Please try to avoid sending the same question multiple times to the
reflector.
> I have been running simulations using Amber 11 on a GPU Tesla C2050
> since a year. I would be running simulations with systems as large as
> 1,00,000 atoms.
Fantastic. Given this experience, you already know the relative
performance of GPUs and also presumably the clusters you have access to.
Given this knowledge, you should be able to answer these questions
yourself. If not, I would suggest looking at the benchmarks on the AMBER
WWW page. You should at least verify that what you were seeing (on your
GPUs / clusters) in the 1,00,000 atom simulations is consistent with what
is reported on the AMBER page.
> 1. If I set up a cluster with the following configurations, what would
> be the performance for a parallel run using Sander?
Use your experience and also look at the AMBER benchmarks. Performance
will depend on how you utilize the resources.
> 2. How many jobs can I run in parallel at a time without affecting the
> performance?
This depends on how you run your job. The description of the resources
you provide is confusing. As I read it, the configutation is:
1 master node. 12-core, 3 TB RAID. Infiniband.
2 compute nodes. Each 12-core. Infiniband.
So, you have one front-end node, and a back-end with two 12-core nodes (or
24 processors total). This will likely be ~1/2 the performance of one of
the GPUs you were running on previously. On this setup, you could likely
run two 12 processor jobs or one 24 processor job. If I had this setup, I
would runit as three nodes rather than dedicating a front-end/master.
If you are only buying three nodes, look at buying two and adding GPU
cards to each. If it was a typo on the number of nodes, and you are
actually buying 10's to 100's of nodes, look at the balanced cost of GPUs
vs. cores. In my experience with the current GPUs, they match ~48-60
cores.
--tec3
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Sat Mar 10 2012 - 10:30:02 PST
