CCL: computer for high-level QC calculations

 Sent to CCL by: Justin Finnerty [justin.finnerty-,-uni-oldenburg.de]
 On Mon, 2008-06-02 at 10:58 +0200, Pablo Echenique echenique.p:gmail.com
 wrote:
 > Dear friends and colleagues,
 >
 > at my laboratory, we are about to buy a computer for performing
 > high-level QC calculations (e.g., coupled cluster) in biological
 > molecules (e.g., peptides). I write to kindly ask you for advice
 > regarding the "best" (in whatever sense you consider appropriate)
 > choice.
 My observations of coupled cluster calculations on our cluster are the
 following.  Most of the the QC codes that I know cannot run totally in
 memory, and often require lots of disk space. We find that local scratch
 disk is a must; we have a 2TB NFS filesystem and have tried parallel
 filesystems but both are very significantly inferior to local scratch
 for these jobs. We have one 16cpu machine with 64Gb RAM and 2.5Tb SCSI
 RAID storage and find that coupled cluster calculations performance
 "improvement" degrades when more than 8 cpus are used and also when
 more
 than one job runs on a node.
 If I was in your situation my recommendation as best value for money
 would be a base node with:
 4 or 8 cpus (eg 2 x Quadcore AMDs)
 2Gb RAM/CPU (eg 16 Gb RAM)
 1 system disk (only ~50 Gb is necessary)
 3 x 350+ Gb sata disks as local scratch (linux soft RAID, best not to
 include the system disk)
 Set up your queue system to run only one job per node.
 Unless you know that the problem sizes are always the same, I would buy
 some nodes with smaller and some with larger scratch disk systems.  This
 would allow you to cater to a wider range of job sizes possibly without
 extra cost.
 To upgrade the performance/capacity to this base system I would
 recommend first buying more nodes rather than upgrading nodes, then
 larger disks, then more RAM and lastly more CPUs/CPU speed.
 Ideally I would ask potential suppliers to provide a test machine and
 actually run test jobs of the size you are actually going to be doing.
 In particular monitor memory and disk activity.
 Cheers
 	Justin
 --
 Dr Justin Finnerty
 Rm W3-1-165         Ph 49 (441) 798 3726
 Carl von Ossietzky Universität Oldenburg