Paper review

Student Name: Ali Akhavan Bitaghsir,

 The paper reports on the Berkeley's "NOW" team project on a realization of
a network of workstations which is capable of performing large scale
parallel computing *and* performing dedicated interactive desktop
applications, based on the newly emerged technologies enabling scalable low
latency networks with low overhead network interfaces.

 The paper comprehensively discuss the proof of the market's higher demand
on smaller cheap workstations rather than super-computers and deduce the
possibility of creating networked workstations with the same performance of
supercomputer but with less price thanks to the fast (low latency & low
overhead & high bandwidth) switched networks invented recently. The work
done is significant in terms of practicality; the NOW team has completely
implemented the required O.S. (distributed GNLinux) and tested O.S. against
performance metrics. Essentially the contribution of the paper and its
effects on the *real *economy seems to be noteworthy since the team has
built the system on top of a commercial operating system (GNLinux).

 There authors have considered some assumptions in the paper but have not
explicitly mentioned them.

 1) It is mentioned that the workstation cost-performance is reducing much
faster than supercomputer's and that's why the authors expect a great impact
on the economy through NOW. However, one should also consider the relative
portion of the companies with a building full of computer to the whole
market of desktop users is not that much, which translates into a
lower-growing market for NOW too.

2) Unfortunately, the authors mention their distributed caching and
especially cooperative file-system algorithms quite vaguely; whereas it
might be the case that their algorithms do not scale into NOWs with many
computers in terms of the complexity of control and data messages needed.
Besides, the added complexity of these algorithms (if they do not scale
simply with the number of workstations) may increase the lag time of product
release for NOW, although they handle this more with software rather than
hardware.

3) The last point to mention which also gives a future direction of research
is the NOW team study on the compromise between interactive dedicated
computing vs. large scale parallel computation. In the current version of
NOW, the parallel computation is done only on computer which are
"completely" idle and this leads to having no-guarantee of performance for
the large scale computation. An possible solution is to divide the
workstations into several computation-classes allowing the users of each
class to use only bounded computational applications (for example for hard
real-time applications and soft ones). . Then, the network can use a portion
of the CPU time of the computers with less computational burden. Another
approach is to send the workstation's job into another workstation and not
tolerating the time for migrating the whole previous process to another
station.
Received on Thu Sep 15 2005 - 11:01:42 EDT