Sawmill: A High-Bandwidth Logging File System
Ken Shirriff
John Ousterhout
University of California, Berkeley
Abstract
This paper describes the implementation of Sawmill, a network file
system using the RAID-II storage system. Sawmill takes advantage of
the direct data path in RAID-II between the disks and the network,
which bypasses the file server CPU. The key ideas in the
implementation of Sawmill are combining logging (LFS) with RAID to
obtain fast small writes, using new log layout techniques to improve
bandwidth, and pipelining through the controller memory to reduce
latency. The file system can currently read data at 21 MB/s and write
data at 15 MB/s, close to the raw disk array bandwidth, while running
on a relatively slow Sun-4. Performance measurements show that LFS
improved performance of a stream of small writes by over a order of
magnitude compared to writing directly to the RAID, and this
improvement would be even larger with a faster CPU. Sawmill
demonstrates that by using a storage system with a direct data path, a
file system can provide data at bandwidths much higher than the file
server itself could handle. However, processor speed is still an
important factor, especially when handling many small requests in
parallel.
Download the full text of this paper in
ASCII (50,665 bytes) and
POSTSCRIPT (197,069 bytes) form.
To Become a USENIX Member, please see our
Membership Information.