> -----Original Message----- > From: [EMAIL PROTECTED] > [mailto:[EMAIL PROTECTED] On Behalf Of coderman > Sent: Thursday, November 30, 2006 4:33 PM > To: theory and practice of decentralized computer networks > Subject: Re: [p2p-hackers] Hypercube topology > > On 11/30/06, Alex Pankratov <[EMAIL PROTECTED]> wrote: > > Does anyone know of or remember seeing any public work on practical > > side of building hypercube'd clusters ? > > the "Building Linux Clusters" book from O'Reilly uses hypercube > topologies for interconnection in the example deployments detailed: > http://www.oreilly.com/catalog/clusterlinux/ > > it's a bit dated at this point, but still informative.
Ok, thanks. I was hoping for more-or-less research/academic papers, but this is not a bad start I guess. > > Specifically, I'm curious about two things - dynamic restructuring > > of the cluster in the event of the connectivity loss between its > > nodes and the handling of non-2^n nodes case. > > the examples in the book used cross-over direct connections between > nodes to avoid switching latencies. > > if you need to support dynamic restructuring you're going to have to > throw a nice switch fabric (or a few of them) in there, and utilize a > much more complicated geometry to retain the benefits of hypercube > inter connectivity while avoiding bottlenecks at switching points. > > you may want to look at Myrinet if you need such switching > capabilities, though this somewhat obviates the need for a hypercube > geometry: > http://www.myri.com/myrinet/overview/ I should've mentioned that I was considering h-cube stuff in the context of overlayed networks. It seems that this can be a neat way to cluster "supernodes" ... From the first glance this should provide a good balance between the number of connections in the supernode cluster and its resiliency. _______________________________________________ p2p-hackers mailing list [email protected] http://lists.zooko.com/mailman/listinfo/p2p-hackers
