Evaluating Distributed File System Performance

Mini Tutorial
Friday, November 13, 2015 - 9:00am-10:30am

Jeff Darcy, Red Hat, Inc.

Jeff Darcy, Red Hat, Inc.

Jeff Darcy has been a UNIX/Linux developer since 1989, with a focus on network and distributed file systems. These have included NFS since version 2, MPFS, Lustre, and—most recently—GlusterFS. Jeff is currently the technical lead for the next major version of GlusterFS, from an undisclosed location at Red Hat.

BibTeX
@conference {208700,
author = {Jeff Darcy},
title = {Evaluating Distributed File System Performance},
year = {2015},
address = {Washington, D.C.},
publisher = {USENIX Association},
month = nov,
}
Download
Description: 

The first part of the tutorial will cover general issues such as the effect of different workloads, measurement pitfalls, and common cheats used by storage vendors. The second part will illustrate how to use common tools such as fio and IOzone to measure storage performance. In the third part, we'll apply this knowledge to two actual distributed file systems—Ceph and GlusterFS—to illustrate the effect and importance of tradeoffs each has made.

Who should attend: 

Storage administrators and developers, site reliability engineers, and anyone else either evaluating alternatives or debugging problems for the storage part of a large computer system.

Take back to work: 

Participants will learn not only how to use performance-testing tools to test two popular distributed file systems, but also how to analyze results. Most importantly, they will learn to recognize anomalies in their own tests, or misleading results from others', so that they can get an accurate picture of each system's capabilities and limitations.

Topics include: 
  • Non-performance requirements for storage
  • Storage workload characterization and sizing
  • Common storage benchmarking mistakes and tricks
  • Usage of IOzone, fio, and other storage-benchmarking tools
  • Basic setup of Ceph and GlusterFS
  • Measuring performance of Ceph and GlusterFS
  • Analysis of benchmark results
  • Suggestions for further analysis and tuning