Dan wrote:
> Hello,
>
> My research group is considering building a linux cluster to replace our
> alpha-based workstation. The question is regarding serial jobs.
>
> It seems that many people use beowulf-type clusters to do massive parallel
> computing, but our research simulations cannot be written parallel, in
> general. (We do Markov chain simulations which have to be linear.) The
> question is, then, will a beowulf-type cluster properly handle a queue of
> serial jobs? Ideally, we would log in to the master node, compile the jobs,
> execute the jobs, and then let the scheduler send one serial job to each
> node processor. Or, if a beowulf-type cluster is not the right type of
> system, what cluster architecture would be appropriate for running
> exclusively serial jobs?
>
> Thanks,
>
> Dan
>
My 256-CPU beowolf cluster is running a mix of serial and parallel jobs.
Around 20% of the jobs are single-cpu and another 20% are two-processor
(in the same box). The rest are mostly 4 or 8 processor jobs, declining
down to a very small number of >= 128 processor jobs. Running the cluster
as a farm of single cpus would actually simplify the running of the
beowolf environment. If you ever want to start running mpi programs it
might be worth it to start with beowolf, otherwise take a look at OpenMosix
(
http://openmosix.sourceforge.net). Beowolf is also quite mature and does
not require kernel patches. Note that I have looked into openmosix but
have not actually run such a cluster.
--
Dr. Douglas O'Neal
Manager, Bioinformatics Center
Delaware Biotechnology Institute
(302) 831-3456