Re: [AMBER] PNEB on GPUs

From: Ross Walker <ross.rosswalker.co.uk>
Date: Wed, 28 Sep 2011 21:01:50 -0700

Hi William,

> I am interested in purchasing a GPU workstation similar to the one
> described
> by Dr Ross Walker at (http://ambermd.org/gpus/index.htm#Hardware). I
> would
> like to begin doing some PNEB simulations to investigate the optimum
> pathway
> between two conformational states of small RNA stem loops. Will a
> workstation with 2x GTX 580 GPUs be able to run using the multisander
> functionality for a pathway of ~32-64 snapshots in a reasonable amount
> of
> time? If not, is this something that is being worked on?

What is meant by the P in PNEB? - AMBER supports Nudged Elastic Band
simulations right now but only on CPUs through the Sander package. While I
would ultimately like to port the NEB code to GPUs it is a way down the
priority list behind things like umbrella sampling restraints, Thermodynamic
integration etc.

Thus at the moment the best you could do is run NEB on CPUs. In this case
the machine I describe has only a single 4 core CPU. Hence with 64 snapshots
you'd be running 16 snapshots per CPU core so at best you would get speed
approximately 1/16th that of a normal single CPU MD simulation of the same
system. Unfortunately NEB simulations are extremely expensive and really
access to large scale clusters, 256+ cores to run effectively.

Your best bet being based in the US is to request a free allocation of
supercomputer time through the NSF XSEDE program (http://www.xsede.org/).
You can request a 12 month duration startup allocation for free with just a
NSF format CV and an abstract describing what you want to run. You then
select the machine you want to use. I would suggest SDSC's Trestles for this
(http://www.sdsc.edu/us/resources/trestles/) - This has 10,368 cores at 32
per node with QDR IB interconnect. You could probably run across 8 or 16
nodes (256 to 512 cores). The queue lets you run for up to 200 wallclock
hours at a time. The maximum startup allocation is 50,000 core hours on
Trestles which would allow you to run on 256 cores for 195 hours. Just over
8 days. If you need more time than this then you can submit a proposal to
one of the quarterly allocation meetings as described here:

https://www.xsede.org/allocations

On another note NEB might not be the optimal approach here. We have seen
issues with it getting stuck in high energy minima. You might have more luck
with an enhanced sampling method such as accelerated MD. We are currently
implementing this on GPUs and provisionally it looks to work well. We plan
to release this feature as part of AMBER 12 sometime in the spring.

All the best
Ross

/\
\/
|\oss Walker

---------------------------------------------------------
| Assistant Research Professor |
| San Diego Supercomputer Center |
| Adjunct Assistant Professor |
| Dept. of Chemistry and Biochemistry |
| University of California San Diego |
| NVIDIA Fellow |
| http://www.rosswalker.co.uk | http://www.wmd-lab.org/ |
| Tel: +1 858 822 0854 | EMail:- ross.rosswalker.co.uk |
---------------------------------------------------------

Note: Electronic Mail is not secure, has no guarantee of delivery, may not
be read every day, and should not be used for urgent or sensitive issues.



 


_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Sep 28 2011 - 21:30:02 PDT
Custom Search