Pastry Paper Review

This paper introduces Pastry as a P2P object location and routing infrastructure which performs efficiently in large-scale peer-to-peer systems. After introducing the model of participating nodes in Pastry, which is much like the Chord P2P node model, the authors talk on the algorithms they use for maintaining the efficiency of routing and more importantly the availability of objects while we have concurrent node arrivals and departures in the network.

>From a high-level point of view, I think the main contribution of this paper to the previous work done in p2p systems is the ability of Pastry to take into account the locality of nodes while routing. The routing tables designated for nodes resemble the "fingers" in the Chord p2p system; yet, we have no notion of locality in Chord. Ofcourse, Pastry has to make some assumptions for this feature, one of them seems to be un-applicable for many metrics : The triangular inequality should hold for the proximity metric, whereas, for some metrics like the number of hops in routing, this inequality does not hold. Even for the distance between nodes, Internet has to grow more to show this property -- Although an study showed that only 5 Percent of triple hosts in the Internet, violates this property. Authors could also do some more experiments to find out the "sensitivity" of their results to their assumption about the triangular inequality -- how much does the efficiency change if the inequality violates to a certain extent.

And the last point is that authors do not address the additional messaging load imposed to the system when a newly joined node gets the state from each of the nodes in its routing table and its neighborhood set. Thereby, it is not specified that how much does it take for the network (after an arrival) to become get back its "Locality" property.
Received on Mon Nov 07 2005 - 10:12:34 EST