Improving the performance of single instance Oracle on file systems, January 2008

Oracle 10.2.0.2 results summary

The second benchmark was performed using Oracle 10.2.0.2. The combined results for four

configurations are shown in Figure 8.

Figure 8 Oracle 10.2.0.2 – Results summary with results normalized to LVM=100%

Results

This review did not intend to use specially modified versions, tuned for optimal performance, to

achieve the highest possible TPM numbers. Hence, off-the-shelf components were used (for example,

the standard operating environment and standard versions of Oracle databases). By carefully

removing bottlenecks and keeping the test configuration consistent, the goal was to compare the

relative performance of each I/O subsystem.

Both benchmarks show the improvement in TPM when ODM is used for Online JFS with single-

instance Oracle.

In a file system without ODM, a small number of data files limit the parallelism that can be achieved

by Oracle processes to perform useful work. Without ODM, even when direct I/O is configured using

Online JFS mount options, HP-UX write file-locking serializes I/O activity, limiting performance. When

ODM is enabled, there is improved performance resulting from asynchronous I/O in the file system.

In Experiment 3, we could not drive down CPU idle time very far. The CPU %idle leveled out at

78%, showing that Oracle processes are mostly blocked waiting for I/Os to complete. Performance

was degraded by about 92% when ODM was not used, or 94% relative to the raw volumes on

VxVM configuration, at a workload of 80 clients.