(hopefully) paper reviews

Bryan Wang (wangbry@hotmail.com)
Wed, 03 May 2000 19:10:11 PDT

Get Your Private, Free E-mail from MSN Hotmail at http://www.hotmail.com
Bryan Wang, 5/3/00

The Sprite Network Operating System

Sprite addresses the need for a high-performance, highly transparent
distributed OS. Authors of this paper, unlike those of previous
papers we've read in class, carefully consider current and future
trends in computing and performance when making design decisions.
Sprite was designed with these forward-looking constraints in mind.

Sprite uses a single file system hierarchy based on prefix tables to
maintain consistent files over the network. It uses client stubs and
RPC transports to make inter-process interaction across networks
completely transparent. It even supports remote kernel calls, so a
process' interaction with its home node remains the same after the
process is migrated.

I found this paper to be the most readable paper yet, and the first to
include detailed performance measurements. One area that Sprite is
lacking in features is protection. Memory sharing is very coarsely
implemented, a process may share all-or-nothing.

With all its forward-looking design decisions, it would be interesting
to see how this OS runs on today's workstations (which, as the authors
correctly predicted, are about a factor of 100x above 1988
workstations in terms of network speed and local RAM).

Bryan Wang, 5/3/00

The Distributed V Kernel and its Performance for Diskless Workstations

The designers at Stanford were trying to promote the use of diskless
workstations by designing a kernel built on short IPC messages
communicating between workstation and a file server. They intended to
prove that a diskless workstation can operate nearly as fast as a
workstation with a local disk, but without the cost and management
overhead that a local disk entails. Cheriton and Zwaenepoel did away
with the layered network model, since efficiency is of the essence on
a diskless workstation. Testing demonstrated severe penalties to
using layers and process-based network servers.

The paper supports its design through a series of sound performance
measurements. This is necessary because the feeling of the time was
that diskless workstations are "bad". They based performance of data
transfers relative to an inherent "network penatly", and the fact that
actual kernel performance was so close to this ideal means thier
design was successful.

A great conclusion of this paper is that the performance of IPC
transfers is so good that it can replace special file access
protocols, transfer protocols, and remote terminal protocals. I
enjoyed this paper because the V Kernel design is so different from
something I would normally consider (diskless workstations, exclusive
use of small packets).

The use of "raw" ethernet data link in the kernel imposes considerable
restrictions on the hardware. What if two years from now (uh, i mean
1983), nobody likes Ethernet anymore? But that was a design trade-off
that they considered and were willing to live with.