[Bioclusters] Cluster specs

Joe Landman bioclusters@bioinformatics.org
Thu, 04 Nov 2004 13:11:05 -0500


(disclaimer: my company designs, builds, integrates, and supports such 

Mike Muratet wrote:

>The local university is preparing a proposal for a cluster and has asked
>for inputs. As is often the case, by the time the request trickles down to
>those who actually have to provide the informaton, the response is due.
>So, if I could impose upon the experience of those on the list, I need to
>refresh my knowledge about clusters and figure out what the state of the
>art is these days. I feel like one of those guys that is trying to get
>their CS homework done on the perl list, but here goes...
>* Platform: I like Red Hat or SUSE linux on Athlons (preferable to me) or
>Pentiums. I've done a lot of teeth-grinding trying to get things to run on
>Macs. But, I think there may be compelling performance arguments for
>Apple. Could anyone make such an argument? Anybody have any comments about
>Blade technologies?

Have a good hard look at Opterons.  Better performers by far on bio 
codes than Athlons, and on chem codes.  The 64 bit OS and builds have 
been shown to give (a free and) significant performance advantage over 
the pure 32 bit binaries.

>* Cluster system: the biocluster webpage talks about approaches for
>bioclusters being "...significantly different from
>traditional HPC and "beowulf-style" approaches" and I certainly buy into
>this. Has anybody thought any about _why_ it's different and would
>they like to share it? I've used Sun GridEngine. Has anybody come up

Short version:

Traditional beowulf uses lots of low latency interconnects and codes 
that depend upon them for high performance MPI.  Bioclusters do not 
generally run/require low latency MPI based programs.  Hence you have a 
traditional beowulf having an additional low latency fabric.

>with anything that has more synergy? Is there any open source BLAST out
>there that exploits the cluster environment?

Yes.  mpiBLAST (see http://mpiblast.lanl.gov and 
http://downloads.scalableinformatics.com/downloads/mpiblast/ for rpm 

There is code on this site to nicely run mpiBLAST through grid engine.


>Bioclusters maillist  -  Bioclusters@bioinformatics.org

Joseph Landman, Ph.D
Founder and CEO
Scalable Informatics LLC,
email: landman@scalableinformatics.com
web  : http://www.scalableinformatics.com
phone: +1 734 612 4615