Re: NAMD CUDA performance

From: Vlastimil Zíma (zima_at_karlov.mff.cuni.cz)
Date: Wed Sep 11 2013 - 04:04:08 CDT

Hi guys,

thanks for help. I noticed I may have gone ahead of myself using too small
(~3 000 atoms) system for benchmark. I will do benchmarks with bigger
systems where it seems the benefit is bigger.

Regards,
Vlastik Z

2013/9/11 Norman Geist <norman.geist_at_uni-greifswald.de>

> Can you show us your namd script and your parallel namd call please?****
>
> ** **
>
> Norman Geist.****
>
> ** **
>
> *Von:* owner-namd-l_at_ks.uiuc.edu [mailto:owner-namd-l_at_ks.uiuc.edu] *Im
> Auftrag von *Aron Broom
> *Gesendet:* Dienstag, 10. September 2013 21:30
> *An:* Vlastimil Zíma
> *Cc:* namd-l_at_ks.uiuc.edu
> *Betreff:* Re: namd-l: NAMD CUDA performance****
>
> ** **
>
> I would suspect the bottleneck is memory bandwidth, since the GPUs must
> communicate with the CPU and adding more GPUs doesn't help this. I see you
> already mentioned a motherboard upgrade, but are you now using PCI 3x for
> the 690?****
>
> I would recommend testing some MD code that is entirely on the GPU and see
> if you get the kind of difference between the 590 and the 690 that you were
> expecting (AMBER or OpenMM).****
>
> Regardless, it may in the end still be that the best use of both GPUs
> within the 690 or 590 is to run multiple simulations (maybe in replica
> exchange if you want to help a particular simulation explore faster, but
> still have limited communication between GPUs).****
>
> ** **
>
> ~Aron****
>
> ** **
>
> On Tue, Sep 10, 2013 at 8:59 AM, Vlastimil Zíma <vlastimil_at_ziima.cz>
> wrote:****
>
> Hi,****
>
> I'm somewhat troubled with the performance of the NAMD, when run on CUDA.
> We have two graphic cards GTX 590 and GTX 690, but it seems that GTX690
> produces only minimal performance benefit.****
>
> Using both GPU cores of the GTX690 produces only about 8 % performance
> benefit, which is not a lot. Neither the upgrade of CPU+motherboard
> produced expected benefit in CUDA simulations (as compared to CPU only
> simulations).****
>
> I can not tell what is the occupation of the GPU's as nvidia does not
> provide any reasonable tool to investigate this on linux, but it seems that
> GPU's are underoccupied. Also I have not noticed a dependency on system
> size.****
>
> Is there any way to tell what is the bottleneck of the performance? I did
> found a note about charm projections, but it seems like a long term
> experiment.****
>
> If needed I can provide more details about my experiments with NAMD
> performance.****
>
> Vlastik Z****
>
>
>
>
> --
> Aron Broom M.Sc
> PhD Student
> Department of Chemistry
> University of Waterloo****
>

This archive was generated by hypermail 2.1.6 : Tue Dec 31 2013 - 23:23:42 CST