[gmx-users] NVIDIA GTX cards in Rackable servers, how do you do it ?

[Téletchéa Stéphane]

Le 24/02/2015 17:18, Szilárd Páll a écrit :
> Thanks! Let me note that those observations are particular to your
> machine. There are multiple factors that cumulatively affect the
> multi-threaded scaling:
> - physical vs "HT" threads
> - crossing socket boundaries
> - iteration time/data per thread
> - GPU and GPU performance
>
> In your case all these three factors are somewhat disadvantageous for
> good scaling. You have two sockets so your runs are crossing CPU
> socket boundaries. The input is quite small and with GPUs the
> HyperThreading disadvatages can increase - especially with a slow GPU.
>
> Also note:
> - your Quadro 4000 can likely not keep up with the 12 CPU cores and
> there is probably some "Wait GPU" time (see log file)
> - if you want to test 1 CPU + 1 GPU using HT vs not using it you
> should run make sure to run with "-pinstride 1 -ntomp 12" in the
> latter case!
> - "-nt" is partially deprecated/backward compatibility flag and should
> only be used if its meaning is "use this many tMPI or OpenMP threads
> and decide which one is better", which is not the case here!
>
> Cheers,
> Sz.

Dear Szilard,

Thanks for the informations, this was a rapid bench, but I have all the 
logs if needed.
I know this is bounded to my system and setup but if that can help 
others, I'd be happy in
extending my tests with required parameters and adding them to the wiki 
if needed.

Concerning the "Wait GPU" time, you are right, the numbers go from 8% to 
72.4% ...

Just let me know if you need more data and logs, I'd be happy to extend 
this benchmark to some
other computers available here with variable setups and hardware 
(including amd too),
to share on "real cases" what should be an "optimal" setting for 
performance/cpu best throughput.

Best,

Stéphane

-- 
Lecturer, UFIP, UMR 6286 CNRS, Team Protein Design In Silico
UFR Sciences et Techniques, 2, rue de la Houssinière, Bât. 25, 44322 Nantes cedex 03, France
Tél : +33 251 125 636 / Fax : +33 251 125 632
http://www.ufip.univ-nantes.fr/ - http://www.steletch.org