Paper #13

A Scalable Content-Addressable Network

Problems

This paper proposes the design of a distributed Scalable Content Addressable 
Network(CAN) that is scalable, fault-tolerant, and completely self-organizing.

Strengths & New Ideas

(1)Completely Distributed:

THe design of the CAN is such that there is no centralized configuration management
required. The nodes are self-organizing and the network adapts as new nodes join,
leave or fail.

(2)Fault tolerant:

The CAN system is highly fault tolerant owing to its design. Based on the dimension
of the space, a node as many neighbors to choose from for routing it's query. This
way if one neighbor fails, there are other neighbors that can be used to rout the
query. The system is also designed to handle the case where all neighbors fails. In
this case a node search message is flooded until another node is discovered and added
to the neighbor list. This feature makes the system highly fault tolerant.

(3)Designed to minimize latency:

The authors have taken a number of measures to reduce query latency in the CAN network.

 i. They do this by adding constraints to where a node is assigned when it joins the network.
Instead of making the decision and topology totally random, they assign the node based
on the underlying IP topology. They used RTT times to well known systems(such as DNS servers)
to make this decision. This allows the clustering of nodes that are physically close to each
other--these have a higher chance of becoming neighbors in the system.

 ii. Instead of routing the query through the node that is nearest to the destination node
based on it's cartesian coordinate, they also keep a list of the RTT times to all the neighboring
nodes and use those instead.

(4) High Data Key Availablity by Caching, Replication

Data availability is high in the CAN through the use of caching and hot-spot management. 
CAN nodes maintain a list of the recently accessed keys. Before forwarding the request 
for a key, a node checks it's cache for the availability of that key and determine whether
it can answer the request without forwarding it further.

A node that is being overloaded with a particular data key request realizes this and 
replicates the data key at it's neighbors to balance the load. Hence, a popular data key
gets replicated in a region surrounding the original storage node. 


Weaknesses:

(1) Even though the nodes use RTT bases system for routing, the mechanism is greedy and
might not always result in the quickest route for hop size > 1. Assume that there are
2 neighbors of Node N, A and B. RTT to A is 10 and RTT to B is 100. The greedy algorithm
will select node A as the routing node. However, it is possible that the successive RTT
times on the route through node A might be greater than the RTT times through node B. This
case is not handled or discussed in the paper.

(2) The CAN results in the paper are just simulations. It would be interesting to see
such a network in an actual peer-to-peer system and then measure the metrics since a live
environment always throws unexpected surprises that reveal weaknesses in the system.

(3) The argument about load balancing in the network is brief. This could
be discussed further.