From: Tarik Nesh-Nash (tarikn_at_microsoft.com)
Date: Wed Feb 25 2004 - 12:14:33 PST
This paper argues the efficiency of a cluster-based scalable network
services. It uses a cluster of workstations of commodity PCs and seems
to be superior on scalability, availability and cost effectiveness
comparing to large workstations, however the implementation and
administration cost may be expansive.
This work describes the design, analysis and implementation of a layered
framework for building network services that addresses this complexity
and promotes an off shelf product that enable the users to focus on the
content rather than the implementation. Three layers are described: The
top layer implements the transformation and aggregation pf the service
(e.g. browser, search engine), the middle layer is used for caching and
transformation between MIME types, etc, the lower layer is used
scalability, load balancing, etc. The layering offers the separation of
content and implementation by supporting layers reuse and modifications.
The clusters offer obvious advantages: scalability (no forklifting),
availability (24/7), and better cost/performance, however this
architecture faces many challenges mainly: Administration, component vs
system replication, partial failure handling and shared state. The
paper them discusses the ACID/BASE differences and argues that the BASE
drawbacks are acceptable for most cases, ACID architecture may still be
used for special cases. BASE choice offers better performance and is
easier to implement.
A detailed description of the implementation of two real cases (Transend
and hotbot) is discussed.
This paper is a good introduction to the cluster networking. I liked
the idea of trading some consistency for simplicity and performance.
However, the administration is still a tough concern.
This archive was generated by hypermail 2.1.6 : Wed Feb 25 2004 - 12:14:43 PST