From: Ankur Rawat \(Excell Data Corporation\) (a-arawat_at_microsoft.com)
Date: Mon Feb 23 2004 - 16:31:22 PST
Andrew is a distributed file system at CMU. I think the file system was
designed to provide better performance to users of the network and make
the network more scalable.
The system on the face of it looks very simple. It had been implemented
mostly in user space and had some optimizations at the kernel level.
The system leverages then existing technologies. Standing in today's
perspective I was not every sure about the intention of the undertaking
of this project. But probably in late 80s disk performance was a huge
bottle neck and justified such a project.
The system operates as a typical client server model. Many server nodes
act as file servers. The system claims to provide BSD file system
semantics with some exceptions, but does not specify which.
The implementation of the final system occurred in 2 cycles. The first
cycle did not meet the performance objectives of the project. The
lessons learnt in the first cycle were used to make some design changes
in the second cycle.
I liked the mechanism of call back to solve the cache validation
problem. But, I did not quite understand how coherency across the client
nodes was managed. I was liked the huge amount of analysis data
presented and used in justifying the design choices made. It was
interesting to note that in spite of moving a lot of work to the clients
the server utilization was high sometimes and a bottleneck.
The second cycle of the project represented sort of the evolution of the
client sever model and the distributed computing model in general.
This archive was generated by hypermail 2.1.6 : Mon Feb 23 2004 - 16:31:19 PST