On Wed, Sep 28, 2016 at 3:40 PM, Kris Feher <kris.feher.yahoo.com> wrote:
> Dear All,
> I am looking for information on the scaling properties of calculations
> with the pmemd.mpi under AMBER 14 and also for MMPBSA.py.MPI for estimating
> needed resources in an application for computing time on a supercomputer
> centre. Could anyone show me to some data or give some estimates? The nodes
> have a maximum of 28 cores (Intel Xeon processors, Sandy bridge, E5-2670,
> 2.6GHz). With how many nodes is a simulation is still scaling acceptably?
> How about the scaling properties of the paralellised MMPBSA Ptython
> calculation?
>
​pmemd.MPI scaling will depend very heavily on interconnect. The answer
could be that the scalability is sufficient up to ~128-256 CPUs (maybe
beyond). Or if the bandwidth is limited and the interconnect has high
latency, you may struggle to see speed-ups with more than 1 node.
The benchmarks for Amber 10 should give you a decent idea:
http://ambermd.org/amber10.bench1.html. It's based on outdated hardware,
and faster CPUs with more cores per node mean the scalability will be worse
unless the bandwidth and latency have improved to match the hardware
improvements of the nodes themselves. I don't think the parallelization in
pmemd.MPI has changed much.
As for MMPBSA.py.MPI, the task is embarrassingly parallel, so as long as
you have enough frames to do several frames per core, you should see
speedup with as many CPUs as you can use. In practice, I think mpi4py will
limit you to maybe a few hundred cores. But you don't gain much by using
more than a few hundred frames, and MM/PBSA-like analyses are incredibly
cheap compared to the initial MD you need to do, so there's rarely a need
to use more than 1 node's worth of cores.
HTH,
Jason
--
Jason M. Swails
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Sep 29 2016 - 11:30:02 PDT
