[Bioclusters] Setting up a small cluster

Joseph Landman bioclusters@bioinformatics.org
12 Jun 2003 02:13:12 -0400


Hi Malay:

  Given your applications and the clusters' focus as a learning/research
tool, I might suggest you have a look at http://www.rocksclusters.org,
or Callident's forthcoming BioBrew (ROCKS for bio-applications).e OSCAR
is somewhat non-trivial to install.  ROCKS is likely to be the simplest
installation of a functional cluster you can use.  If your machines can
run RedHat 7.3 you should have no problem with ROCKS.

  Alternatively, you could simply install RedHat, SUSE, or some other
distribution upon the cluster, and hand compile mpich and other tools. 
I typically recommend this, as the mpich options you want are invariably
not the ones built in to the binary.  I would recommend SGE as a
relatively simple install (and it is built into ROCKS).  You could use
LSF if you have the budget, or perhaps Condor.  Many cluster vendors
have been using RedHat 8.0 and more recently RedHat 9.0.  Due to ABI
(not API) changes in the core libraries, this may render some
pre-compiled programs non-functional between 8 and 9, and would require
a re-compile at minimum.  RedHat 8.0 isn't a great distribution to build
a cluster on, I know of many folks reloading clusters originally based
upon RH8 with RH7.3, or similar.  Library and stability issues made 8.0
a significant problem for a number of sites.  RH9.0 isn't old enough for
lots of cluster users to have experienced it as of this time.  Moreover,
changes in the end-of-life support model for these distributions may
have the side effect of driving cluster users to other supported
systems.

  One intriguing distribution I have not yet played with, which gives
you a functional OpenMosix (or so it claims) is clusterKnoppix.  You
might also look at Scyld, though it will cost you real money.

  OpenMosix will happily run in the above scenarios.  You would have to
patch NCBI BLAST to disable use of the POSIX threads to get it to
migrate (there is a patch on David States' group page at University of
Michigan, written by Carlos Santos, see 
http://stateslab.bioinformatics.med.umich.edu/software/OM_BLAST_patch/openmosixpatch.html)

  OpenMosix is more of a load balancer than a job scheduler, and might
require some additional tweaking to work well with MPI, PVM, and a job
scheduler.  The philosophies between a load balancer and a job scheduler
are somewhat different, as they are trying to solve different problems.

  Feel free to pull down my sge_mpiblast script
(http://scalableinformatics.com/downloads/sge_mpiblast) to run mpiBLAST
under the SGE MPICH environment.  When I get a chance to get LSF up on
my cluster, I will see if I can adapt it to LSF (shouldn't be hard).

  Good luck and feel free to post questions.

Joe

On Thu, 2003-06-12 at 00:42, Malay Kumar Basu wrote:

[...]

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