reliable network for cluster

[Bob Drzyzgula]

Don't know how others do it, but one way is to assign
the canonical IP address for a host to an internal
loopback-type virtual interface, possibly lo itself. This
IP address should appear to be on a separate IP network
from the public production network (this means that you'll
need to have to assign a separate IP network for every
host). Assign each of the physical, external interfaces IP
numbers on the main network, and enable ip routing to
the internal interface. Finally, run gated and the OSPF
routing information protocol, assigning equal costs to
the routes from all the interfaces to the internal net.
You'll wind up with reasonably gnarly routing tables, but
OSPF should be able to handle it and automatically load
balance among the interfaces, mostly by giving a different
answer every time a remote host looks for a route to the
internal net.

Sorry, I'm quite certain this could be explained better...

--Bob

On Wed, Jul 04, 2001 at 08:20:11PM -0400, Brian Beuning wrote:
> Many people on this list seem to be interested in the comptational
> power of clusters.  If their network goes down, they might just restart
> the problem they are crunching on.
> 
> For some of us, the beauty of clusters is their reliability.  If one
> node
> fails, the rest of the nodes can continue processing.  Of course, a
> cluster
> is only as reliable as the network that connects the nodes.  Channel
> Bonding seems great for getting more bandwidth but only helps with half
> of the network reliability issue.  (It lets a node send packets out in a
> broken
> network, but does not necessarily let clients get to the box using
> Channel
> Bonding.)
> 
> What I would like is a way to have multiple NIC cards in a node have
> the same IP address, each NIC is connected to a different switch/hub
> and have the routing figure out which paths are up and down.  It should
> also do load balancing (aka scale the network).  Since our clusters all
> use commodity hardware, paying for some esoteric network routers is
> not part of the plan.
> 
> How do you folks solve this issue?
> 
> Thanks,
> Brian Beuning
> 
> 
> 
> 
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf