USENIX Technical Program - Abstract - USENIX Annual
Conference, General Session - June 2000
A Comparison of File System Workloads
Drew Roselli, Jacob R. Lorch, University of California, Berkeley;
Thomas E. Anderson, University of Washington
NOTE: This paper has an error in the proceedings. In Figure 2 (Block Lifetime), the NT line is taken from a different time period than that stated in the paper. As a result, the graph does not match the text description of it. These versions of the paper have this error corrected.
Abstract
In this paper, we describe the collection and analysis of file system
traces from a variety of different environments, including both UNIX and
NT systems, clients and servers, and instructional and production
systems. Our goal is to understand how modern workloads affect the
ability of file systems to provide high performance to users. Because of
the increasing gap between processor speed and disk latency, file system
performance is largely determined by its disk behavior. Therefore we
primarily focus on the disk I/O aspects of the traces. We find that more
processes access files via the memory-map interface than through the
read interface. However, because many processes memory-map a small set
of files, these files are likely to be cached. We also find that file
access has a bimodal distribution pattern: some files are written
repeatedly without being read; other files are almost exclusively read.
We develop a new metric for measuring file lifetime that accounts for
files that are never deleted. Using this metric, we find that the
average block lifetime for some workloads is significantly longer than
the 30-second write delay used by many file systems. However, all
workloads show lifetime locality: the same files tend to be overwritten
multiple times.
|