[Bioclusters] Flat Network Neighborhoods

Indraneel Majumdar indraneel@indialine.org
Mon, 3 Sep 2001 08:51:40 -0500


I tried adding up the costs. It did not seem like a cost effective
approach. The cost for a large switch was always less than the cost for
the total number of smaller switches required. It would make sense for a
giant cluster though where you cannot get a switch so large (say 1024
nodes or something). I also tried to find a way of implementing a meshed
network (see Becker's paper on alternative network topologies) but a
person on the beowulf list pointed out that the queing software etc may
not support this (since a single host gets 2 IP addresses. older
gethostbyaddr() and gethostbyname() did not return the list of all
addresses). Implementation of a meshed network becomes more difficult
with the FNN than with a simple network topology with a larger switch
(in part due to the fact that you have only 3 or 4 PCI slots on a


On Fri, Aug 31, 2001 at 02:16:48PM -0400, jfreeman wrote:
> Hi Folks,
> I was talking to the folks yesterday's BAMCAT
> (http://xanadu.mgh.harvard.edu/bambct/) meeting and the cost of switches
> came up as a possible problem, so I remembered seeing the Flat Network
> Neighborhood (FNN) model used on the computer KLAT2.
> See:
> http://aggregate.org/KLAT2/
> In short they use cheap 8 port switches to connect large numbers of
> cluster computers and put two nic cards in each box to handle the
> routing.  They wire the network based on the output from a genetic
> algorithm which takes the number of nodes, the number of ports in the
> switch and the number of nics in the computer and gives out the
> switching fabric.
> I haven't tried it myself, we have big enough switch at the moment, but
> it seemed like an interesting approach.
> Good Luck,
> Jim
> _______________________________________________
> Bioclusters maillist  -  Bioclusters@bioinformatics.org
> http://bioinformatics.org/mailman/listinfo/bioclusters