From: Axel Kohlmeyer (akohlmey_at_cmm.chem.upenn.edu)
Date: Thu Jul 23 2009 - 09:59:10 CDT
On Thu, 2009-07-23 at 10:43 -0400, Sadhu, Shubho (NIH/NCI) [F] wrote:
> Just to clarify, does this mean individual NAMD threads are competing
> for GPU resources if number of processors requested > number of GPUs?
in a way, yes.
since the GPU support in NAMD currently only affects the non-bonded
interactions, there is still work left that is CPU-only. depending on
the individual problem set size, having a CPU-to-GPU ratio of 2:1 would
allow much better utilization of the GPUs.
> For me NAMD runs faster as more processors are added (over the number
> of GPUs), but scaling is much worse.
i don't understand what you mean by that.
NAMD gives you a concise message about how many tasks you
are using and and which tasks are sharing what GPUs.
to make efficient use of GPUs is still depending on many
factors and there is still room for improvement in soft-
and harware/driver support.
cheers,
axel.
> Shubho
> ________________________________________
> From: owner-namd-l_at_ks.uiuc.edu [owner-namd-l_at_ks.uiuc.edu] On Behalf Of Axel Kohlmeyer [akohlmey_at_cmm.chem.upenn.edu]
> Sent: Thursday, July 23, 2009 9:04 AM
> To: David Chalmers
> Cc: namd-l_at_ks.uiuc.edu
> Subject: Re: namd-l: NAMD CUDA on dual Nvidia 295 GPUs
>
> On Thu, 2009-07-23 at 22:20 +1000, David Chalmers wrote:
>
> david,
>
>
> > Can I use all four GPUs?
>
> yes.
>
> > Should I be using one, four (or some other number) of cores?
>
> you need at least one cpu core per GPU. NAMD can oversubscribe
> GPUs and then you can hope to better utilize the individual GPUs
> by running something else on a cpu core, while another task is
> accessing the GPUs. the are some issues with this due to limitations
> of the nvidia drivers.
>
> > Which version of NAMD should I be using, multicore or the charmrun version?
>
> i've been able to use the charmrun version. the multi-core version often
> gave me segfaults (even without using GPUs).
>
> > Is there some way that I can better understand how many GPUs NAMD is using?
>
> there should be a message in the output telling you exactly
> how the individual GPUs are enumerated and bound to tasks.
>
> cheers,
> axel.
>
> > Thanks for any advice,
> >
> > David
> > ________________________________________________________________________
> >
> > David Chalmers
> > Faculty of Pharmacy, Monash University
> > 381 Royal Pde, Parkville, Vic 3053. Australia
> > ________________________________________________________________________
> >
>
> --
> =======================================================================
> Axel Kohlmeyer akohlmey_at_cmm.chem.upenn.edu http://www.cmm.upenn.edu
> Center for Molecular Modeling -- University of Pennsylvania
> Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
> tel: 1-215-898-1582, fax: 1-215-573-6233, office-tel: 1-215-898-5425
> =======================================================================
> If you make something idiot-proof, the universe creates a better idiot.
>
-- ======================================================================= Axel Kohlmeyer akohlmey_at_cmm.chem.upenn.edu http://www.cmm.upenn.edu Center for Molecular Modeling -- University of Pennsylvania Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323 tel: 1-215-898-1582, fax: 1-215-573-6233, office-tel: 1-215-898-5425 ======================================================================= If you make something idiot-proof, the universe creates a better idiot.
This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:53:04 CST