Review - Locality-Aware Request Distribution

This paper presents a locality-aware request distribution strategy that
attempts to leverage high cache hit rates coupled with good load balancing
to achieve high throughput in request based cluster systems. The authors
assert that a front-end can take into account both the service and content
requested to forward requests to a node that has a high probability of
having the necessary objects in cache memory. The most interesting aspect of
the research is that this is possible, and efficient, without the front-end
having explicit knowledge of caching schemes, or even the current cache
state, in the back-end nodes.

Pai et al. present a fair comparison to the weighted-round-robin scheme,
which was most used in production systems at the time of publication. In the
case where the working set of data being served exceeds that of the local
memory on the back-end systems, their concept is intuitive and seems to
achieve the desired result under simulation. The design and implementation
of the TCP hand-off protocol, which was not completely covered in this
paper, seems to be far along in development as well.

Although this paper presents a very novel approach and is coupled with a
great analysis, it does suffer several weaknesses. For example, the
front-end does require some application level knowledge, which would mean
that it may need a local implementation of each service offered by the
cluster (e.g. HTTP server, FTP server, SQL Server, etc.) Also, there is no
Secure Socket Layer (SSL) consideration. Specifically what performance hit
this would result in, and how certificates would me managed (A common
element in HTTP websites).

It is also clear that further simulation is required in order to consider a
larger set of services offered on the current Internet. This scheme may only
be valid/useful in client/server protocols where the client is pulling
finite size chunks from the server. The simulation doesn't seem to take, for
example, FTP servers into consideration. In this case, client connections
often sit idle, which would invalidate the load calculations.

Although the this locality-aware strategy does not consider many of the
common services available on the Internet today, it does make a strong
argument for locality-aware systems in clustered server solutions and is
well presented in this paper.
Received on Mon Sep 26 2005 - 09:40:51 EDT