Fujitsu-Siemens Proves PrimePower's Performance

by Timothy Prickett Morgan

Server makers Fujitsu and Siemens combined their IT units in October 1999 to create Fujitsu-Siemens and a unified line of computers that spanned from laptops to supercomputers. The Sparc64-based and Solaris-compatible PrimePower line of RISC/Unix servers are the crown jewel in their partnership. While Fujitsu-Siemens has run many benchmarks on these machines, it has resisted running a TPC-C online transaction processing test, using the Oracle database, to allow comparisons with other big boxes. That just changed.

In the RISC/Unix server market, there are only two databases that matter: DB2 running on AIX and Oracle running on AIX and every other Unix flavor. While Fujitsu-Siemens is to be commended for being an enthusiastic supporter of the TPC-C test throughout the five-year life of the PrimePowers, which span up to 128 processors in a single-system image and support Solaris 8 and Solaris 9, the PrimePowers generally have been put through the TPC-C paces running Fujitsu's own SymfoWare database. This is exactly what IBM does--running its own databases on its own iron--so what's the big deal, you ask? Well, there is a wealth of performance information about DB2 running on all kinds of machines, and everyone in the Unix business knows that Oracle performed better on AIX boxes than DB2. (IBM has closed that performance gap somewhat with DB2 8.1, compared with Oracle 10g.) No one knows who uses SymfoWare in production, and aside from the TPC-C tests, there is no public benchmark information about it. So assessing how Oracle would run on the PrimePower boxes with the TPC-C test compared with SymfoWare was not possible. Fujitsu-Siemens had a good idea, and even had internal documents that it shared with customers to show them what to expect when using Oracle8 and Oracle9i. But, publicly, it said nothing.

With its latest TPC-C benchmark on its 1.3 GHz Sparc64-V processors used in the PrimePower 2500 servers, Fujitsu-Siemens has pretty much put the issue to rest. A 64-way version of the PrimePower 2500 was tested running Solaris 8 and the new Oracle10g Enterprise Edition database, and the Oracle results are consistent with what many had expected: that Oracle does seem to run a little bit more slowly on the boxes compared with SymfoWare (if you trust performance extrapolations in the absence of hard data, that is). But the difference is not really that significant, as it turns out. It is in the same realm as the performance benefit that Oracle has over DB2 on AIX boxes, in fact.

In any event, Fujitsu-Siemens equipped the PrimePower 2500 with only half of its full complement of processors, and then added 512 GB of main memory to the box and 57.6 TB of disk capacity. The PrimePower server was able to handle 595,702 transactions per minute (TPM). After a 19 percent discount dropped the price of the configured server to $7.4 million, this PrimePower 2500 running Oracle10g delivered a price/performance of $12.43. When rivals IBM and Hewlett-Packard do their TPC-C tests on big iron, their discounts run into the 40 to 50 percent range, which is why they show better bang for the buck.

By my estimates, a 64-way machine, configured the same way but running SymfoWare, would have delivered approximately 12 percent more oomph and a little bit better than 15 percent price/performance, because SymfoWare is a little less expensive than Oracle10g. I also reckon that a 128-way machine running Oracle10g and using the new 1.3 GHz Sparc-V chips would have broken the 1 million TPM barrier, which both HP and IBM have done with their biggest boxes. Why would Fujitsu-Siemens hold back? It may not yet be able to support 1TB of main memory, which the 128-way configuration probably would require to run in a balanced fashion.

Fujitsu-Siemens held the top spot in the TPC-C test for a long time, but IBM and HP have been slugging it out for the top spot. Earlier this year, IBM launched the 1.9 GHz Power4+ processors for its 32-way pSeries 690 servers, and it ran the TPC-C test on this box with AIX 5.2 and DB2 8.1 to take the top spot. This server was configured with 1 TB of main memory and 74 TB of disk capacity, with the whole setup costing $10.6 million and delivering 1,025,486 TPM. After a 47 percent discount, this pSeries 690 was able to show a price/performance of $5.43 per TPM.

In November 2003, HP was the first vendor to break through the 1 million TPM ceiling when it delivered a TPC-C benchmark result for its 64-way Integrity Superdome servers running its HP-UX operating system and Oracle10g. That Superdome machine used the 1.5 GHz "Madison" variant of the 64-bit Itanium processors from Intel. That machine was also configured with 1 TB of main memory and 38.3 TB of disk capacity, which allowed it to deliver a performance of 1,008,145 TPM at a cost of $8.33 per TPM after a 48 percent discount.

It is hard to say how much the main memory and the CPU count are affecting performance on this big box. It might turn out that a 64-way PrimePower with 1 TB of main memory can do 1 million TPM, just like a 64-way Itanium 2 machine can. But, even if that is the case, IBM is delivering the most powerful processors in the world and would need only half the processor count of an HP or Fujitsu-Siemens box to deliver the same throughput. This is important when software pricing is based on processor count, not system performance.