[p2p-hackers] DHTs in highly-transient networks

[Saikat Guha]

On Thu, 2005-12-01 at 18:38 +0000, Salem Mark wrote:
> in "highly transient networks", where the number of nodes 
> appearing and disappearing are very high, maintaining the DHT becomes hard 
> and introduces considerable overhead.
> 
> I am trying to find out what exactly "highly-transient" means. A file 
> sharing network like Gnutella, seems to be highly transient, where peers 
> join/leave the network frequently. 

True; the answer depends on the particular application and protocol.

In Gnutella (without ultra-peers), activity of all clients would affect
the network equally. With an intelligently chosen subset of nodes,
(ultrapeers in Gnutella, supernodes in Kazaa and Skype) the effects of
churn can be mitigated. This relies on the assumption that this subset
of nodes is more stable than the rest. The assumption appears to be
borne out in Gnutella (Daniel's paper), and in Skype [1]. 

[1] An Experimental Study of the Skype Peer-to-Peer VoIP System 
    http://www.guha.cc/~saikat/pub/cucs05-skype-abstract.php


> Could somebody elaborate on this? is 
> there a node departure/arrival/failure rate (per sec? per min?) that 
> identifies "highly-transient" networks ?

FWIW, in [1], we found that the supernode turnover is typically less
than 5% / 30min. Median supernode session time is 5.5 hours; session
time is heavy-tailed (Pareto) and not exponential. Supernodes are much
more stable than regular nodes.

Btw, if anyone wants a copy of [1], please email me directly. It has
data on Skype supernode lifetimes, churn rates, comparison between skype
supernodes and regular nodes, Skype VoIP and file-transfer workload
characterization, etc. In short, we found that Skype differs
considerably from filesharing networks (different usage model, much
higher median lifetimes etc).

cheers,
-- 
Saikat
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: This is a digitally signed message part
Url : http://zgp.org/pipermail/p2p-hackers/attachments/20051206/26715c26/attachment.pgp