The Four Hundred

At last week's Intel Developer Forum, Japanese server maker NEC took the wraps off a 32-way server running Microsoft's forthcoming "Whistler" Windows .NET Server 2003 version of Datacenter Server, based on the "McKinley" Itanium 2 processor. The results of the TPC-C online transaction processing benchmark put the Wintel platform in the same power class as the enterprise Unix and proprietary platforms (like OS/400) that have been ahead of Windows since Windows NT 3.1 debuted, in 1993.

The NEC TX7 series of computers are based on the "AzusA" chipset that the company created for the Itanium family of 64-bit processors from Intel. A 16-way version of the AzusA machines debuted with the 733 MHz and 800 MHz "Merced" Itaniums, the first generation of 64-bit chips from Intel. These 16-way servers were sold by NEC in Japan, and by Hewlett-Packard in North America and Europe, running HP's HP-UX 11i variant of the Unix operating system or the open-source Linux operating system. According to HP, at the time they were launched last summer, the 16-way AzusA servers were rated at about 140,000 transactions per minute on the TPC-C test (this is an estimate, not an actual benchmark result) and at about 51 gigaflops (that's billions of floating point operations per second). That puts the first generation of AzusA servers in the same power class as the most powerful Unix and proprietary servers (such as IBM's AS/400 and iSeries line and its S/390 and zSeries mainframe line) of early 2000, when Unix machines with 24 or 32 processors were able to hit around 150,000 TPM.

Intel said that servers using the Itanium 2 processors should yield between 1.5 and 2 times the performance of servers using the original Itaniums because of various improvements in the chip. It is hard to say whether or not this is the case with the NEC TX7/AsuzA servers, since NEC has not shown a like-for-like comparison. The original 16-way machines that got that 140,000 TPM estimated rating were running HP-UX and the eponymous database from Oracle, not the future 64-bit and NUMA-enabled versions of Windows .NET Server 2003 Datacenter Edition and SQL Server 2000. The AsuzA machines tested this week also had 32 processors, instead of 16 in the early generation. Based on very skinny Intel estimates, you might think that an AzusA server with 16 1-GHz Itanium 2 processors should be able to handle between 200,000 and 250,000 TPM running Unix, and with 32 processors, maybe as much as 300,000 or 350,000 TPM. You might also guess that that Windows .NET Server would be able to perform a little lower than this level, given its relative inexperience with high-end SMP (symmetric multiprocessing) scaling, unlike Unix, which has been scaling up to this number of processors for about five years or so, depending on the platform.

That would be a pretty good guess. NEC's test results on the 32-way TX7 server, which supports HP-UX and Linux today and will support Windows .NET Server 2003 by the end of the year, posted an impressive 308,621 TPM on the TPC-C test, at a cost of $14.96 per TPM. That NEC TX7 server was equipped with Itanium 2 chips with 3 MB L3 caches, 256 GB of main memory, and 23.6 TB of disk capacity. That server cost around $1.4 million, including the license for Datacenter Server 2003, and the storage cost just under $2 million. SQL Server cost another $529,312, and client hardware and three years of support brought the total cost to $5.3 million, a figure that dropped to $4.6 million after a 13 percent discount. This is Unix-class performance at a substantial price break, compared with Unix.

For instance, HP just announced that its 64-way Superdome server using PA-8700+ processors, using 875 MHz chips and running on HP-UX and Oracle9i, could crank through 423,414 TPM at a cost of $15.64 per TPM. That machine had 256 GB of main memory and 15 TB of disk capacity. However, that Superdome server and its storage cost over $9.1 million, Oracle cost nearly $1.3 million, and the whole shebang cost about $12.6 million, at list price. It was only after a staggering 44 percent discount that HP could get the cost of the Superdome configuration down to the $15.64 per TPM price/performance rating. You can bet that upgrades on Superdome machines are not discounted at this level. Similarly, IBM's 32-way pSeries 690 "Regatta-H" server using 1.3 GHz Power4 processors was able to handle 403,255 TPM on the TPC-C test when equipped with AIX 5L 5.2 and Oracle 9i on a machine rigged with 256 GB of main memory and 20 TB of disk capacity. The server and storage in this Regatta-H box cost $11.3 million, and Oracle cost another $1.5 million. The total price tag for the TPC-C n-tier network under test came to $15.8 million. Like HP, IBM offered a big discount on this Regatta-H configuration, to show better bang for the buck, in this case a 49 percent discount. Just try to get that price on a Regatta-H upgrade and see how far you get with IBM.

