REVIEW: "Pastry: Scalable, decentralized object location and routing for large-scale peer-to-peer systems"

From: Nilton Bila <nilton_REMOVE_THIS_FROM_EMAIL_FIRST_at_cs.toronto.edu>
Date: Mon, 7 Nov 2005 03:05:21 -0500

"Pastry: Scalable, decentralized object location and routing for
large-scale peer-to-peer systems"

The paper presents Pastry, a scalable architecture for object location and
routing of messages in wide-area peer-to-peer applications. Pastry maps
keys to nodes with Ids numerically close to the keys. Every node is
assigned a numeric Id, and every value is assigned a numeric key. Unlike
similar systems, Pastry maps each key to k nodes numerically nearest to
the key itself instead of mapping to just one node. This provides
replication, and since the node Ids are randomly assigned, such k nodes
will be geographically diverse, providing room to explore locality. When
Pastry routes messages to nodes it takes into account the locality of the
nodes. Messages reach nearest local nodes before they can be forwarded to
far ones.

The paper demonstrates that messages in Pastry are routed with good
locality, according to proximity metrics supplied by the user level
application. Experimental results demonstrate this locality as, on
average, messages travelled only 4 hops to reach destination nodes in a
network of 10,000 nodes. The deviation from optimal routing was also found
to be only 30%.

The paper also provides experimental evidence that Pastry handles node
failures well. Results show that Pastry maintains near optimal routing
tables even after failure of 10% of nodes. Handling of nodes arrival is
also addressed in the paper.
An additional strength of Pastry is the provision of a simple API which
allows applications to easily integrate Pastry into their code base.

A weakness arises, however, as proximity is the only metric considered in
the paper for network performance. Bandwidth is also very important in
determining the performance of a distributed application. Proximity (e.g.
latency) not always determines the best site to retrieve data from, as for
large data, higher bandwidth even if with higher latency is preferred.
Received on Mon Nov 07 2005 - 03:05:35 EST

This archive was generated by hypermail 2.2.0 : Mon Nov 07 2005 - 03:21:36 EST