8-Way RAID-5 Setup

In the RAID-5 environment, the normalized response time values, across all four profiles, disclose that the deadline scheduler provides the most efficient solution on Ext3 as well as XFS (see Figures 19-2 and 19-3). While executing in an Ext3 environment, all four I/O schedulers are within 4.5%, with the AS I/O scheduler trailing noop and CFQ by approximately 2.5%. On XFS, the study clearly discloses a profound AS I/O inefficiency while executing the metadata benchmark. The deltas among the schedulers on XFS are much larger than on Ext3, as the CFQ, noop, and AS implementations trail the deadline scheduler by 1%, 6%, and 145%, respectively. As seen in the single-disk setup, the AS scheduler provides the most efficient sequential read performance. The gap between AS and the other three implementations is reduced significantly compared to the single-disk scenarios. The average sequential read throughput for the other three schedulers is approximately 20% less on both Ext3 and XFS, respectively. The sequential write performance is dominated by the response time of the CFQ scheduler that outperformed the other three solutions. The delta between the most- (CFQ) and least-efficient implementation is 22% (AS) and 15% (noop) on Ext3 and XFS, respectively.

Figure 19-2. Ext3aggregate response time (normalized).

Figure 19-3. XFSaggregate response time (normalized).

In Figure 19-2, the x-axis depicts the I/O schedulers; 0 = AS, 1 = noop, 2 = deadline, and 3 = CFQ (for Ext3). The front row reflects the nontuned environments, and the back row reflects the tuned environments. The y-axis discloses the normalized response time (over the four profiles) per I/O scheduler.

In Figure 19-3, the x-axis depicts the I/O schedulers; 0 = AS, 1 = noop, 2 = deadline, and 3 = CFQ (for XFS). The front row reflects the nontuned environments, and the back row reflects the tuned environments. The y-axis discloses the normalized response time (over the four profiles) per I/O scheduler.

Table 19-2 shows data from a RAID-5 8-way setup with no tuning.

Table 19-2. RAID-5 8-Way SetupNo TuningMean Response Time in Seconds (AS, Deadline, Noop, CFQ)

	AS - Ext3	DL - Ext3	NO - Ext3	CFQ - Ext3	AS - XFS	DL - XFS	NO - XFS	CFQ - XFS
File Server	77.2	81.2	86.5	82.7	83.8	90.3	96.6	90.7
Metadata	147.8	148.4	133	145.3	205.8	90.8	101.6	100.8
Web Server	70.2	58.4	66.2	59.2	82.1	81.3	78.8	75.2
Mail Server	119.2	114.8	115.3	119.3	153.9	92.1	100.7	92.2
Seq. Read	517.5	631.1	654.1	583.5	515.8	624.4	628.7	604.5
Seq. Write	1033.2	843.7	969.5	840.5	426.6	422.3	462.6	400.4

In the second phase, all the I/O scheduler setups are tuned by adjusting the tunable nr_requests from its default value of 128 to 2,560. The results reveal that the CFQ scheduler reacts in a positive way to the adjustment and can provide an efficient solution on Ext3 as well as on XFS. The tuning results in decreasing the response time for CFQ in all the conducted benchmarks on both file systems. Although CFQ benefits from the tuning, the results for the other three implementations are inconclusive. Based on the profile, the tuning results in either a gain or loss in performance. Because CFQ is designed to operate on larger sets of I/O requests, the results reflect the scheduler's design goals. This is in contrast to the AS implementation, where by design, any read intensive workload cannot directly benefit from the change. On the other hand, for sequential write operations, AS can take advantage of the tuning as the response time decreases by 7% and 8% on Ext3 and XFS, respectively. The conducted benchmarks reveal another significant inefficiency behavior in the I/O subsystem, because the write performance for all the schedulers on Ext3 is significantly lower (by a factor of approximately 2.1) than on XFS. The issue here is the Ext3 reservation code. The Ext3 patches to resolve the issue are available at http://www.kernel.org.

Table 19-3 shows data from a RAID-5 8-way setup with nr_requests.

Table 19-3. RAID-5 8-Way Setupnr_requests = 2,560Mean Response Time in Seconds (AS, Deadline, Noop, CFQ)

	AS - Ext3	DL - Ext3	NO - Ext3	CFQ - Ext3	AS - XFS	DL - XFS	NO - XFS	CFQ - XFS
File Server	78.3	72.1	87.1	70.7	94.1	75	89.2	76
Metadata	127.1	133	137.3	124.9	189.1	101.1	104.6	99.3
Web Server	62.4	58.8	75.3	57.5	79.4	72.83	80.6	71.7
Mail Server	110.2	92.9	118.8	99.6	152.5	100.2	95.1	81
Seq. Read	523.8	586.2	585.3	618.7	518.5	594.8	580.7	594.4
Seq. Write	968.2	782.9	1757.8	813.2	394.3	395.6	549.9	436.4

Finally, a benchmark was performed on a 16-way SMP system that utilizes a 28-disk RAID-0 configuration.