cse221: paper evals

Octavian Luca (oluca@cs.ucsd.edu)
Thu, 4 May 2000 04:37:47 -0700 (PDT)

V Kernel

This paper discusses the V Kernel, a message-oriented kernel that provides
uniform local and network IPC. It differs from previous distributed
kernel efforts in it's choice of a network environment made up of diskless
workstations linked via fast ethernet connections to other workstations
and file servers. This topic is interesting and important of late since
the advent of fast network connections has motivated designers to
reevaluate distributed system architectures, sometimes leading to
interesting solutions such as diskless workstations in this case.

The paper discusses the major decisions made in the design of the system
and argues that they give good performance when compared with more
orthodox approaches. The authors argue that the 2 main unique decisions
they made of having diskless workstations and using general purpose
network IPC facility yielded significant benefits. This system
consequently has lower hardware costs per workstation, simpler
maintenance, little memory overhead, and fewere problems with replication,
consistency and distribution of files. They also point out major
disadvantages people usually associate with their design decisions. The
paper first explains the implementation details and the system structure
which focuses on optimizations for obtaining maximum performance. Then
the paper describes the methods used to measure the performance of the
kernel according to various criteria and comparing them to their
counterparts in conventional system. The measuremens compare the cost of
remote operations versus the cost of similar local operations, and the
cost of file access network operations versus cost of other types of
network access. The claim of the researchers that their system has
minimum attainable network penalty is supported by their measurements.
However these tests are based on synthetic benchmarks rather than on
observations system use over time which has to be taken into account when
considering their findings.

>From this paper I learned how to successfully argues the viability and
soundness of a design through the use of experimental results based on
logically formulated synthetic benchmarks. Also this shows the importance
of reconsidering old decisions when new hardware developments such as fast
networks expand the envelope of design possibilities.

SPRITE

The paper discusses the design decisions made in the implementation of the
Sprite OS. It is a distributed operating system whose goals are to hide
it's distributed nature as much as possible from the users, provide the
sharing and communication of time-sharing systems on a network of high
performance workstations and servers linked by a high speed network. The
OS itself is meant to take advantage of multiple processors for even
distribution of system services. Similarly to V kernel, this design is
meant to take advantage of hardware advances such as the availability of
cheaper memory and fast ethernet connections.

A major goals of this system is facilitating resource sharing. To
accomplish this, they strive to achieve a transparent network file system,
memory sharing among processes on the same workstation, and migrating
processes among workstations to use idle machines. The paper discusses in
detail the mechanisms built into the kernel and argues their performance
merits when compared to other previously proposed solutions. Performance
measurements are provided to practically illustrate their arguments.

Of note is the use of multithreading to take advantage of workstations
with multiprocessors, and RPC's to allow a simple way for different
workstations to invoke each other's services. Also the paper extensively
discusses the impact of caches and decistions made in their implementation
on this system.

The paper was useful to further illustrate ways of taking advantage of
distributed systems, and how to effectively take advantage of powerful
workstations in a cooperative environment.