Locality-Aware Request Distribution in Cluster-based Network Servers
--------------------------------------------------------------------------------------
V. Pai, M. Aron, G. Banga, M. Svendsen, P. Druschel, W. Zwaenepoel, E. Nahum

This paper introduces the LARD strategy for cluster-based network servers. The 
idea behind this strategy is using content in addition to load. 
Simply put, in a cluster composed of a front-end and a number of back-ends, the 
front-end distributes requests based on service/content requested and the current 
load of back-end nodes. The aim of this scheme is obtaining a high cache hit rate 
while maintaining the load balance that state-of-the-art cluster servers (based on 
WRR) provide. 

The strength of the paper lies in this locality-aware novel approach. The authors 
design a scheme that achieves both high cache hit rates as well as load balancing 
among back-ends. Their simulation results in comparison with other schemes 
demonstrate that LARD is a successful strategy. They achieve scalability in 
throughput, a low percentage of requests missed and good utilization of the nodes. 
Another interesting approach is the TCP handoff protocol and its transparency to 
the client (no changes on the client side). 

However, in their scheme the weakness is represented by the front-end being a 
single-point of failure. They could have proposed for example a replication scheme 
for the front-end in order to eliminate this drawback. Anyway, this maybe 
complementary to their work and subject to future research.  

Overall, I find this paper to be a good technical one. They present a new strategy 
for cluster-based network servers, that is using content when assigning requests to 
nodes. They also introduce the handoff protocol. Their results show that the LARD 
simple scheme has good performances.