|
The normalized results across the four workload profiles reveal that the deadline, the noop, and the CFQ schedulers perform within 2% and 1% percent on Ext3 and XFS (see Figure 19-1). On Ext3, the CFQ scheduler has a slight advantage, whereas on XFS the deadline scheduler provides the best aggregate (normalized) response time. On both file systems, the AS scheduler represents the least-efficient solution, trailing the other I/O schedulers by 4.6% and 13% on Ext3 and XFS, respectively. Not surprisingly, among the four workloads benchmarked in a single-disk system, AS trailed the other three I/O schedulers by a significant margin in the web server scenario, which reflects 100% random read operations. On sequential read operations, the AS scheduler outperformed the other three implementations by an average of 130% and 127% on Ext3 and XFS, respectively. The sequential read results clearly support the discussion in this chapter of where the design focus for AS is directed. In the case of sequential write operations, AS reveals the most efficient solution on Ext3, whereas the noop scheduler provides the best throughput on XFS. The performance delta for the sequential write scenarios among the I/O schedulers is 8% on Ext3 and 2% on XFS. Figure 19-1. Aggregate response time (normalized).In Figure 19-1, the x-axis depicts the I/O schedulers; 0 = AS, 1 = deadline, 2 = noop, and 3 = CFQ. The front row reflects the Ext3 setup, whereas the back row shows XFS. The y-axis discloses the aggregate (normalized) response time over the four benchmarked profiles per I/O scheduler. Table 19-1 shows data from a single disk, single CPU.
Next, we will describe a benchmark on a midrange 8-way NUMA RAID-5 system. |
|