IO Schedulers and Performance

I/O Schedulers and Performance

As a workload generator, this case study utilizes the flexible file system benchmark (FFSB) infrastructure. FFSB represents a benchmarking environment that facilitates I/O performance analysis by simulating a diverse set of I/O patterns. The benchmarks are executed on multiple individual file systems, utilizing an adjustable number of worker threads, where each thread may operate either out of a combined or a thread-based I/O profile. Aging the file systems, as well as collecting system utilization and throughput statistics, is part of the benchmarking framework. In addition to the more traditional sequential read and sequential write benchmarks, this case study uses a file server, a web server, a mail server, and a metadata-intensive I/O profile (discussed in the preceding section).

The file and mail server workloads are based on the Iometer benchmark from Intel, whereas the mail server transaction mix is loosely derived from the SPECmail2001 I/O profile. The I/O analysis in this case study is composed of two distinct focal points. One emphasis of the study is on aggregate I/O performance achieved across the four benchmarked workload profiles, whereas a second emphasis is on the sequential read and write performance behavior. The emphasis on aggregate performance across the four distinct workload profiles is based on the claim made that an I/O scheduler has to provide adequate performance in a variety of workload scenarios and hardware configurations. All the conducted benchmarks are executed with the default tuning values, if not otherwise specified, in an Ext3 and XFS file system environment. In this chapter, the term response time represents the total runtime of the actual FFSB benchmark, incorporating all the I/O operations that are executed by the worker threads.

The following section discusses the benchmark conducted on the single-CPU single-disk system.