IBM's eServer p5s Rock the TPC-C Benchmark

by Timothy Prickett Morgan

The performance and price/performance games in the Unix server market have always been more like leapfrog than the horse races, which the Unix players continuously make analogies to. With its Power5 "Squadron" servers, IBM has jumped out in front again, as it did three years ago with its "Regatta-H" Power4-based machines. The comparison is important because the Regatta machines allowed IBM to eat Unix market share like crazy as rivals Hewlett-Packard and Sun Microsystems struggled with product transitions and other business issues.

Regardless of what the naysayers say, the TPC-C online transaction processing benchmark test is an important performance and price/performance proof point, along with the SAP, Oracle, and PeopleSoft application tests and the many tests put out by the Standard Performance Evaluation Corp. And once again IBM is jumping ahead of its rivals in the midrange and high end of the server market, in terms of bang for the buck and raw performance.

IBM has not yet run TPC-C tests on its two-way p5 520 or four-way p5 550 servers, but it has run tests on four different configurations of its midrange p5 570 server, which scales up to 16 processor cores in a single system image. To demonstrate that it is getting a significant performance boost from using its own DB2 8.1 database, compared with running Oracle's 10g database on its machines, IBM ran two benchmarks (on a four-way and on an eight-way configuration) on Oracle 10g and two other benchmarks (on an eight-way and on a 16-way) on DB2 8.1

On the four-way p5 570 configuration, which had four 1.9 GHz Power5 cores activated, IBM was able to crank through 194,391 transactions per minute (TPM) on a machine equipped with AIX 5L V5.3 and Oracle 10g in a three-tiered configuration. (In the three-tiered TPC-C test, which is pretty much the only one anyone uses anymore, the system under test only runs as a database server for the TPC-C transaction processing test, which simulates the OLTP environment you might expect to find in a warehouse distribution operation.) The server, equipped with 128 GB of main memory, cost $527,000; the 15.2 TB of disk storage used in the test cost another $1.05 million. Adding in operating system, database, and other systems software, plus client hardware (a bunch of X86 boxes running Windows 2000 Server and Visual Studio.NET), and support for the whole shebang, cost $1.79 million. After a 39 percent discount on the setup, which included the prepayment for three years of maintenance and system discounts, the price of the configuration was $1.09 million, or $5.62 per TPM. This time last year, the cost of a Regatta box in the same power class was around $8 per TPM, and that was after discounts approaching 50 percent.

On the eight-way p5 570 (again using 1.9 GHz cores) with 256 GB of main memory and 25.5 TB of disk, IBM was able to handle 371,044 TPM running Oracle 10g, at a cost of $5.26 per TPM, while a setup running DB2 8.1 with 256 GB of main memory and 35.2 TB of disk could pump out 429,900 at a cost of $4.99 per TPM. That extra disk space added a few hundred thousand dollars to the cost of the system under test, but the increase in performance was enough to still see the DB2 machine offer a lot better bang for the buck. DB2 showed a 16 percent performance advantage on this machine, and if IBM had been able to hold disk capacity constant (the TPC-C test requires disk storage to scale with the number of users), it would have been able to demonstrate about a 14 percent price/performance benefit of using DB2 8.1 over Oracle 10g. Both machines had about the same discount, too. So IBM wasn't doing anything weird on that front to make DB2 look better than Oracle.

On the top-end, 16-way p5 570, IBM installed AIX 5L V5.3 and DB2 8.1, crammed 512 GB of main memory into the box (as much as IBM and rivals were able to support last year on 32-way and 64-way servers), and pushed out 809,144 TPM, at a cost of $4.95 per TPM. This server had an unbelievable 614.3 TB of disk capacity, which is an absurd amount.

In fact, the TPC-C tests demonstrate the old adage that disk arm count, not disk capacity, is one of the main determining factors in OLTP performance. Other vendors add lots of disk arms to their TPC-C systems as well, for just this reason. But it would be interesting to see, for instance, what a p5 570 with 16 cores activated would do, in terms of TPC-C performance, with fewer than 1,640, 15K RPM Fibre Channel disk drives whirring around under their frames. Say, with something closer to 5 TB or even 10 TB of disk capacity, which would still be a lot of capacity for a database server. Perhaps disk capacity is as far ahead of server needs as memory capacity is behind it.