Hi Dean,
> The place I'm running sander.MPI limits np to 48.
>
> I'm seeing a big performance hit with np 48 compared to np 32.
>
> I don't see this performance hit with pmemd. Is this normal?
>
> I'm doing SGLD, so I'm using sander.MPI
Yes this normal. Sander uses an all reduce to sum the forces on all the
nodes. If you have a power of 2 mpi threads it uses a binary tree to do
this. If you do not have a power of 2 then it switches to a less efficient
algorithm. PMEMD on the other hand has a spatially decomposed data
architecture so communications are point to point rather than all reduces /
broadcasts. Hence the algorithm used for communication does not require a
power of 2 cpus to be efficient. I would suggest not using more than 32
threads with sander. You could run two jobs at once, one of 32 cpus and one
of 16.
If however, your job is supported by the PMEMD feature set then you should
ALWAYS use PMEMD since you can use all 48 cpus and will get much better
performance.
All the best
Ross
/\
\/
|\oss Walker
| Assistant Research Professor |
| San Diego Supercomputer Center |
| Tel: +1 858 822 0854 | EMail:- ross.rosswalker.co.uk |
|
http://www.rosswalker.co.uk |
http://www.wmd-lab.org/ |
Note: Electronic Mail is not secure, has no guarantee of delivery, may not
be read every day, and should not be used for urgent or sensitive issues.
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Fri Nov 06 2009 - 18:00:02 PST
