On Sat, 7 Nov 2009, Scott Ehrlich wrote:
I have a 10 blade cluster of just hardware - I can install what I
want, how I want. What options are there if I wanted to build the 10
blades as one large beast, but _NOT_ necessarily for someone doing
grid-type work? Some users don't now how to program that way, but
they'd like to have their program run on something that acts like a
single processor, but a massive single processor with lots of memory.
Some code is designed for a single, powerful workstation. I'd like
this cluster to act as one powerful workstation.
I have had okay experiences with ScaleMPs vSMP software. Depending on
workload it may or may not be good for your needs but it does work,
however, you'll also need to ensure that the nodes are connected via a
high speed interconnect like Infiniband. The vSMP system is based around
a COMA architecture. As such some memory is reserved for caching to hide
the latencies. We had issues with job affinities. Jobs would migrate
from CPU to CPU frequently leading to rather poor performance. All of
these findings were reported to them as we were an early adoptor.
While this can for for some of your jobs, I would advise that if you are
looking to the long term, consider taking advantage of MPI, OpenMP or
concurrent languages like Erlang. You'll be happier for it.
--
James A. Peltier
Systems Analyst (FASNet), VIVARIUM Technical Director
HPC Coordinator
Simon Fraser University - Burnaby Campus
Phone : 778-782-6573
Fax : 778-782-3045
E-Mail : jpeltier@xxxxxx
Website : http://www.fas.sfu.ca | http://vivarium.cs.sfu.ca
http://blogs.sfu.ca/people/jpeltier
MSN : subatomic_spam@xxxxxxxxxxx
Treat your password like your toothbrush. Don't let anybody
else use it, and get a new one every six months.
- Clifford Stoll_______________________________________________
CentOS mailing list
CentOS@xxxxxxxxxx
http://lists.centos.org/mailman/listinfo/centos
