Review - Pastry: Scalable, Decentralized Object Location and Routing

Peer-to-peer systems consist of a collection of nodes, each having equal
capability and role. Pastry is a scalable protocol that offers decentralized
object location and efficient routing path assignment. In this
implementation, each node is assigned a random identifier and joins a mesh
network. Objects are replicated on numerically adjacent nodes so as to
obtain optimal location for any requesting node. Pastry is able to guarantee
O(log N) routing hops to a requested resource.

Pasty assumes that adjacent nodes are diverse in geography, jurisdiction,
ownership and network attachment. It is through this that the authors claim
high locality leading to efficient routing. Locality in this sense is based
on the number of routing hops. By geographically replicating content
uniformly, any node should be able reach requested content in an efficient
number of lookups.

However, as is stated in the paper, nodes are not uniformly distributed
across the Internet. They tend to cluster. Moreover, it is well known that
triangulation inequality does not hold when considering geographically
disperse network nodes. Unfortunately, Pastry looses much of its locality
claim in this case. The simulation experiments performed by Rowstron et al.
use uniformly distributed nodes with emulated distance metrics.

While Pastry suffers from some assumptions, it would be interesting to
experiment with a real-world implementation. It may be the case that the
assumptions made here do not have a tremendous effect on the overall
performance. There are clear advantages to this system, including
fault-resilience (through replication and decentralization).
Received on Mon Nov 07 2005 - 02:16:25 EST