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

8
• A 1 Gb dedicated private network between driver and SUT was used to provide sufficient network
bandwidth. (Typically, this can handle up to 10
6
TPM, two orders of magnitude higher than the
TPM measured in these experiments.)
• Because we wanted to compare the relative performance of I/O subsystems (above the level of the
disk array), two high performance HP StorageWorks EVA 3000 disk arrays were connected to the
SUT. Each disk array contained 55 spindles (each 36 GB @15K rpm). RAID 0 was chosen for this
experiment so we could maximize the benefits of a high spindle count and so results were not
limited by disk arrays. In a production environment, RAID10 (Mirroring and Striping) is normally
recommended, for its performance and redundancy characteristics.
• Sufficient SAN bandwidth between SUT and disk arrays was achieved by dedicating two fiber
channel links from SUT to the dual controllers of each EVA 3000.
• The EVA 3000 controllers were upgraded to firmware version CR0C33runp-4004 for active/active
configuration. Each controller had the following: read cache of 512 MB, on and mirrored; write
cache of 256 MB, with write-back; mirror cache of 256 MB; and control cache of 256 MB.
The experimental setup is illustrated in Figure 1. Appendix B contains more details of the
configuration and the test method used.
Figure 1 Experimental configuration used for Oracle 9i benchmark
Test results
Two benchmarks were conducted: the first on Oracle 9.2.07 and the second on Oracle 10.2.0.2.
In both benchmarks, identical I/O subsystem configurations were set up using the same server
hardware and disk array storage configuration for the database server.
The Oracle 10gR2 benchmark used an updated version of the same OLTP application used for the
Oracle 9i benchmark. In the 10gR2 environment, the driver system was a four-way Itanium rx4640
server; in the 9i environment the driver system used was an eight-way PA N4000 server.
In these two benchmarks, environment factors such as kernel parameters, Oracle initialization
parameters, database sizing, and log file sizes were the same. For example, the Oracle System
Global Area was set at 10 GB for both benchmarks.
Driver
System under test
Private network
• Four-way rx4640
• 16GB phys memory
• HP-UX 11.23.0609
MCOE
• Hardware enablement
patches for 11i v2 0609
• Oracle 9i R2 9.2.0.7
• Eight-way
N-4000
• 8 GB physical
memory
• HP-UX 11.23 PA
• HP Stora
g
eWorks EVA 3000
• 55 spindles – RAID0
• 36 GB 15K RPM
• 32 LUNs – db
• 32 LUNs – redo logs
• 2 controllers per EVA
• 1 port connected for each
controller
1GB Ethernet