CSC 2231 Review 1st Week NOW

This paper predicts that Network of Workstations (NOW) would replace
mainframes and supercomputers since NOW has better cost-performance.
NOW project proposed several techniques, such as low overhead
communication, GLUnix, xFS, to provide a global computing environment.

However, this prediction does not fully come true after 8 years.
Mainframes are alive in large organizations, and today there are still
hundreds of thousands of powerful desktops that are only used to edit
documents and are idle most of time. I think this paper fails to
consider some important factors as follows:

First, parallel paradigms are still challenging for normal programmers
although many parallel libraries, such as DSM, MPI, OPEN MP, are provided.
More important, even if we write some parallel programs for workstations,
the overhead could be very high. NOW proposed to use ATM LAN, Myrinet, etc.
techniques to reduce communication and message processing overhead,
nevertheless, they are not widely used. Hard to use makes NOW lose users.

Second, there are not enough proper applications to drive regular
enterprises to build NOW. Parallel programming tasks are often scientific
applications, which are not widely needed by general departments or
companies. In addition, some applications, for instance, database servers,
are hard to achieve good scalability in NOW due to consistency or
dependency issues. Thus, big enterprises, such as banks, bureau of
meteorology, still buy mainframes to run their applications.

Last, but not the least, there are some social reasons. People usually are
not willing to share resources without getting benefits. Also, sharing
resources could bring many security and performance issues.

On the other hand, NOW project does have a big impact on the popularity of
distributed systems because it realizes the tremendous power of PCs. In
particular, as a subsequent project of NOW, network storage systems
like SAN achieve a success.

We can know some lessons from NOW. Distributed systems may not all target
at replacing super computers; they could offer applications that super
computers cannot have. For example, P2P file sharing or P2P streaming
needs a large number of computers to involve. For enterprises, dedicated
clusters are more suitable than using users' desktops. Moreover, we do need more
incentives, i.e. attractive applications or other approaches, to make end
PC users to contribute their resources. Finally, from parallel platforms'
view, workstations are promising with the appearance of G-bit network card
and fast fiber optic networks, but we do expect better parallel paradigms.
Received on Wed Sep 14 2005 - 16:54:37 EDT