review of LARD

In cluster-based network servers, a front-end directs user requests to
back-ends, which process the queries and forward data to clients. This
paper focuses on content-based request distribution and proposes a
practical strategy for locality-aware request distribution (LARD).

The main contribution of this paper is the efficient LARD strategy which
aims to combine good load balancing and high locality in the back-end’s
main memory caches. A trace-driven simulation is implemented to compare
LARD with existing request distribution policies under different
conditions. The simulation results demonstrate that LARD-based strategies
have a more efficient performance in terms of load balancing, throughput
and cache hit rate. The authors also discuss a TCP handoff protocol which
is a client-transparent connection handoff protocol for TCP-based network
services. The paper presents performance results of LARD strategy and TCP
handoff protocol.

LARD uses the caches in the back-ends according to the requests content.
Requests for a given target are sent to a particular back-end in order to
improve the hit rates in the back-end’s main memory caches. The target
assignment takes the load balancing into account. The target is assigned
to a single node in basic LARD, and is assigned to a set of nodes in LARD
with replication (LARD-R). LARD-R improves the back-end overload situation
in basic LARD.

However, the following problems are not fully addressed in this paper.
First, back-end failure may affect the performance of LARD. If a back-end
is failed, its cache content is lost. This may decrease the cache hit
rates and cause the target re-assignment in LARD. The relationship between
failure and LARD performance is not discussed. Second, the performance of
LARD and LARD-R is workload-depended. For example, if a workload includes
some high-frequent accessed objects then LARD has a better performance.
Furthermore, this paper does not discuss the target assignment based on
the relationships among targets. Third, this paper focuses on static data
served in HTTP servers. If the HTTP service provides dynamic data, that is
the content in the back-end caches is deferent from the real data
according to time, the caches need to be updated for each request. In this
case, LARD needs to be improved for dynamic content.

Overall, this paper is a well-written paper which provides sufficient
technique details and solid evaluation results.
Received on Sun Sep 25 2005 - 23:59:22 EDT