Hi Fatima,
> what combination of CPU + GPU one should have for a better parallel
> scaling of PMEMD.cuda in Amber 11
>
> 1. one node with 4 GPU cards;
> 2. one node with 2 GPU cards;
> 3. one node with 1 GPU card.
>
> Do you have any opinions on any of the above or any permutations?
It depends what you want to do. If you want to run 4 individual calculations
at once them 4 GPUs in a single box will be fine, tests to date show the
performance is unaffected by running multiple calculations on multiple GPUs
in one node as long as you have enough CPU cores.
If you want to run in parallel - using the soon to be released MPI version
of the GPU code then you will get best performance with 1 GPU per node but
you MUST connect the nodes together with QDR infiniband. I don't have exact
numbers yet for 4 nodes x 1 GPU per node vs 2 nodes x 2 GPU per node vs 1
node x 4 GPUs per node as we are still tuning the code and I also don't
readily have enough variations in hardware available to test all the
combinations. However, I hope to add some info to the website shortly.
Right now though for single GPU calculations you should see that 4 nodes x 1
GPU === 1 node x 4 GPUs.
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 |
|
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 Tue Aug 17 2010 - 18:00:03 PDT
