Re: [AMBER] pmemd with mpi number of cpus

From: Ross Walker <>
Date: Wed, 15 Dec 2010 11:59:54 -0800

Hi Stefano,

Just to expand upon this a little. There is no actual power of two limit on
the number of threads with sander. It is simply a performance issue. When
you have a power of 2 CPUs sander uses a binary tree for the reduce scatter.
If you have a non-power of 2 cpus then it just calls MPI_REDUCE_SCATTER. At
a very basic level this means that using 8 threads can often be faster than
using 12 and using 16 might be more efficient than using 24. However, this
comes with the BIG caveat that it depends very much on your hardware, your
MPI installation AND the simulation parameters. In practice this means that
you should probably just test it for the setup you have. Try running it on 8
threads or 12 threads and see what the performance difference is. With AMBER
11 you can get a very good estimate very quickly by just looking at the
mdinfo file.

PMEMD on the other hand does its MPI communication via point to point
operations and thus the communication pattern does not change if you have 8
or 12 threads. Hence there is not a distinct benefit of using a power of 2
threads with the CPU PMEMD code.

All the best

> -----Original Message-----
> From: case []
> Sent: Wednesday, December 15, 2010 6:22 AM
> To: AMBER Mailing List
> Subject: Re: [AMBER] pmemd with mpi number of cpus
> On Wed, Dec 15, 2010, Stefano Pieraccini wrote:
> >
> > I have until now used the parallel version of pmemd employng a number
> of
> > processor that was a power of two (In the older version of sander there
> was
> > this restrictions and I went on like this). Now I'm using a cluster with
> > cpus on each node, and it would be useful for me to use a number of
> > processors multiple of 12. Is the restriction of the power of 2 still
> > or anyway does it still assures better performance, or can I safely use
> > arbitrary number of cpu?
> pmemd (CPU version) has no power-of-two limitations, and I have used it
> quite
> successfully on 12 cores.
> ...dac
> _______________________________________________
> AMBER mailing list

AMBER mailing list
Received on Wed Dec 15 2010 - 12:30:03 PST
Custom Search