Hypercube routing

Skylar Thompson <(E-Mail Removed)> wrote:
> I'm an admin for a college CS department. We've been working on turning one
> of our regular Beowulf clusters into a cube cluster with eight machines.

I'm not really conversant on the structure of hypercube routing, so you
may need to alter this.

> [1] For others confronting this problem, the solution turned out to involve
> changing the broadcast address from {0,1,2}.255 to 255.255 and enabling
> proxy ARP.

This would reflect poor or incorrect network design wrt subnets. What
are the subnet masks of your machines? I would suggest, since you only
have two nodes on each subnet, that you give each link a /30, which
means that there are 2 possible host addresses (plus the network and
broadcast address.

$ ipsc -a 10.0.0.0/30 | head -30
[output trimmed]
Full subnet mask: 255.255.255.252
Hosts per subnet: 4

Subnet 1: 10.0.0.0 10.0.0.3 *
Subnet 2: 10.0.0.4 10.0.0.7
Subnet 3: 10.0.0.8 10.0.0.11
Subnet 4: 10.0.0.12 10.0.0.15
Subnet 5: 10.0.0.16 10.0.0.19
Subnet 6: 10.0.0.20 10.0.0.23
....

So for the two hosts in Subnet 1, you can give one the address 10.0.0.1,
and the other 10.0.0.2, with a netmask of 255.255.255.252. The broadcast
address (which will be important to check, you are using RIP after all),
will be 10.0.0.3

You should be able to just start routed on each node and be away
laughing (just remember that RIP has a largish failover/failback time,
but works quite well for a small network).


Another possible solution would be advertise a /32 gateway via RIP. I
think you could do this using /etc/gateways.

--
Cameron Kerr
(E-Mail Removed) : http://nzgeeks.org/cameron/
Empowered by Perl!

I'm an admin for a college CS department. We've been working on turning one
of our regular Beowulf clusters into a cube cluster with eight machines.
Each machine has three network interfaces, each of which is plugged into
one other machine. Each machine exists on three networks: 192.168.0.0,
192.168.1.0, and 192.168.2.0. After slaving away at the project for awhile, I
finally got machines more than one hop away from each other to ping[1].
With static routing, I can now configure the entire cluster as planned. I
am curious, however, as to whether I could setup dynamic routes so I can
minimize the impact of a single-machine failure on networking (I'm not
really worried about compute failure right now). I have played around with
routed and RIP, but I can't manage to get it to set the correct routes on
the correct interfaces. What other solution would people recommend other
than RIP?

[1] For others confronting this problem, the solution turned out to involve
changing the broadcast address from {0,1,2}.255 to 255.255 and enabling
proxy ARP.

--
-- Skylar Thompson ((E-Mail Removed))
-- http://www.cs.earlham.edu/~skylar/