Workload Dependent Performance Evaluation of the Linux 2.6 I/O Schedulers

The 2.6 release introduced the option to select a particular I/O scheduler at boot time. The 2.4 Linus elevator was retired, incorporated are now the anticipatory (AS), the deadline, the noop, as well as the completely fair queuing (CFQ) I/O schedulers. Each scheduler has its strengths and weaknesses. The question is under what workload scenarios does a particular I/O scheduler excel, as well as what is the performance gain that is possible by utilizing the available tuning options.

This study quantifies the performance of the 4 I/O schedulers under various workload scenarios (such as mail, web, and file server based conditions). The hardware is being varied from a single-CPU single-disk setup to machines with many CPUs that are utilizing large RAID arrays. In addition to characterizing the performance behavior and making actual recommendations on which scheduler to utilize under certain workload scenarios, the study looks into ways to actually improve the performance through either the existing tuning options or any potential code changes/enhancements.


Download PDF.


So, it seems to suggest that RAID-5 performance with XFS and the AS scheduler is quite poor. I wonder if software-based RAID-5 would be effected similarly?