Hi Kshatresh,
This is to be expected. Scaling of sander.MPI can be very poor and is
highly simulation specific. It is also extremely hardware specific - i.e.
it depends on how good your interconnect us, how many cores you have per
node etc. Typically a rule of thumb is don't expect sander to scale beyond
32 "cores" for most moderate sized proteins (50 to 100K atoms). If you
switch to pmemd.MPI you will see better speed and scaling but again you
can't assume that performance will always increase with node count.
Typically it will increase to a point and then fall off a cliff. If you
have AMD CPUs (or older Intel chips, pre sandy bridge) then the memory
bandwidth sucks so it can actually be beneficial to leave cores idle. E.g.
if you have 16 cores per node try running with just 8 cores per node.
Alternatively if you GPUs available you would be far better served using
them that some large scale cluster. See
http://ambermd.org/gpus/ for info.
There are benchmarks on there. For a point of reference for a 23,000 atom
simulation you see the following:
PMEMD.MPI
64 cores (4 nodes), QDR IB 2.6GHz Sandybridge E5-2670 = 58.19ns/day
(more cores than this slows down)
PMEMD.CUDA
1 x GTX680 = 72.49ns/day
1 x GTX-Titan = 111.02ns/day
Hope that helps/
All the best
Ross
On 5/12/13 9:48 PM, "Kshatresh Dutta Dubey" <kshatresh.gmail.com> wrote:
>Dear Users,
>
>I have installed amber 12 on a high performance super computer in users
>home. Serial and parallel installations were successful. But when I am
>running job using sander.MPI on 32 nodes, it runs very slow, even slower
>than a single core machine. Within 5 days it has completed only 200 ps
>simulations of 40,000 atoms.
> I am unable to figure the problem. Please suggest me whether it is
>installation problem or machine related problem.
>
>--
>With best regards
>**************************************************************************
>*****************************************
>Kshatresh Dutta Dubey, PhD
>Post Doctoral Fellow
>Indian Institute of Technology Kanpur
>Kanpur, India
>_______________________________________________
>AMBER mailing list
>AMBER.ambermd.org
>http://lists.ambermd.org/mailman/listinfo/amber
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Mon May 13 2013 - 03:30:02 PDT
