[gmx-users] gigabit scaling, DPPC benchmark

[Jason de Joannis]

> Date: Mon, 06 Jun 2005 19:47:17 +0200
> From: Christian Burisch <burisch at bph.ruhr-uni-bochum.de>
> Subject: Re: [gmx-users] gigabit scaling, DPPC benchmark
> To: Discussion list for GROMACS users <gmx-users at gromacs.org>
> Message-ID: <42A48C25.40403 at bph.ruhr-uni-bochum.de>
> Content-Type: text/plain; charset=us-ascii; format=flowed
>
> Diego Enry wrote:
>
> Hi everyone!
>
> > Jason, I don't know if that's the problem but when I use a P4EE 3Ghz,
> > Gromacs runs  faster without hyperthreading.
>
> Are you using kernel 2.6 (if you are using Linux at all...)? We are
> just now upgrading our cluster, and the effect of HT with kernel 2.6
> on Gromacs was one thing I wanted to benchmark. But if you - or anyone
> else reading this - have done so already I would be very glad if you
> could share the info!
>

Sorry, 2.4

> @Jason: Another way to decrease network latency - which is probably
> the limiting factor in your case - may be GAMMA (it was mentioned a
> long time ago on this list):
>

It looks like this is just a variation on MPI. It mentioned having a
separate interconnect for regular IP traffic. I wonder if that alone
might improve performance?

> http://www.disi.unige.it/project/gamma/
>
> We never tried it ourselves since we are using Myrinet in most of our
> nodes, but it sounds very interesting!
>
> Regards
>
> Christian
>
>