Review - LARD

            This paper proposes a strategy to improve the performance and
scalability of clusters by exploiting locality and an aggregated view of
main memory cache of a cluster. The system, LARD, is composed of a front-end
node which examines request contents and forwards the requests to a cluster
of back-end nodes which then respond to those requests. It improves main
memory cache hit rates, has good load balancing, low system idle time and
better throughput than traditional weighted round robin (WRR) systems
currently (1999) deployed on clusters.

            The paper is fairly well written and provides an elaborate study
of their proposed system which is supported by simulations (although with
idealistic assumptions) and a prototype system to validate their ideas and
simulation results. The system is evaluated on real-life traces and is
compared to WRR, showing that LARD has better performance than currently
deployed WRR. The system's other strong point is its scalability and its
compatibility of the proposed TCP handoff mechanism with standard
client-server protocols, which is essential for a real-life deployment of
such system.

            In my opinion, the weak point of the paper is the lack of
description of the hashing algorithm they use to partition the database in
LARD on system startup. It is one of the essential parts of the system since
they make assumptions, like good partitioning of name space and working set.
Their description only deals with steady-state of how to redistribute the
load, etc. They also fail to give a better description of their LB/GC system
but used it for comparison with LB.

            At the time of writing (1999), this paper seems to present an
idea not explored by others (or at least not implemented and evaluated
[13]). The ideas explore an important problem of speed and scalability at
the peak of the Internet-based economy.

 
Received on Mon Sep 26 2005 - 02:24:30 EDT