Re: [AMBER] Good cluster configuration for running PMEMD.MPI?

From: Matthew Zwier <mczwier.gmail.com>
Date: Tue, 17 Apr 2012 09:00:08 -0400

Hi,

Be sure to know what you're buying with the new 16-core AMD chips.
They have 16 integer units but only 8 FPUs, so they're really more
like 8-core CPUs with Hyper Threading, at least for the purposes of
floating-point-heavy applications like AMBER. (See
http://en.wikipedia.org/wiki/Bulldozer_(microarchitecture) ).

MZ


On Mon, Apr 16, 2012 at 5:37 PM, Jason Swails <jason.swails.gmail.com> wrote:
> On Mon, Apr 16, 2012 at 9:19 AM, Juan Carlos Muñoz García <
> juan.munioz.iiq.csic.es> wrote:
>
>> Hello,
>>
>> I'd like to ask you for your advise regarding a good/fast cluster for
>> running PMEMD.MPI of AMBER. Our group has the possibility of adquiring a
>> computer cluster with the following configuration:
>>
>> - 4 Processors Opteron Interlagos 16 Cores/each 6276 at 2,3 GHz, 8 Mb
>> cache L2 and 16 Mb cache L3, and 4 Hipertransport vers. 3 channels of 6,4
>> GT/s.
>> - Dual circuit board H8QGI+-F 3 SATA Hot Swap, VGA MATROX G200.
>> - 16Mb, 2x PCI-e 2.0 x16, 2 PCI-e x8, 2 network cards 1000 (Gigabit) IPMI
>> 2.0 (KVM over LAN).
>> - Memory: 128 GB DDR3/1600 ECC.
>>
>> According to what I've read in AMBER benchmarks site, it seems it's better
>> to use a higher number of processors/node but with a lower number of cores
>> per node.
>>
>
> This is highly system-dependent and depends on several factors.  For
> instance, you will get poor scaling between multiple nodes if your
> interconnect is slow (infiniband is best here).  Also, the more processors
> you have sharing the same memory bandwidth, the worse it will scale within
> a node that has a lot of processing cores.
>
> The benchmarks are examples on supercomputing facilities aimed to give a
> general idea of what you can expect with a few different hardware
> configurations.  However, the only way to determine the optimum scaling for
> your cluster is to run some benchmarks on your cluster.  The benchmark
> suite is available on ambermd.org (I think), so you can download them and
> run them to determine how your system performs.
>
> HTH,
> Jason
>
> --
> Jason M. Swails
> Quantum Theory Project,
> University of Florida
> Ph.D. Candidate
> 352-392-4032
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber

_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue Apr 17 2012 - 06:30:04 PDT
Custom Search