PAPER REVIEW: Lessons from Giant-Scale Services

PAPER REVIEW: Lessons from Giant-Scale Services

Summary:
The paper surveys existing approaches to large scale web services and
points out useful techniques and considerations to make in designing such
systems. It makes no new technological contributions as these techniques
are already in place by the time it is published, however it points out
existing trade-offs in performance goals for internet systems. It is thus
a survey of the author's experiences.
According to the paper, the way to deploy such large scale web services is
by use of clusters of commodity workstations.

Strength
The paper highlights the relative insignificance of uptime and that
instead yield and harvest are more inetersting to system administrators as
these variables more accurately measure the user experience provided. It
also shows well a tradeoff between these two perfomance goals through the
use of simple but powerful mathematical model, the DQ principle.
The paper further shows how partitioned systems can be used to achieve
better yields while replicated ones better can achieve higher harvests in
case of failures, and further points out how one could employ both methods
for better performance.

Weaknesses
The paper argues that the performance achieved with both node replication
and partitioning is the same. However, this is only true for cases of high
utilization, and no evidence is given to show that the sites in question
operate constantly under hight utilization levels. Under less than high
utilization, replication yields better performance as multiple queries
execute independetly with concurrency while in the partitioning scheme the
a longer query at a partition node will slow down all subsequent queries,
causing a bottleneck, among other overheads.

The mathematical expression in (1) used to represent uptime appears to be
slightly unconventional. Traditionally, uptime is given as:
uptime = MTBF /(MTBF + MTTR)
Received on Mon Sep 19 2005 - 10:12:46 EDT