After years of trials, tribulations, and technological innovations, the PC server seems to have finally grown up into a full-fledged midrange server. By the time you start reading this, Intel and its PC server partners will have started peddling the high-end midrange servers that use the Profusion chipset developed by Intel’s Corollary unit in conjunction with Compaq, the reigning champion of the PC desktop and server markets. Corollary, which Intel purchased in 1997, used to make high-end symmetric multiprocessing (SMP) machines that went beyond the SMP clustering that Intel offered at the time, which was limited to two-way servers. Intel bought Corollary and solicited the help of Compaq, which, through its Digital and Tandem server groups, had plenty of experience building machines with multiple processors that present an operating system and applications with a single system image. Intel may be the highest-volume server board manufacturer in the world, but it has definitely been a technology laggard behind long-time SMP expert IBM, which has been building multiprocessors since the 1980s, and upstarts Sun Microsystems, Hewlett-Packard (HP), and Compaq’s Digital unit in the UNIX server area, which all perfected their SMP designs in the mid-1990s.
With the new Profusion chipset, Intel is now able to glue two of its four-way Pentium III Xeon processor cards together to create an eight-way server, much as Apache and Northstar AS/400e servers glue together three four-way AS/400 processor cards to create 12-way servers and the forthcoming I-Star servers will glue together up to four sixway AS/400 processor cards to create 24-way servers. Even though the Intel Profusion chipset is far behind the SMP work of the big server vendors like IBM, Sun, and HP, the new eight-way machines put PC servers, which can run various UNIX and Linux implementations as well as Windows NT, in the same power class as the belly of the AS/400 market: Model 720 and 730 machines. By reaching this point, Intel and its server partners may not be able to show the mind-boggling scalability of an AS/400 I-Star or the new RS/6000 S80, but these Profusion machines will have enough oomph to run the kinds of e-business applications that IBM would like very much to keep on the AS/400 as native workloads.
The Profusion machines like IBM’s 8500R, Compaq’s 8550-2M, HP’s 8500 LXr, and Dell’s 8450 will be very attractive alternatives to proprietary and open-system servers that offer as much or more scalability. And they will be attractive not just because they offer near-commodity pricing—Intel hates when I say that, because it doesn’t want you thinking
about these high-powered electronic parts as commodities and especially doesn’t want you to think the same thing about next year’s 64-bit IA-64 Merced processors—but also because they offer good engine power in addition to more than enough scalability to meet the data center processing needs of small, medium, and, depending on the workloads, large businesses. Up until now, only specialized servers using non-uniform memory access (NUMA) clustering from Sequent and Data General could use Intel iron to support big midrange and high-end mainframe-class workloads. As Figure 1 shows, the 550-MHz Pentium III Xeon processor that is used in conjunction with the Profusion chipset in PC servers is only modestly more powerful than the current 262-MHz Northstar PowerPC processor used in the IBM AS/400 line. (Compared with the 32-bit Pentium III Xeon, the PowerPC is a much more elegant 64-bit processor, which is why it can do almost as much work with half as many processing cycles. But elegance is not what sells computers anymore; price is.) The irony is that the 637-MHz Turbo Opera G6 processor used in
IBM’s S/390 mainframes is only about 60 percent more powerful than the 550-MHz Pentium III Xeon chip. The new 450-MHz Pulsar processor that IBM announced in the RS/6000 “Condor” S80 servers—next year’s I-Star chip will be a silicon-on-insulator implementation of the copper-based Pulsar chip—is almost twice as powerful as the Pentium III Xeon, and subsequent I-Star processors will be more than two and three times as powerful as the Intel chips, even after Intel jacks up clock speeds to 600 MHz, 800 MHz, and higher on its processors during 2000 and 2001. It is clear that Intel will always be behind IBM in raw processing power. It is also clear that IBM will be able to show better scalability on its midrange server lines, as Figure 2 shows. But none of this may matter a whole lot.
With Profusion, big iron is going mainstream, and that will have a profound effect on the midrange market, of which the AS/400 is still a major part. Very few mainframe customers, much less midrange customers, will need more processing power than the Profusion servers provide. And with Intel cranking up the clock speed and extending SMP support on the boxes and with partners like IBM providing server clustering for
them—IBM has just adapted its RS/6000 SP clustering hardware and software to support 14-way clustering of Profusion and other kinds of Netfinity servers—customers will never run out of gas on their Intel iron again. This is an unprecedented moment for the PC server market because PC servers have always been too puny to handle big jobs and grow with them. It took IBM almost eight years to get the AS/400 line to that point with the first generation 530 “Muskie” RISC boxes, and only now have IBM’s mainframe customers gotten big iron that exceeds their needs with the G6 servers. (Because of the nature of how UNIX machines are used as network backbones, many RS/6000 customers will need all the power IBM can muster for several years to come.)
