[CSC2231] Paper Review: Locality-Aware Request Distribution in Cluster-based Network Servers

This paper describes LARD, a locality-aware request distribution
mechanism that uses a single front-end node to redirect incoming
requests to a cluster of back-end nodes. LARD achieves a good balance
between load balancing the back-end machines and maximizing memory-cache
performance. This is done by redirecting requests of the same type to
the same back-end node unless the back-end node is overloaded.

Given that previous works aim at either balanced load or cache locality,
being able to balance the two extremes of the spectrum is a significant
contribution of LARD. The TCP connection handoff protocol introduced in
this paper is also novel. An initial handoff allows a client
communicates with a back-end node transparently, whereas multiple
handoffs allows a client to be serviced by multiple back-end nodes to
further improve the balance between balanced load and cache locality.
LARD performs well under simulation and experiments.

Although LARD achieves good performance, its fundamental problem is with
the front-end node. Not only the front-end node is a bottleneck, it is
also a single point of failure of the design. The front-end node can
support up to 10 back-end node, which limits the horizontal growth of
the cluster if more processing resources are needed. Other than that,
the TCP connection handoff approach requires the front-end node to
handle all incoming data on behalf of the back-end nodes. If the
front-end node is overloaded or fails for whatever reasons, all requests
cannot reach a server to service them. So, I think using LARD somewhat
defeats the purpose of using a cluster because of these problems.
Received on Mon Sep 26 2005 - 08:58:44 EDT