[Beowulf] One network, or two?

Tony Travis ajt at rri.sari.ac.uk
Tue Sep 23 13:22:53 PDT 2008


Prentice Bisbal wrote:
> [...]
> My new cluster, which is still in labor, will have InfiniBand for MPI,
> and we have 10 Gb ethernet switches for management/NFS, etc. The nodes
> only have 1 Gb ethernet, so it will be effectively a 1 Gb network.
> 
> I'm also curious as to whether the dual networks are overkill, and if
> using a slower network for I/O will cause the system to be slower than
> doing all traffic over IB, since I/O will be slower and cause the nodes
> to wait longer for these ops to finish.

Hello, Prentice and Alan.

I've built a Beowulf based on EPCC BOBCAT:

	http://bioinformatics.rri.sari.ac.uk/bobcat/

What attracted me to BOBCAT is that it used two completely separate 
network fabrics: One for 'system' and one for 'applications'. Before I 
used this approach, it was very easy to lose control of the Beowulf 
cluster because it's easy to saturate the 'application' IPC, but if you 
have completely separate network fabrics, you can still control the 
Beowulf even when the 'application' network is saturated. This works 
extremely well in practice on our openMosix Beowulf with 88 PXE-booted 
nodes using NFSROOT over 100Mb and 'application' IPC over Gb ethernet.

	Tony.
-- 
Dr. A.J.Travis, University of Aberdeen, Rowett Institute of Nutrition
and Health, Greenburn Road, Bucksburn, Aberdeen AB21 9SB, Scotland, UK
tel +44(0)1224 712751, fax +44(0)1224 716687, http://www.rowett.ac.uk
mailto:ajt at rri.sari.ac.uk, http://bioinformatics.rri.sari.ac.uk/~ajt



More information about the Beowulf mailing list