From: Tarik Nesh-Nash (tarikn_at_microsoft.com)
Date: Wed Feb 25 2004 - 18:02:47 PST
This paper presents a highly scalable cluster based mail service and
discusses its motivation, design and performance. In order to achieve
these goal, three requirements are met:
- Manageability: Porcupine is self managed, it automatically
handles nodes failures and recovery, nodes addition and network
failures. The system changes its configuration dynamically and the
administrator role is very limited to adding new machines and replacing
broken machines.
- Availability: Regardless of the state of an individual node,
Porcupine does not fail on delivering mail services to the user. This
is achieved by duplicating the data in nodes as a back up from
individual node failure. The nodes update occur slowly and is not
automatically persisted. This may cause some erroneous behavior as
duplicate emails or existence of deleted message. This problem is minor
comparing to the gained benefit.
- Performance: The performance is approximately linear to the
number of nodes. The work load is dynamically load balanced. The
decision is decentralized to the node level, so it may be based on some
outdated information. The performance, though, is better than fixed
user mapping clusters
Four techniques were used to achieve this implementation: Functional
homogeneity, automatic reconfiguration, dynamic scheduling and
replication.
This system is a good variation of the BASE mechanism discussed on the
previous paper and provides an improved solution than the prototype
Grapevine.
This seems more like a description of an application rather than
operating system paper.
This archive was generated by hypermail 2.1.6 : Wed Feb 25 2004 - 18:03:16 PST