AS Versus Deadline Performance

team bbl


In the case of tuning AS to approach deadline behavior, the tuning consists of setting nr_requests to 2,560, antic_expire to 0, read_batch_expire to 1,000, read_expire to 500, write_batch_expire to 250, and write_expire to 5,000. Setting the antic_expire value to 0 (by design) disables the anticipatory portion of the scheduler. The benchmarks are executed utilizing the RAID-5 environment, and the results are compared to the deadline performance results reported in this study. On Ext3, the non-tuned AS version trails the nontuned deadline setup by approximately 4.5% across the four profiles. Tuning the AS scheduler results in a substantial performance boost, because the benchmark results reveal that the tuned AS implementation outperforms the default deadline setup by approximately 6.5%. The performance advantage is offset, though, while comparing the tuned AS solution against the deadline environment with nr_requests set to 2,560. Across the four workload profiles, deadline again outperforms the AS implementation by approximately 17%. As anticipated, setting antic_expire to 0 results in lower sequential read performance, stabilizing the response time at deadline performance. On XFS, the results are based on the erratic metadata performance behavior of AS, which is inconclusive. One of the conclusions is based on the current implementation of the AS code that collects the statistical data; the implemented heuristic is not flexible enough to detect any prolonged random I/O behavior. This is a scenario where it is necessary to deactivate the active wait behavior. Further, setting antic_expire to 0 should force the scheduler into deadline behavior, a claim that is not backed up by the empirical data collected for this study. One explanation for the discrepancy is that the short backward seek operations supported in AS are not part of the deadline framework. Therefore, depending on the actual physical disk scheduling policy, the AS backward seek operations may be counterproductive from a performance perspective.

Table 19-6 shows data from a RAID-5 8-way with default AS, default deadline, and tuned AS comparison.

Table 19-6. RAID-5 8-WayDefault AS, Default Deadline, and Tuned AS ComparisonMean Response Time in Seconds
 

AS Ext3

DL Ext3

AS Tuned Ext3

AS - XFS

DL - XFS

AS Tuned - XFS

File Server

77.2

81.2

72.1

83.8

90.3

84.5

Metadata

147.8

148.4

133.7

205.8

90.8

187.4

Web Server

70.2

58.4

62

82.1

81.3

75.9

Mail Server

119.2

114.8

103.5

153.9

92.1

140.2

Seq. Read

517.5

631.1

634.5

515.8

624.4

614.1

Seq. Write

1033.2

843.7

923.4

426.6

422.3

389.1


This benchmark study has revealed that the tuned CFQ setup is the proven remedy for the RAID-5 environment.

    team bbl



    Performance Tuning for Linux Servers
    Performance Tuning for Linux Servers
    ISBN: 0137136285
    EAN: 2147483647
    Year: 2006
    Pages: 254

    flylib.com © 2008-2017.
    If you may any questions please contact us: flylib@qtcs.net