(no subject)

Jing Su - jingsu_REMOVE_THIS_FROM_EMAIL_FIRST_at_cs Review: NOW

SUMMARY

This is a concept and ``vision'' paper, which argues that the
resources of commodity workstation computers can be pooled together to
achieve significant computing power -- even rivaling supercomputers.

The primary focus of the paper's argument is sharing the large pool of
available memory and processors. Fast network layers can allow memory
paging to other NOW systems to be significantly faster than paging to
disk. The paper also argues that many modern workstations spend
significant amounts of time idling (or waiting for IO), thus creating
opportunities for running pieces of highly parallel computation tasks.

Other issues such as sharing disk, security, and handling rashes are
mentioned in this paper, though only presented as a high-level vision.

KEY STRENGTHS

I think the key strength of this paper is the broad vision which it
presents. As a disclaimer, I'm too young to know whether this 1995
paper is seminal in terms of bringing together a big vision that
launches projects such as cluster computing, SETI, and beowulf.

KEY WEAKNESSES

As a systems paper, the biggest weakness of this paper is that no
working prototype was presented. The paper does mention a working
distributed memory paging system (GLUnix), though little information
regarding its implementation and performance is given. For example,
one expects issues such as permission handling, checkpointing and
rollback, and crash recovery to be more challenging than the paper
suggests.

This paper also makes the assumption that progress will not be
significantly impeded by synchronization issues and program linearity.
The feasibility study presented in the paper is a best-case upper
bound, which assumes no synchronization stalls and memory barriers, no
checkpoint and crash-recovery and roll-back costs, and perfect
parallelism.

COMMENTS

There is also a 1995 paper which presents an implementation for global
memory sharing across a computing cluster: Implementing Global Memory
Management in a Workstation Cluster. Michael M. Feeley, William
E. Morgan, Frederic H. Pighin, Anna R. Karlin, Henry M. Levy, and
Chandramohan A. Thekkath. Proc. of the 15th ACM Symposium on Operating
Systems Principles , December 1995.

Thus, my gripe with this paper is not that they did not present the
GLUnix prototype, but that their prototype does not come close to
addressing the large problem space they present. The above paper
tackles memory sharing and memory sharing alone. This paper tackles
so much more.
Received on Wed Sep 14 2005 - 14:59:37 EDT