From: Ankur Rawat \(Excell Data Corporation\) (a-arawat_at_microsoft.com)
Date: Mon Feb 23 2004 - 16:12:09 PST
This paper is about implementation of a log structured file system
called sprite. The authors leveraged concepts and technology from the
dbms world in implementing this file system. What is most interesting is
that though the paper is 10 years old - we still face the exact same
problems in secondary memory performance today.
I liked the clear objectives the author had in what the file system
intended to achieve. The paper very clearly analyses the then current
state of the art of secondary memory technology and identifies the
issues that are the bottlenecks.
The main issues are:
* Synchronous writes to the disk
* Low utilization of the disk I/O bandwidth
* High number of seeks
* Fragmentation
The authors focus on improving the write performance of there File
system.
They do so by:
* Organizing data as a log structure
* By compacting the layout of the data written
* By writing large amounts of data at one shot, therefore
improving the disk I/O bandwidth utilization
There is a tradeoff between disk utilization and write performance in
implementing their system. I found that to be unrealistic because the
majority of computer users intend to use their disk at high utilizations
and do not so much care about write performance.
Sprite file system might make sense in a shared file system on the
network where there are thousands of users and writing a lot of data to
the disk.
This archive was generated by hypermail 2.1.6 : Mon Feb 23 2004 - 16:12:09 PST