From adam@cypherspace.org Thu Mar 7 15:37:35 2002 From: adam@cypherspace.org (Adam Back) Date: Thu, 7 Mar 2002 15:37:35 +0000 Subject: (fwd) Hypernets - Good (G)news for Gnutella In-Reply-To: ; from oskar@freenetproject.org on Wed, Feb 20, 2002 at 05:40:31PM +0100 References: Message-ID: <20020307153735.A1067072@exeter.ac.uk> Perhaps we should email the author and ask? A few thoughts: perhaps he is assuming a more efficient and scalable way to do a broadcast amongst a chosen number of fraction of the network. If the subset broadcast were random (say 100 neighbours in some hypercube routing algorithm), it could scale in the sense that overhead doesn't increase as more nodes join, though it remains to be seen how efficient it would be. The gnutella network in contrast broadcasts 7 hops deep into the network I believe? (Or was that Freenet). The network topology is more randomised and loops may occur (?), so duplicate squashing is employed to squash originated or already seen search requests, I speculate the topologically close subset broadcast algorithms would be less efficient with more randomly organized networks like gnutella. To reach the number of hosts which happen to be reachable via a 7 hop deep flood fill of gnutellas typical structure could be more efficiently achieved in a hypercube. Fewer hops would be employed leading to lower global bandwidth consumption and lower latency. Scale in the above sense doesn't make any guarantees about being able to do a complete search of the network. I think it may help for nodes to pre-distribute their own content indexes globally with random cache replacement up to the reserved content index cache size of each machine. Adam On Wed, Feb 20, 2002 at 05:40:31PM +0100, Oskar Sandberg wrote: > > I don't understand how they can go from discussing the graph diameter to > saying anything about the scalability of a Gnutella network. One could > have a completely connected network (ie graph diameter 1) and Gnutella > would still not "scale" as long as the only way of finding out whether > node contains a resource is by asking it. > > Nice characteristics of the graph give nothing without a way of routing > and sorting within it (of course, for binary identifiers, one can route > in a hypercube using the Gray paths, but that is old news and just a > special case of what Plaxton, Tapestry, Pastry etc do.) > > On Wed, Feb 20, 2002 at 01:00:17PM +0000, Adam Back wrote: > > Paper about scalability of distributed search networks arguing that > > gnutella's scalability limitations come from it's use of tree > > structure in the topology of the self organising network, rather than > > any inherent non-scalability of the application. They propose and > > analyse the use of higher dimension networks such as hypercubes. > > > > Adam > > -- > > Oskar Sandberg > oskar@freenetproject.org > > --- > You are currently subscribed to bluesky as: adam@cypherspace.org > For list information visit http://www.transarc.ibm.com/~ota/bluesky/ From gohnesorge@lh-computertechnik.de Sun Mar 17 06:13:20 2002 From: gohnesorge@lh-computertechnik.de (g'o'tz ohnesorge) Date: Sun, 17 Mar 2002 07:13:20 +0100 Subject: t3st References: Message-ID: <3C943400.11D58DEF@lh-computertechnik.de> .. sorry .. :)