Re: [AMBER] Number of cores for a parallel job

From: David A Case <>
Date: Tue, 28 Jan 2014 07:55:18 -0500

On Tue, Jan 28, 2014, FyD wrote:
> From tests carried out in the past on SMP hardware (with shared
> memory) it looks like sander better runs on 2^n core systems; i.e.
> with 2, 4, 8, 16, 32, 64... cores. This means the scaling is not good
> for instance with 24 cores. Is it correct? On the contrary, on the
> same SMP systems it looks like Gaussian 09 scales quite well on 16 or
> 24 cores.

You are correct about sander working best with 2^n cores, although when n>5,
sander generally will not scale well under any circumstances. The pmemd code
does not have this limitation, and is always faster. (Sander uses a simple
tree code to communicate coordinates and forces, and this code is optimal for
powers of two).

"Psander" was the name of an experimental code that preceded pmemd. I had
forgotten that it was ever even distributed.

I can't help with info about gromacs....maybe you can just run some
experiments yourself....


AMBER mailing list
Received on Tue Jan 28 2014 - 05:00:03 PST
Custom Search