL doing?

Just try :)
There are small parts of code in between timed parts,
currently these end up in rest.
GMX_CYCLE_ALL will cause all those parts to be timed as well.


> Roland
> --
> ORNL/UT Center for Molecular Biophysics cmb.ornl.gov
> _______________________________________________
> gmx-developers mailing list
> gmx-developers at gromacs.org
> http://www.gromacs.org/mailman/listinfo/gmx-developers
> Please don't post (un)subscribe requests to the list. Use the
> www interface or send it to gmx-developers-request at gromacs.org.This email
> was Anti Virus checked by Astaro Security Gateway. http://www.astaro.com

More information about the gromacs.org_gmx-developers mailing list