Oracle Claims Top Cluster Performance on Linux, Windows
September 20, 2002
A report submitted to the TPC (Transaction Processing Performance Council) puts Oracle in the lead in database performance on 32 processor clustered server systems. But you'll have to wait until next spring for the test system that was used to generate the results to become available.
The TPC-C benchmark is used as a standardized method to measure the overall performance (measured in Transactions per Minute, or tpmC) and price performance ratios (system cost/tpmC) of variously configured hardware and software. TPC-C benchmark reports are submitted by members to the TPC organization for review and publication via the TPC Web site.
A report submitted earlier this week and now appearing on the TPC site states that the combination of Oracle 9i R2 Enterprise Edition with Red Hat Linux Advanced Server achieved a tpmC throughput rating of 138,362. With a total test system cost of 2,380,546 (US), the price performance rating rings in at $17.21 per transaction. The report is listed as being "In Review."
The report prompted Oracle to claim their system was "...14% faster and 10% less expensive than Microsoft SQL Server 2000's best result on identical processors and memory," based on a similar TPC-C report with a Windows based system running MS SQL Server.
As reported on Oracle's site, the Microsoft figures were retrieved from a TPC-C report based on a configuration consisting of IBM eServer xSeries 370 servers with 32 Intel Pentium III 900 MHz processors running Microsoft SQL Server 2000 on Windows 2000 Advanced Server. A total throughput of 121,319.23 tpmC was generated on this system with a price performance ratio of $18.97/tpmC.
Oracle has provided DatabaseJournal with a copy of the original report they used, which does in fact verify those figures. However, as of press time a report that duplicated the numbers reported by Oracle could not be located on the TPC site. Instead, the most similar report found there was a report that did indeed result from a clustered 32 processor IBM eServer system running Windows software. In that report, the hardware and software utilized was identical, as was the throughput result of 121,319. But the price performance rating was significantly worse at $25.47/tpmC. When questioned about the discrepancy, Oracle explained that the report they originally used to publish their figures from earlier in the week had since been withdrawn from the TPC site.
The difference between the two reports is significant, as the older Microsoft report (the one still appearing on the TPC site) was based not on the current TPC-C version 5 benchmarks, but instead was based on the previous v3.5 benchmarking specs. The newer Oracle report is based on the v5.0 specs. There are multiple differences between the two specifications, enough to prompt the TPC to warn readers that "Version 5 results may not be compared to results from earlier versions." The report provided by Oracle directly to DatabaseJournal, however, is indeed based on the version 5 specification, and agrees with Oracle's statements: tpmC is 121,319, while $/tpmC is $18.97. Officials from the TPC could not be reached for comment at press time.
In addition to the specification version, other differences between the Oracle and MS reports include the availability of the test system (May 31, 2001 for the MS system, March 5th of next year for the Oracle/Linux system), and the base hardware used: IBM eServer xSeries 370 for the MS results and HP ProLiant DL580-PDC 32P C/S for Oracle.
There is little else, however, that Oracle can use as a valid comparison for their results currently on the TPC site. Of all the reports currently available, only one sports a clustered system with 32 processors running the TPC-C version 5 benchmark. That report is also from HP and features Oracle 9i running on Windows 2000 Advanced Server to the tune of 137,261 tpmC and price/perfomance of $18.46/tpmC.
While an additional IBM/MS entry does come close to the numbers boasted by HP/Oracle (136,767 and price performance of $16.93), the system used to generate these results consisted of 4 8 CPU servers in addition to a separate DTC server, itself with 4 more 700Mhz CPUs, for a combined total of 36 processors in 5 servers.