From: Prasanna Kumar Jayapal (prasak_at_winse.microsoft.com)
Date: Mon Feb 02 2004 - 17:39:31 PST
This paper ("Experience with Grapevine - The growth of distributed
systems") describes an early distributed, replicated system, "Grapevine"
and presents a report on what the authors have learnt from using this
system in a large and practical environment. The authors also talk about
the potential improvements that were not carried out, hoping that their
experience may offer some help to designers of new systems.
A Grapevine system is composed of a wide area network dotted with
Grapevine servers, one or more of which is typically located (and
services) within a given local area network. It mainly provides
messaging and registration services. The registration data which
controls the messaging activities is stored in the registration
database. This database consists of entries (called RNames) and there
are two types: group entries (contains set of info about groups and
individuals like distribution lists, ACLs, etc) and individual entries
(like passwords, inbox sites, etc). Replication of key data makes for a
reliable, reconfigurable system.
The authors discuss remarkably well the effect of a large number of
users on the system and how small design decisions may turn to be system
bottlenecks. One issue with a distributed system of this type is its
scalability, which is given due attention in this paper. Among
particularly interesting problems discussed are how and when to increase
the number of servers, performance issues of distributing mails for long
(and possibly nested) distribution lists, the increasing flow of mails
among far away nodes (and how multistep forwarding improves efficiency
and robustness under unreliable links), and the different
tradeoffs between safety and higher communication bandwidth (or update
delays) when control data (registries) are replicated. Nevertheless,
many of the paper's suggestions are based on only quantitative empirical
evidence, and hence, it is not clear whether some of the solutions can
be extrapolated to other distributed systems.
One interesting aspect of the paper is that it cleverly introduces
solutions to several (currently standard) issues presented in both
e-mail server configuration and network topology design (like mailing
groups and recursive mail aliasing, distributed mail delivery,
replication of services and space distribution organized under domain
names - registry -).
The success of the authors in achieving the goals of a distributed,
reliable, performance-oriented meta system is best reflected in the
comment: "most users and registrars treat Grapevine as if it were
implemented on a single, large, reliable computer ... to which all
workstations were connected through high-speed links."
Unfortunately, it doesn't appear as though Grapevine will "become the
source of authentication and access control information for the
internet." This may partly be due to freezing the Grapevine design and
implementation before its time. I wonder if they ever thought of a
distributed Operating system at that time. Overall, I think this was a
good paper to read through and understand the necessities of a
distributed system.
This archive was generated by hypermail 2.1.6 : Mon Feb 02 2004 - 17:39:28 PST