CSC2231 LARD review

This paper makes web clusters fast by increasing locality in memory cache
of back end servers. It combines load balancing and content-aware concepts
into the design of front-ends.

This work is a fine grain content-based request distribution. However, it
increases the complexity of front ends since the dispatcher must be aware
of semantics of requests. Moreover, their simulation only shows the algorithm
scales well for tens of machines; it is unclear about the scalability of this
algorithm for a larger scale clusters.

This algorithm ignores the locality among targets. Because close targets
may reside in the same page of memory, assigning close targets to the same
server may also improve the cache hit rate. In addition, this paper does
not discuss the granularity for a target for different applications.

The performance gain of LARD depends on whether the working set size
exceeds the memory size. This is also related with concrete traffic
pattern. In some database dynamic partition scheme, they divide data into
several parts so that the working set of each part can roughly fit into
memory. As an alternative approach, the latter may also achieve good
performance. No comparison with such data partitioning is given in the
paper.

In general, it is among one of pioneering papers that broadly extend the
functions of the load balancer for web clusters.
Received on Sun Sep 25 2005 - 22:47:19 EDT