Re: Re: Using the DHT to find download sources

vvpng wrote:
> Clustering will change the topology of the network. And we should be
> careful. Large scale simulation is needed before deploy if we are
> going to do this.
> 
> It's true that having a strong clustering may result in islands in the
> Gnutella network. And one may have trouble to find things its Gnutella
> clusters do not have. We do have the DHT, however, which will help us
> to "jump" to the clusters having the files that we cannot find in our
> clusters. In addition, we can always reserve some connections, say
> 5~10 out of 30, to random ultrapeers in the Gnutella network to
> improve connectivity between clusters. 

I'd originally imagined this as something that only leaves would do:
they'd connect to the same ultrapeers as their friends, or directly to
their ultrapeer friends, but connections between ultrapeers would still
be random. (In addition to its social connections, each leaf might
maintain one connection to a random ultrapeer to get a greater variety
of search results at the risk of receiving more spam.)

However, even if we extend the idea to connections between ultrapeers, a
relatively small number of random connections between clusters should
serve to keep the network connected and the diameter low, as in Duncan
Watts' "connected caveman" graphs. If, for example, each ultrapeer makes
half its connections socially and half randomly then the whole network
should still be reachable in a few hops, but the initial low-TTL parts
of the search will tend to return more results from the requester's
local cluster, which hopefully means more relevant content.

Cheers,
Michael