david at aminal.com <david at aminal.com> wrote:
> What we have is a routing problem. Not a caching problem. The mesh is
> getting 'lumpy', a few nodes becoming responsible for a large portion
> of keyspace. I've seen it in my data store a couple of times now, refs
> disappearing until most point at one machine. In a situation like that,
> datastores will turn over very fast, with much of the traffic moving
> through a minority of nodes.
> 
> Hypothesis:
> 
> Any node that starts with an initial advantage, will gradually attract
> references to ever greater portions of keyspace, at the expense of other
> references.
> 
> Don't know how to test this, though, other than to continue to watch my
> datastore.

Well, setting the closest-key comparisons aside for a moment, a random walk
through a power-law network will tend to gravitate to the nodes that have
the most references, giving them more references -- "the rich get richer."
In a poorly-specialized network, routing could very well be nearly a random
walk.

Bernardo Huberman sent me a preprint along these lines -- I'm going to ask
if I can post it to the list.

theo

-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 49 bytes
Desc: not available
URL: 
<https://emu.freenetproject.org/pipermail/devl/attachments/20010407/966e2d6c/attachment.pgp>

Reply via email to