[p2p-hackers] Symmetric Replication for Chord

[Matthew Leslie]

Hi Michael,

I would suggest that replication factors which do not divide the
keyspace are not going to cause problems, simply make sure you round any
fractional numbers of keys in a consistent manner. A greater issue might
be replication factors which do not divide the number of nodes, as this
will lead to an uneven distribution of replicas over nodes.  This is an
issue I have run into myself - perhaps this is what you meant? I don't
have any useful suggestions on this myself, yet.

The practicalities of replication in Chord quickly become somewhat
fiendish.  We've found that although using a second hash function to
place replicas can increase performance, it is at the cost of
considerable extra complexity. You can see my comparison of various
replication schemes, including a form of symmetric replication in a
paper I wrote for DASP2P-06. 

http://urchin.earth.li/~mleslie/mleslieStorage.pdf

The 'Block' replication algorithm in my paper is an instance of
symmetric replication, though it divides the nodes into equivalence
classes in a slightly different way to that proposed in the paper by
Ghosi et al. The most interesting thing we have found (I think) is that
symmetric replication leads to a more reliable system for a given
replication factor than DHash style replication - the authors of the
original paper do not make this point.


Matt


PS: Sorry it took me so long to get round to replying, I hope this isn't
irrelevant now.
	
 
 

________________________________

From: p2p-hackers-bounces at zgp.org [mailto:p2p-hackers-bounces at zgp.org]
On Behalf Of Michael Hartmann
Sent: 23 February 2006 17:20
To: p2p-hackers at zgp.org
Subject: [p2p-hackers] Symmetric Replication for Chord


Anyone actually tried implementing the replication scheme suggested by
this
paper: 
Symmetric Replication for Structured Peer-to-Peer Systems,
A. Ghodsi, L-O Alima, S. Haridi, DBISP2P2005, 2005.
http://dks.sics.se/pub/replication.pdf 

I am trying to implement it for Chord, it seems better than the
successor-list approach, but I am wondering if it is possible to
make it work for replication degrees that do not divide the size of the
key
space?

Mike