With the same level of discounting as the NEC 32-way AzusA box using Itanium 2 processors, the HP machine delivers about $26 per TPM and the IBM machine yields about $34 per TPM. This is obviously not as good of a price/performance as the NEC box is delivering--not by a long shot. The fact that the HP box has twice as many processors as the NEC machine, yet delivers only 37 percent more performance, is something that will weigh heavily on daring customers who want to move to big Wintel iron, since application software and middleware are often priced on a per-processor basis these days. And here's an even funnier thought: If Intel could crank up the clock speed on the Itanium 2 processor to the same 1.3 GHz that IBM is using for the Power4 chips in the Regatta-H machines, and if the AzusA chipset and its crossbar switch could deliver the bandwidth, then the AzusA machine running Windows and SQL Server would be able to hit about 400,000 TPM--the same, more or less, as the Regatta-H machine. To say that Microsoft and Intel are turning up the heat on the Unix vendors is an understatement. There is a reason why HP is throwing in the towel on PA-RISC and concentrating on making chipsets and Itanium-based servers.

Until now, the TX7 server line from NEC has been an obscure product that was aimed only at technical computing customers. The company started shipping the 16-way TX7/i9010 and the 32-way TX7/i9510 at the end of August, and it plans to ship the eight-way TX7/i6010 and 16-way TX7/i6510 in the fourth quarter. Both the 900 MHz and the 1 GHz versions of the Itanium 2 processors are available in the TX7 line. These machines will offer differing numbers of slots and memory expansion. It is unclear if HP will be selling them or not, since it has its own chipsets under development.

So how does this 32-way NEC server running Windows compare with an iSeries box, in terms of price/performance and performance? Good question. But IBM has not yet run a TPC-C test on the top-end iSeries Model 890, so it is hard to say. But I can make some estimates, to give you some perspective, and I am, however reluctantly, willing to do that in the absence of recent hard data from IBM, and refer you to past TPC-C results on the Model 840.

An iSeries Model 840-2420 with a CPW rating of 16,500 was rated at 152,346 TPM, at a cost of $44.52 per TPM. This machine had 24 of IBM's 500 MHz I-Star processors and was configured with 128 GB of main memory and 13.5 TB of disk. The core Model 840 server and its storage had a list price of $6.3 million, and the total price tag for the network under test came to $8.8 million. IBM put a 22 percent discount on the TPC-C database and application servers and its operating system and database software. It offered similar discounts on networking and maintenance costs. The final discounted price for the whole TPC-C configuration was just under $6.8 million, or 23 percent off list price. Since this configuration was tested in March 2001, IBM has cut memory, disk, and Model 270 prices, which might reduce the cost of the machine to somewhere in the range of $35 per TPM. IBM tested a non-standard configuration of the Model 840-2420 with 450 MHz I-Stars and was able to get 163,776 TPM out of the box, which is, paradoxically, closer to the predicted TPC-C rating you would expect, if you knew that CPW is estimated TPC-C, based on an internal IBM implementation of it, divided by 10.

IBM did not test a Model 840-2461 using its 600 MHz S-Star processors on the TPC-C test, but a 24-way machine should be able to, in theory, hit about 202,000 TPM. Similarly, a 32-way iSeries Model 890 with 1.3 GHz Power4 processors should, in theory, be able to hit about 374,000 TPM. If the bigger iSeries boxes perform about 8 percent slower on the real TPC-C test, compared with the CPW tests, then the Model 840-2461 would see its performance reduced to about 186,000 TPM and the Model 890 would see its performance drop to about 345,000 TPM. I have a feeling that with OS/400 V5R2, IBM has fixed whatever performance problems had been limiting its top-end performance in the V4R5 generation. In other words, I think that CPW is once again the predictor of TPC-C performance. IBM has had two years to fix whatever the problem was, and considering that its biggest customers use its biggest boxes, this is not something IBM would let linger for long.

As for pricing, it is hard to say where the iSeries Model 890 might be. First, a Model 890 with 256 GB of main memory and about 30 TB of disk capacity would be very pricey indeed. A Model 890-0198 (which cannot take any interactive processor features) with all 32 processors turned on has a list price of $1.75 million, 256 GB of main memory costs $4.72 million, and 30 TB of disk would run to $2.39 million. The 32 Model 270 servers needed to feed that Regatta-H server would cost about $2.5 million with the recently announced memory and disk price cuts. Maintenance on the whole shebang would run another $1.65. Add all of that up, and it comes to $13 million. Lopping off a 23 percent discount comes to $11.4 million, which would yield a price/performance of around $30 per TPM.

That's roughly where the big Unix iron is, as you can see. And the Windows server is, in terms of bang for the buck, about where it has historically been since the TPC-C test was launched, in 1993. Midrange OS/400 and Unix servers in a given power class have always cost about twice as much as a Wintel solution, and IBM mainframes generally cost about twice as much as OS/400 and Unix servers on tests like the TPC-C.