Based on white papers re:Hitachi USP-V best practices from Oracle, I expect to see some response time increases but the performance increases due to cache and the way the USP-V stripes the data across the storage should make the response time issue nearly undetectable. I ran 14 performance tests on multiple tiers of storage - SATA, Fibre Channel (in Raid 5 and 10 configurations) and SSD, each tier was tested using Orion (simple, matrix and dss modes) and Swingbench running on 11Gr2. From the database perspective, the most important item in the results is the Max Transactions per Minute (Max TPM) reported from Swingbench running on an actual 11.2.0.2 Oracle database using ASM. Why ASM? Raw storage means no file system caching, so the results I get are truely from the storage.
I started the first test and right away I noticed the storage wasn’t my bottleneck so I increased the virtual CPU’s in my LPAR to 8 and after test 2, I lowered the db_cache to a tiny 256MB. At this point, when I ran test 3, nearly all the IO was physical IO.
The results of test 3 will be my baseline for future tests.
To see the virtualization overhead and how effective the USP-V cache is, I ran Swingbench on a database with Clariion storage in a USP-V cache “pass-through” configuration:
The next test results show the Clariion behind the USP-V’s massive 512GB cache, all other things equal to the test above. The Clariion has 4GB of cache itself-this is Hitachi cache feeding the Clariion cache, feeding 10K FC spindles.
As you can see above, the Max TPM increased and the avg response time improved. The next test is similar to the previous, but this time using a Hitachi AMS unit loaded with 7.5k sata disks. This was more of a functional test of the AMS unit...obviously you can't compare its performance to the EMC Clariion since the Clariion test above had 10k FC spindles. Ignoring that, I wanted to see a "worst-case" scenario of using sata disks for the databases.
The next set of tests were done on storage internal to the USP-V, virtualized and thin-provisioned with a 4GB clipper (cache partition) for the first 3 and the full 512GB clipper for the last 2:
In this series:
HDS USP-V Overview for Oracle DB
HDS USP-V Performance for Oracle DB
HDS USP-V Features for Oracle DB
HDS USP-V Thin Provisioning for Oracle DB
I started the first test and right away I noticed the storage wasn’t my bottleneck so I increased the virtual CPU’s in my LPAR to 8 and after test 2, I lowered the db_cache to a tiny 256MB. At this point, when I ran test 3, nearly all the IO was physical IO.
Test Number | Storage | Spindle Type | Raid Level | Virtualized | Cache(GB) | IOPS | MBPS | Max TPM | Max TPS | Avg TPS | Avg Resp | Disk Busy % | CPU% |
1 | Clariion FC | FC | 5 | No | 4 | 1403 | 117.46 | 69413 | 1294 | 1083 | 65 | 30 | 95 |
2 | Clariion FC | FC | 5 | No | 4 | 1403 | 117.46 | 102911 | 2182 | 1450 | 51 | 24 | 87 |
3 | Clariion FC | FC | 5 | No | 4 | 1403 | 117.46 | 73011 | 1368 | 1087 | 66 | 100 | 84 |
The results of test 3 will be my baseline for future tests.
To see the virtualization overhead and how effective the USP-V cache is, I ran Swingbench on a database with Clariion storage in a USP-V cache “pass-through” configuration:
Test Number | Storage | Spindle Type | Raid Level | Virtualized | Cache(GB) | IOPS | MBPS | Max TPM | Max TPS | Avg TPS | Avg Resp | Disk Busy % | CPU% |
4 | Clariion | FC | 5 | Yes | pass-thru | N/A | N/A | 62895 | 1190 | 984 | 78 | 42 | 71.4 |
The next test results show the Clariion behind the USP-V’s massive 512GB cache, all other things equal to the test above. The Clariion has 4GB of cache itself-this is Hitachi cache feeding the Clariion cache, feeding 10K FC spindles.
Test Number | Storage | Spindle Type | Raid Level | Virtualized | Cache(GB) | IOPS | MBPS | Max TPM | Max TPS | Avg TPS | Avg Resp | Disk Busy % | CPU% |
5 | Clariion | FC | 5 | Yes | 512 | 1207 | 120.28 | 72666 | 1319 | 1087 | 65 | 94 | 85 |
As you can see above, the Max TPM increased and the avg response time improved. The next test is similar to the previous, but this time using a Hitachi AMS unit loaded with 7.5k sata disks. This was more of a functional test of the AMS unit...obviously you can't compare its performance to the EMC Clariion since the Clariion test above had 10k FC spindles. Ignoring that, I wanted to see a "worst-case" scenario of using sata disks for the databases.
Test Number | Storage | Spindle Type | Raid Level | Virtualized | Cache(GB) | IOPS | MBPS | Max TPM | Max TPS | Avg TPS | Avg Resp | Disk Busy % | CPU% |
6 | AMS | SATA | 5 | Yes | 32 | 428 | 53.96 | 31853 | 963 | 476 | 149 | 100 | 85 |
The next set of tests were done on storage internal to the USP-V, virtualized and thin-provisioned with a 4GB clipper (cache partition) for the first 3 and the full 512GB clipper for the last 2:
Test Number | Storage | Spindle Type | Raid Level | Virtualized | Cache(GB) | IOPS | MBPS | Max TPM | Max TPS | Avg TPS | Avg Resp | Disk Busy % | CPU% |
7 | Internal | FC | 5 | Yes | 4 | 1396 | 133.12 | 64168 | 1155 | 954 | 75 | 35 | 69.1 |
8 | Internal | FC | 10 | Yes | 4 | 2077 | 179.77 | 66451 | 1185 | 979 | 73 | 45 | 70.6 |
9 | Internal | SSD | 10 | Yes | 4 | 5122 | 214.70 | 64044 | 1171 | 966 | 75 | 43 | 69.9 |
10 | Internal | FC | 10 | Yes | 512 | 9486 | 308.92 | 66277 | 1166 | 956 | 74 | 45 | 71.3 |
11 | Internal | SSD | 5 | Yes | 512 | 8060 | 252.61 | 58844 | 110 | 935 | 81 | 43 | 72.8 |
In this series:
HDS USP-V Overview for Oracle DB
HDS USP-V Performance for Oracle DB
HDS USP-V Features for Oracle DB
HDS USP-V Thin Provisioning for Oracle DB
No comments:
Post a Comment