|
Utilizing the 28-disk RAID-0 configuration as the benchmark environment reveals that across the four workload profiles, the deadline implementation can outperform the other three schedulers. However, the CFQ, as well as the noop scheduler, slightly outperforms the deadline implementation in three out of the four benchmarks. Overall, the deadline scheduler gains a substantial lead processing the web server profile, outperforming the other three implementations by up to 62%. On Ext3, the noop scheduler reflects the most efficient solution while operating on sequential read and write requests, whereas on XFS, CFQ and deadline dominate the sequential read and write benchmarks. The performance delta among the schedulers for the four profiles is much more noticeable on XFS (38%) than on Ext3 (6%), which reflects a similar behavior as encountered on the RAID-5 setup. Increasing nr_requests to 2,560 on the RAID-0 system leads to inconclusive results for all the I/O schedulers on Ext3 as well as XFS. Table 19-4 shows data from a RAID 0 16-way setup.
As we continue to the next section, we further illustrate the performance behavior of the AS scheduler design that views the I/O subsystem based on a notion that an I/O device has only one physical (seek) head; this study analyzes the sequential read performance in different hardware setups. The results are compared to the CFQ scheduler. |
|