PAPER REVIEW: A Case for NOW (Networks of Workstations)

PAPER REVIEW: A Case for NOW (Networks of Workstations)

SUMMARY:
The paper argues that Network of Workstations (NOW) will be the primary
infrastructure for processing demanding computations while still
servicing low end interactive applications as NOWs provide better
price-performance ratio than do supercomputers and Massively Parallel
Processor systems (MPP), which were then the primary infrastructure for
demanding computing. This argument is propelled by observing that
performance improvements in workstations and personal computers are much
higher(80% price-performance improvement a year) than in the
supercomputers (20-30% a year) market and that faster interconnect
technologies are emerging.

CRITIQUE:
Strengths
The paper masterfully argues that given the faster rates of performance
improvements in commodity workstations (due to lower development costs
to the user as these are amortized over a large market), and given the
latency involved in introducing technologies improvements from
workstations into the MPP market, for equal computing performance the
cost of an MPP or large multiprocessor server is at least twofold that
of a NOW of cost effective workstations. It is thus argued that NOWs
will grab much of the market for supercomputers.
The paper also points out the challenges in implementing NOWs such as
the overhead and latency incurred in message passing, reliability
concerns for the local user and for the dedicated process, among
others, and attempts to address these through the use of emerging
faster networks, the use latency time for processing, as well as provide
guarantees for each type of user. It does however point out that
difficulties with the overhead of message passing are not easily overcome.

Weaknesses
The paper provides experimental as well as simulated evidence in support
of claims, however fails to explain how such experiments and
simulations were conducted, for example in Tables 2 and 3.
The paper makes the assumption that a workstation is available when it
has been idle for 1 minute, however fails to assess how significant is
the overhead of migrating processes if on average 1 minute was the
average time of availability in each instance, i.e. worst case scenario.
It would also be beneficial to point out in detail the disadvantages
inherent to the system as it attempts to adapt hardware and software
built for different purposes.

COMMENTS:
It is worth noting that the paper was originally published more than a
decade ago (1995) and since then, the Berkeley NOW project in general,
propelled wide use of NOWs in large scale computing applications today,
as it demonstrated that such infrastructure was feasible. A quick
survey of today's fastest supercomputers includes a number of fairly
inexpensive NOWs such as the National Center for Supercomputing
Applications (NCSA) cluster of PlayStation 2 's [1], IBM clusters,
Beowulf, etc.
We note, however, that a decade later, the goal of sharing the computing
resources between dedicated processes and interactive ones while
providing guarantees to users has not been as successful yet.

[1]http://arrakis.ncsa.uiuc.edu/ps2/
Received on Thu Sep 15 2005 - 01:13:42 EDT