PVFS vs. file servers
Josip Loncaric
buying a big file server vs.
using PVFS (see http://parlweb.parl.clemson.edu/pvfs/)
on his cluster. He has a cluster of Alphas running Linux, and each node
already has about 10GB of disk space available (>100GB total). Buying a
big file server would add more disk space (ignoring the space already
available), but it would allow his nodes to be dedicated to computing
free of file serving tasks.
He runs big CFD codes and needs lots of disk space to store results and
input files.
Any comments on pros and cons?
Sincerely,
Josip
--
Dr. Josip Loncaric, Senior Staff Scientist mailto:jo...@icase.edu
ICASE, Mail Stop 132C PGP key at http://www.icase.edu./~josip/
NASA Langley Research Center mailto:j.lon...@larc.nasa.gov
Hampton, VA 23681-2199, USA Tel. +1 757 864-2192 Fax +1 757 864-6134
_______________________________________________
Beowulf mailing list
Beo...@beowulf.org
http://www.beowulf.org/mailman/listinfo/beowulf
Brian Haymore
testing purposes). We have had our system up in production use now for
nearly 3 months without any major issues. The performance is quite
good. We have a single meta server and 4 io nodes. We use this space
as a global scratch space and an NFS server for home directory space.
We also use local disk space on the compute nodes for another level of
scratch. Kind of like thinking of local node scratch as "L1 scratch"
and global PVFS space as "L2 scratch" and then NFS for the final
repository. PVFS scales quite well and is easily expanded down the
road. The cons would be that the system was not designed to have a lot
of redunancy in it. This for my site isn't really a con at all as we
have selected it for use as a global scratch, which doesn't require much
if any redundance. After all it is scratch space.... A pro would be
that you can use very cheap hardware or current existing hardware to try
it out before you opt to push it for production use. You could easily
move a few of your compute nodes out of the loop for a bit and set them
up as a pvfs system using 3 or 5 nodes (1 meta, 2 or 4 io nodes). This
is similar to how we got started.....
--
Brian D. Haymore
University of Utah
Center for High Performance Computing
155 South 1452 East RM 405
Salt Lake City, Ut 84112-0190
Email: br...@chpc.utah.edu - Phone: (801) 585-1755 - Fax: (801) 585-5366
Greg Lindahl
> input files.
You didn't say anything about speed of access. It should be *cheap* to buy a
x86 box and add a couple of big IDE drives to it. That gives him more disk
space than the 100 GB on his various internal disks, doesn't cost much, and
it won't exactly be fast, but it may well be fast enough for him.
-- g
Joel Jaeggli
On Wed, 20 Sep 2000, Greg Lindahl wrote:
> > He runs big CFD codes and needs lots of disk space to store results and
> > input files.
>
> You didn't say anything about speed of access. It should be *cheap* to buy a
> x86 box and add a couple of big IDE drives to it. That gives him more disk
> space than the 100 GB on his various internal disks, doesn't cost much, and
> it won't exactly be fast, but it may well be fast enough for him.
We've gotten as much as 35MB/s out of ibm 75gxp ide disks in local
testing. across a fast ethernet switch we've seen as much as 8.5MB/s which
isn't to shabby for an ide based fileserver...
joelja
> -- g
>
>
> _______________________________________________
> Beowulf mailing list
> Beo...@beowulf.org
> http://www.beowulf.org/mailman/listinfo/beowulf
>
--
--------------------------------------------------------------------------
Joel Jaeggli joe...@darkwing.uoregon.edu
Academic User Services con...@gladstone.uoregon.edu
PGP Key Fingerprint: 1DE9 8FCA 51FB 4195 B42A 9C32 A30D 121E
--------------------------------------------------------------------------
It is clear that the arm of criticism cannot replace the criticism of
arms. Karl Marx -- Introduction to the critique of Hegel's Philosophy of
the right, 1843.