IBM's p5 595 Tops the TPC-C Charts

by Timothy Prickett Morgan

As many have been predicting for the better part of a year, IBM's 64-way "Squadron" p5 595 server has overwhelmed the competition on the TPC-C online transaction processing benchmark test, breaking handily through the 3 million transactions per minute (TPM) barrier. The p5 595 is only one of four TPC-C configurations to ever break through 1 million TPM, a club that includes IBM's 32-way pSeries running AIX, Hewlett-Packard's 64-way Superdome running HP-UX, and HP's cluster of four-way Itanium servers running Linux and a cluster-aware relational database.

IBM reported a few weeks ago that a 64-way Squadron box with 1.9 GHz Power5 cores was able to process 3,210,541 TPM, more than three times the work of a prior generation 32-way Regatta pSeries 690 could handle. The p5 595 had twice the main memory (2 TB, if you can believe it) and twice the number of processors as the pSeries 690, and that accounted for probably a little more than 1 million TPM of additional performance. However, the p5 595 database server IBM used at the heart of its TPC-C test had some other tricks. The simultaneous multithreading in the Power5 chips, which is exploited by the new AIX 5L V5.3 release of IBM's Unix, probably added another 30 to 35 percent to overall performance. Moreover, IBM ran the test with a future DB2 8.2 release, which is expected in the middle of May 2005 and which undoubtedly included performance tuning for SMT and other Power5 features. That p5 595 server was equipped with a stunning 243 TB of disk capacity, which cost $14.4 million. The main memory in the server cost just under $9 million, and the 64 processor cores cost just over $2 million. The remaining $1 million in the database server's hardware cost was for the frames and basic peripherals, and on top of this was added another $1.5 million for AIX and DB2. When the TPC-C application servers which simulated 2.56 million end users, and maintenance for the whole shebang were added to the priced configuration, the total cost of the network under test came to $28.6 million, to which IBM offered a 48 percent discount, dropping the price of the system to $16.7 million. That works out to $5.19 per TPM, which, incidentally, is not a lot better than the pricing on the pSeries 690. But that is only because of the ridiculous amount of storage required by the TPC-C test, which skews the pricing.

That p5 595 price/performance, however, is considerably lower than the $8.33 per TPM that HP posted in early 2004 with a 64-way Superdome using 1.5 GHz Itanium 2s and running HP-UX 11i v2 and Oracle 10g Enterprise Edition. If HP could get mx2 dual Madison modules out the door with the new 1.6 GHz/9 MB Madisons, and if HP-UX can scale to 128-way processing in a single system image (and there is not reason to believe it cannot), then HP might be able to break 2 million TPM on the TPC-C test. And it would only be able to reach IBM's Squadron bang for the buck by cutting its list prices for Integrity servers, mx2 processors, main memory, and disks by about 35 percent and then giving a 48 percent discount off list just like IBM and HP have been doing for years on their big iron tests.

IBM, as you might imagine, is very big on this performance leap, which leaves rivals HP, Sun Microsystems, and Fujitsu-Siemens in the dust. "Once in a great while, a technology comes along and changes the competitive landscape," boasted Adalio Sanchez, general manager of IBM's pSeries Unix server division, as he announced the results. He said that this was not just about a benchmark test, but about IBM's demonstration that it is committed to full system designs, not just cranking up the clock speeds on server processors (that barb is aimed at you, Intel), or talk about new, flashy operating systems (that barb's for you, Sun). "To lead in the next generation of Unix, you have to focus on the system," Sanchez says.

This strategy has already worked for IBM, which launched its Power4-based "Regatta" servers three years ago and has eaten 10 points of Unix revenue market share in that time. The Power4 chips were the first processors in the world to put two cores on a single chip, but more importantly, the chip and system design of the Regatta systems allowed IBM to do from two to four times the amount of work per processor core on OLTP workloads as competing Unix system designs. The former makes great public relations, but the latter makes money in a world of CPU-based software pricing. IBM's share gains in Unix are well deserved, and they have prompted competitive responses from Sun and HP. (Sun has shifted to Opterons on the low end and Fujitsu-designed servers at the high-end as it finishes out its massively multicored processors; HP has stayed competitive mainly by cutting prices on its Integrity line when forced to.) IBM is not letting up now that it has attained such performance with the Power5 servers. While Sanchez talked up the p5 595 engineering feat, he forewarned competitors that IBM was "well into" Power6 systems and starting development of Power7 systems. IBM wants to get a 40 percent share of the Unix server market within the next several years, and if IBM can keep selling proprietary iSeries midrange and zSeries mainframe servers at the premium they command compared to Unix servers, there is little to stop IBM from undercutting its Unix rivals in a price war for midrange and big Unix iron.

What IBM does not point out is that the p5 595 performance results, and the impressive bang for the buck they deliver, also make its flagship--or dare we say formerly flagship?--zSeries mainframes look slow and overpriced compared IBM's own Unix gear. IBM does not do TPC-C tests on its mainframes (at least it does not talk about them publicly), but using back-of-the-envelope math, I reckon that a top-end 32-way zSeries 990-332, which is rated at 9,060 MIPS, can do about 450,000 TPM on the TPC-C test. That's it.

The base zSeries 990-332 machine, without disk and memory, costs around $15 million. If you had to beef it up with an appropriate amount of memory and disk, the mainframe hardware might cost $30 million. If you have to add monthly software fees for three years to this machine, it is probably on the order of $50 million for this machine over three years, including maintenance. Mainframes don't have list prices--which used to be against the law for IBM--so it is hard to say for sure. Even if you assume a 50 percent discount, after adding in the software costs, you are talking about $55 per TPM. That's a 10 to 1 price premium. And it will get worse as the Power6 and Power7 generations roll out, unless IBM consolidates the zSeries into one of these future Power-based servers. And that is why many people believe IBM will do just that, as it has already done with its proprietary OS/400-based servers.