Next message: Ankur Rawat \(Excell Data Corporation\): "Andrew Distributed File system"
Scale and Performance in a Distributed File System
describes
the experiences and improvements in the Andrew File System developed at
Carnegie Mellon
University.
When I read the first portion of the
paper I kept thinking that the approaches chosen would not work well
either for
scaling or even low usage performance. I
was wondering why we were even reading this paper except possibly as an
example
of a bad design. I then had to laugh at
myself when I realized the point of the paper was actually to describe
the successful
improvements made to achieve the scalability and performance goals.
The cache coherence strategy used is
not nearly as strong as
is used for memory cache coherence.
However, I’m not sure that it needs to be. I’m
a little uncomfortable with the burden of
synchronizing file access being placed on the applications given one of
the
goals of the system is transparency. I
did like the transaction-style strategy for broadcasting updates. The changes are only visible when the file is
closed. I also liked the choice of
allowing existing open handles to continue using the older data. I know of transaction support going into a
new OS and that implementation will make the modifications visible to
existing
handles as soon as the transaction is finalized. I
believe that strategy to be flawed. I prefer
either invalidating the handles or
using old data as was implemented.