REVIEW: A Case for NOW

This paper deals with an important and interesting problem of achieving
highly scalable, highly available and high performance systems quickly and
at modest costs by using "full" workstations as building blocks. It
demonstrates how commodity workstations can be interconnected to tap into
idle resources of workstations and combine them to give the processing power
of a supercomputer or massively parallel processors.

The key strength of the paper is that it describes a practical system at a
time when enabling technologies were around, i.e. a "killer" network and a
"killer" processor. It also clearly outlines the goals of the system and
important performance considerations in implementing such a system. I also
liked the use of cooperative caching, low-cost software RAID and load
balancing in their system to achieve good performance.

The weak side of the paper lies in its idea of security. The proposed
solution addresses security of the cluster that's not available to the
outside world. Even in 1995, when Yahoo! and Amazon were a couple years old,
they were fairly popular and would need some other form of security against
network attacks if they were to adopt a Network of Workstations (NOW) for
their business needs. Therefore a potential follow-on, at the time, would be
security considerations for NOW. The other weakness of the paper is that the
authors give the impression that the system is infinitely scalable by simply
adding non-interactive machines to the NOW. This is definitely not realistic
and a scalability study would make their scalability claim much stronger.

Overall it's a well written paper that presents a solid idea supported by
implementation and measurements. It starts by providing a good motivation
and follows by outlining the concepts, of what forms the basis of modern
clusters. Their idea was drawn out 11 years ago and accurately predicts that
"NOWs will be the systems of choice for large scale computing within a
decade".
Received on Wed Sep 14 2005 - 17:57:47 EDT