Re: AMBER: parallel AMBER on cluster

From: Robert Duke <rduke.email.unc.edu>
Date: Thu, 15 Dec 2005 13:37:22 -0500

Kateryna -
I think that unless your hardware is either very bad, or your software is
configured very badly, you should do much better than what your sysadmin is
saying, especially for pmemd, but even sander scales better than indicated.
You should look at the various benchmarks posted at amber.scripps.edu.
Also, I have posted various pages about configuring linux systems on this
web site (helps with mpich settings, system buffer settings, etc.). Now,
what you do need is server-grade network interface cards, and if you have
your cluster interconnected with a switch it must also not be something that
is slow and cheap. For my main test setup, I run 2 dual 3.2 GHz pentium 4
machines with server network cards connected via a crossover cable (gets rid
of the switch entirely, but you are limited to 4 processors). I have run
mpich 1.252, 1.26, mpich 2, lam; all work without huge differences in
performance. For a really large system on these machines, pmemd 8 gives me
the following performance for a 90906 atom constant pressure pme simulation
benchmark, with results shown in wallclock seconds for 500 steps:

#proc wall clock secs
1 757
2 470
4 272

(these are values for pmemd 8 compiled with ifort 8, run on RHEL 3 in
October of this year; I see slight variations with
 compiler and OS versions).

The distributed fft transposes here kill you for systems of around 100,000
atoms, but still we are running about 3 times as fast on 4 processors as on
1. The scaling will be better for smaller problems in the 10K - 50K atom
range.

Mind you, if the sys admin does not get things right, you won't obtain this
performance. There is lots of dinking with various factors, including
hostfile setup. I would recommend reading amber.scripps.edu info carefully,
as well as the relevant mpich or whatever other mpi manual.

Regards - Bob Duke

----- Original Message -----
From: "Kateryna Miroshnychenko" <kateryna_mirosh.ire.kharkov.ua>
To: <amber.scripps.edu>
Sent: Thursday, December 15, 2005 1:16 PM
Subject: AMBER: parallel AMBER on cluster


> Dear Ambercommunity,
> We recently obtained an access to cluster and want to install our
> AMBER8 there. But system administator doesn't want to install
> AMBER on parallel queue. He told me that the improvement in the
> speed of calculations will be only about 5-10%, because the nodes
> in this cluster are connected by Gigabit network and this limits
> the performance of AMBER. I would like to khow whether this is true.
> Any opinions on this subject would be helpfull.
>
> With many thanks in advance,
> Katya
>
>
> ********************************************
> Kateryna Miroshnychenko
> post-graduate student
> Department of Biological Physics,
> Institute of Radiophysics and Electronics,
> National Academy of Sciences of Ukraine,
> 12, Proskura st., Kharkiv, 61085, Ukraine
> E-mail:kateryna_mirosh.ire.kharkov.ua
> ********************************************
>
> -----------------------------------------------------------------------
> The AMBER Mail Reflector
> To post, send mail to amber.scripps.edu
> To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
>


-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber.scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
Received on Thu Dec 15 2005 - 18:53:01 PST
Custom Search