Re: [AMBER] parallel MMPBSA: no of cores versus no of frames analyzed

From: Jason Swails <>
Date: Mon, 1 Oct 2012 12:32:04 -0400

On Mon, Oct 1, 2012 at 12:10 PM, Vlad Cojocaru <> wrote:

> Dear AMBERs,
> We are experiencing some issues with the parallel version of MMPBSA. We
> use AMBER 12 on AMD Opterons (with 12 or 48 cores/node).
> In a test with 9 frames to process, we noticed that the parallel MMPBSA
> does not run on 12 cores (it hangs at the cpptraj step with no error).
> We actually realized that it runs on a number of cores equal or lower
> than the number of frames to analyze.
> From this test, we suspect that the number of frames to analyze has to
> be a multiple of the number of cores to be used in the parallel MMPBSA run.
> does not support using more processors than you have frames,
since it splits the workload by frame. Even though you could potentially
scale past nframes by assigning the complex and receptor of the same frames
to different cores (and this is before trying to call parallel external
programs), the gain for such an approach was not worth the cost of
programming. does handle non-multiple core counts, although it does so
naively. For a 10-frame analysis, for instance, 5 cores will run 2 frames
on each core. 6 cores will run 2 frames on 4 of the cores and just 1 frame
on the other 2 (so you see virtually no speed-up relative to 5 cores).
 Similarly with 9 cores, 1 frame will be run on each of 9 cores, and 2
frames will be run on the first core (so again, no speedup from 5 cores).
 However, it should work.

We could not find documentation regarding this issue so I would like to
> confirm whether this is a known issue of the parallel MMPBSA.
> Thank you for any clarifications on this issue,

Does stall in parallel when the number of cores is 9 or
fewer? Or is it only when nproc>nframes?


Jason M. Swails
Quantum Theory Project,
University of Florida
Ph.D. Candidate
AMBER mailing list
Received on Mon Oct 01 2012 - 10:00:03 PDT
Custom Search