Trace Based Analysis of File System Effects on Disk I/O

In this paper, we present a study of low-level disk access patterns for three workloads for a variety of file systems. While there has been a great deal of work on disk architectures, there has been very little work measuring actual low-level disk access. We have extended previous work to also consider different file systems to examine their effect on the workload. The analysis is done based on traces of disk access that have been captured at the driver level just before the data is transferred to disk. The workloads we examined were an email server, file server emulation and a database server and each workload was regenerated for different file systems. We found that depending on the workload, the access patterns could be read dominated or write dominated. The frequency of access in some cases was very high – with inter-arrival times well under 5ms. We also examined the burstiness behavior and read/write distribution. We found that the choice of file system was for the most part insignificant but in some cases can greatly influence the access characteristics. We believe the resulting traces and summary analysis can be used by storage system designers in their analysis of file system and disk organization designs. The consideration of file system also allows storage system designers to consider the style of file system when making design choices.