The Four Hundred - pSeries Power4+ Tops TPC-C, but Where's iSeries?

A few weeks ago, when the "Madison" Itanium 2 processors were announced by Intel, the company spent a lot of time bashing IBM's Power4+ line of pSeries RISC/Unix servers. These processors are without question the touchstone against that Madison--and indeed any RISC processor--is measured against these days. IBM knew Intel was going to do this, so it did something about it.

Sources at IBM had been saying for weeks that the company was gearing up to deliver updated results on the TPC-C online transaction processing benchmark test that would knock rival Hewlett-Packard's 64-way Integrity Superdome machines from the top spot. The word I kept hearing was IBM would beat 720,000 transactions per minute (TPM), about 6 percent more oomph than its initial test on the updated 32-way pSeries "Regatta-H" servers, which now use 1.7 GHz Power4+ processors and can support up to 512 GB of main memory. IBM and HP have been leapfrogging each other on these tests since late last year, and both are determined to hold the top spot. IBM reckoned that there was going to be a lot of Power4+ bashing by Intel and HP, and so its performance gurus in its Austin, Texas, pSeries performance labs worked all through the weekend, tuning AIX and DB2 and configuring a new storage subsystem to the Regatta box and pushed the performance of the 32-way server to 763,898 TPM. This was, to say the least, unexpected--and very good timing. Some people at IBM on the pSeries team are going to get nice bonuses because Big Blue stole at least some of Intel's and HP's thunder on Madison announcement day. But the advent of these improved Regatta machines for the pSeries and some other pending iSeries announcements slated for later this year (see the separate story in this issue for more on those) lead me to believe that IBM will debut Power4+ iSeries machines before the year is out.

Before we get into that, let's examine the pSeries TPC-C tests. Both the new and old Regatta-H machines IBM has tested recently were configured with AIX 5L V5.2 and IBM's DB2 Universal Database 8.1 software. The Regatta-H server had 32 1.7 GHz Power4+ processor cores (with a 1.5 MB shared, on-chip L2 cache for every two cores), 512 MB of L3 cache, and 512 GB of main memory. The initial configuration from May had 43 TB of IBM's 7103 SSA disk arrays, but this time around IBM could do the test using only 33.3 TB of its FAStT EXP 900 disk arrays. The SSA disk subsystems used in the original test back in May cost $6.86 million at list price, while the FAStT 900 arrays only cost $3.44 million. IBM says that the SCSI disks did not do all that much to help performance, but such a radical cut in price had a big impact on bang for the buck.

In the TPC-C tests, the prices of the Regatta-H servers were the same at $3.27 million, with $1.38 million going for main memory alone and $1.5 million going for processors. AIX and DB2 for the server cost $632,725. With application servers and three years of maintenance thrown in, the whole TPC-C configuration had a list price of $14.57 million for the machine doing 680,613 TPM, but after a 48 percent discount, the price of the whole shebang dropped to $7.57 million or $11.13 per TPM after the discounts. With the latest result, which appears to be driven mostly by tuning the DB2 database, the Regatta-H server with the FAStT 900 arrays cost only $10.7 million at list price. But IBM tossed in a 41 percent large systems discount, which dropped the cost of the machine down to $8.31 per TPM.

The 64-way Madison-based Integrity Superdome HP has tested running Windows 2003 most recently came in at 707,102 TPM at a cost of $9.13 per TPM after a 38.5 percent discount for large systems buyers. With this being the first 64-way Windows-based machine on the market, HP might be able to catch IBM. But then again, it may be relying on HP-UX, which HP sources say will perform better than Windows 2003 on any Integrity machine, to meet or beat Big Blue's best result on a high-end server. Even a 10 percent performance advantage with HP-UX over Windows 2003 would put the Superdome ahead of IBM's Regatta-H in terms of raw performance on the TPC-C test. HP may just be holding out to see what IBM would be able to do before running and publishing its HP-UX results.

HP might also be waiting until it can demonstrate the performance of HP-UX running on the 128-way PA-8800 versions of the Superdome machines, which are expected late this year or early next year. The PA-8800, like the Power4 and Power4+ processors from IBM, puts two cores and shared cache memory on a single chip. These two-core processors will plug into the 64-wau Superdome chassis, yielding a server with effectively 128 processors. (IBM's 32-way way Regatta-H server is really a 16-way server with dual-core processors.) It's hard to say where such a Superdome might be in terms of performance, but 900,000 to 1 million TPM is the ballpark. But the game isn't over there.

As we reported back in February, IBM's 64-way "Squadron" Power5 servers, due sometime in the spring of 2004, are expected to have about four times the performance of the first generation 32-way pSeries Regatta-H boxes using the 1.3 GHz Power4 processors. That's about 1.6 million TPM. Back in February, the IBM executives who first talked about that 4X multiple gave the impression that the performance leap was compared to the then-current results posted on the 1.3 GHz Power4 Regatta-H machines using Oracle databases, which suggested that maybe 2 million TPM was possible. IBM has since that comparison.

Based on the past performance of S-Star and Power4 servers running OS/400 and its integrated DB2/400 database, a 32-way Regatta using the new 1.7 GHz Power4+ processors and the High Speed Link-2 remote I/O connections (which is probably the same as, if not very similar to, the RIO-2 connection used in the pSeries line) should be able to hit about 625,000 TPM based on the latest benchmark for the pSeries versions of this machine. There is a big performance difference on IBM Power4-based servers when comparing Oracle or DB2 running on AIX and DB2/400 running on OS/400. IBM's DB2 database running on Unix has traditionally hit somewhere in the middle of these two, in terms of raw TPC-C performance, but IBM seems to have gotten DB2 on par with Oracle on AIX machines. IBM may yet do this with DB2/400 on OS/400. We won't know until IBM debuts these Power4+ machines, which I think it could do by the end of the year. (I have not heard any rumors that it is going to do this. It is more of a gut feeling.)

With the pSeries Squadron machines and their Power5 processors slated for April or May 2005 and the Power4+ chips probably shipping in volume in November or December to a magnitude that would be sufficient to provide chips for both the pSeries and iSeries line, this would be a logical time to do a high-end iSeries announcement. Such a big iSeries Regatta-H machine would affect very few customers, but could bring in hundreds of millions of dollars in extra sales in the fourth quarter of 2003. IBM may hold out and just wait for the Power5 Squadron machines in the spring of 2004. IBM never tested the original iSeries Regatta-H servers on the TPC-C test (which is bad behavior on the part of IBM, to be frank), but a 32-way with 1.3 GHz processors probably would have handled about 350,000 to 360,000 TPM. So it looks like a 64-way iSeries Squadron box should be able to hit around 1.3 million TPM, and maybe even a little more.