[gmx-users] Scaling on SGI Altix 3700 Bx2

[Berk Hess]

>From: "Dallas B. Warren" <Dallas.Warren at vcp.monash.edu.au>
>Reply-To: Discussion list for GROMACS users <gmx-users at gromacs.org>
>To: Discussion list for GROMACS users <gmx-users at gromacs.org>
>Subject: [gmx-users] Scaling on SGI Altix 3700 Bx2
>Date: Wed, 29 Jun 2005 10:32:25 +1000
>
>I have just run some scaling of one of my systems on the new APAC SGI
>Altix 3700 Bx2 (http://nf.apac.edu.au/facilities/ac/hardware.php), 32
>CPUs on a single image node.
>
>For a 330,000 atom system using PME:
>
>CPUs		Scaling
>1		100%
>2		82%
>4		66%
>8		51%
>16		31%
>24		21%
>32		20%
>
>Is there anything that can be done on the software side of things to get
>it to scale better?  Other than not using PME and waiting for the new
>release which has rumours to have improved things here with PME.
>
>I attempted to use the sort and shuffle options to see if that would
>improve (found it did with cut-offs on another system but don't use the
>options as subsequent analysis is a pain) things, but it immediately
>crashes with triclinic box errors.

This bad scaling is completely due to PME.
The CVS code scales much better.

As for -shuffle and -sort, this should not be such a pain for analysis.
One only needs to run trjconv with the deshuf.ndx file to
obtain deshuffled trajectories (this functionality should of course
be built into mdrun in a future release).

Berk.