From: Norman Geist (norman.geist_at_uni-greifswald.de)
Date: Wed Aug 20 2014 - 01:38:51 CDT
Hi,
the timings you get with 0.01 s/step doesn't seem to be very good. Before
you move to a higher benchmark pls try in namd script:
twoawayx yes
if the above helps also try in addition:
twoawayy yes
if that helps try in addition:
twoawayz yes
This options artificially increase the number of patches (box slices) so
that the large amount of cores a GPU has can be better utilized.
Also to the namd2 command try:
+ignoresharing
To come back to your original question, the stmv benchmark is quite large.
Mit freundlichen Grüßen
Norman Geist.
> -----Ursprüngliche Nachricht-----
> Von: owner-namd-l_at_ks.uiuc.edu [mailto:owner-namd-l_at_ks.uiuc.edu] Im
> Auftrag von Phil Greer
> Gesendet: Dienstag, 19. August 2014 18:31
> An: Maxime Boissonneault
> Cc: Namd Mailing List
> Betreff: Re: namd-l: "Beefier" benchmark
>
> Sometimes, the number of seconds/step is off if the run completes too
> quickly. This happened on a GPU system we were building in my
> department. Increasing the number of steps gives a bigger sample for
> computing seconds per step. In reality, that is just a fist pass test
> without finding a bigger system. If it is still claiming ~0.01s/step
> then a bigger system is needed.
>
> -Phil Greer
>
> On Aug 19, 2014, at 12:14 PM, Maxime Boissonneault
> <maxime.boissonneault_at_calculquebec.ca> wrote:
>
> > That will increase the total run time, but why would it change the
> number of seconds per step ?
> >
> > Maxime Boissonneault
> >
> > Le 2014-08-19 12:10, Phil Greer a écrit :
> >> In the apoa1.namd file edit the numsteps parameter to 10000 or
> higher.
> >>
> >> -Phil Greer
> >>
> >> On Aug 19, 2014, at 11:48 AM, Maxime Boissonneault
> <maxime.boissonneault_at_calculquebec.ca> wrote:
> >>
> >>> Hi,
> >>> I am running the apoa1 benchmark on our GPU cluster. Each node has
> 8 GPUs and 20 cores. I am compiling Namd with MPI and SMP support. It
> seems I reach a limit of about ~0.01s/step with one node, and cannot go
> quicker with more than one node.
> >>>
> >>> Is there a larger benchmark out there that would scale better on
> this type of system ?
> >>>
> >>> Thanks,
> >>>
> >>> --
> >>> ---------------------------------
> >>> Maxime Boissonneault
> >>> Analyste de calcul - Calcul Québec, Université Laval
> >>> Ph. D. en physique
> >>>
> >
> >
--- Diese E-Mail ist frei von Viren und Malware, denn der avast! Antivirus Schutz ist aktiv. http://www.avast.com
This archive was generated by hypermail 2.1.6 : Thu Dec 31 2015 - 23:21:08 CST