It is safe to say that IBM, HP, and Compaq business partners will be under increasing pressure to sell Profusion boxes, where they might have otherwise been happy to push a big AS/400 or UNIX box, which costs more but yields higher margins. Most business partners will be willing to sacrifice high margins on an AS/400 sale for the relatively low margins on a Netfinity 8500R because the Profusion box is an easy sell. An eight-way IBM Profusion server with 2 MB of L2 cache per processor and 8 GB of main memory costs $110,000; Microsoft Windows NT 4.0 Enterprise Edition costs $4,000; and SQLServer costs $29,000. (I haven’t configured disks to the box, but obviously, customers will have to.) That comes to $143,000. This Profusion server can support 40,266 TPC-C transactions per minute (TPM), and at $3.55 per TPM for a base server, many midrange customers will buy it. Contrast this with a base AS/400e Northstar server with eight processors, 8 GB of main memory, the OS/400 operating system, and the DB2/400 database bundled in. The eight-way 740-2069 costs $580,000 and, based on extrapolating from the performance results for the larger 12-way Northstar 740-2070, can support only about 35,000 TPC-C TPM. That comes to $16.57 per TPM. This is a modest gap in performance that the AS/400e I-Stars will easily bridge but a very big gap in price/performance that they almost certainly will not.
However shortsighted, many midrange customers pick their machines based mostly on price. As one MIS manager recently explained it to me, it is much more important to tell his boss that he has saved him five 0s than it is to say he has five 9s of reliability. At small and medium private companies that are the backbone of the midrange, any money not spent on computers goes right into the owner’s pockets. And that, more than any other factor, explains why people are clamoring for Windows NT. Profusion servers will only accelerate this idiotic trend unless IBM completely rethinks the way it sells AS/400s. According to market researcher International Data Corp. (IDC), Intel-based servers will account for $21 billion, or 31 percent of the aggregate revenues in the $68 billion worldwide server market, in 1999, before Profusion sales even kick in. By 2003, IDC reckons that the Intel server market—pushed by 32-bit and 64-bit processors and Profusion and follow-on chipsets that will expand SMP capabilities to 16, 24, and maybe even 32 processors in a single box—will more than double to $46 billion, more than half of the $89 billion worldwide server market. And by the way, IDC expects that the non-Intel portion of the server pie will contract from $47 billion to $43 billion over the same term. A lot of high-end server vendors are going to be fighting pretty hard for that shrinking piece of the pie.
I think AS/400s must be only modestly more expensive than NT servers, regardless of reliability. NT is controlling the momentum in the midrange, not OS/400, and that makes the AS/400 an underdog regardless of its reliability and scalability benefits. AS/400s have to be an easier sell as well. That means unbundling the operating system, database, and other features from the hardware and cutting AS/400 hardware and software prices. IBM has to be able to demonstrate that its hardware and software are not much more expensive than Profusion iron and Windows NT software on an item-by-item list. Its advertisements have to hammer home the point that for only a 10 percent premium, AS/400 hardware is many times more reliable. This would, I believe, attract customers.
But odds are IBM will stick to the old marketing message of reliability and scalability and try to get as much money for the AS/400 as it can. IBM will have some impressive performance and much-improved price/performance statistics to show with its upcoming AS/400 lines. (Go to www.midrangecomputing.com/mc/ to see my estimates for various workloads.) This is great news for midrange customers who want to stick with their AS/400s, which is the vast majority of the installed base. But the continuing price/performance advantages of the PC server market will continue to make those platforms more attractive for everything but core financial applications. IBM may have built all the Internet and e-business extensions into the AS/400 and attracted vendors of popular enterprise resource planning (ERP) suites to the box, but the AS/400 Division has a long way to go if it wants to get even a 20 or 30 percent market share on these new workloads against Intel servers, even at AS/400 sites. IBM has to change, plain and simple, and if the Profusion servers and Windows NT can’t make it change, maybe nothing short of a customer and Business Partner uprising can.
Breaking News
As we go to press, Intel has discovered a bug in the Pentium III Xeon processors used in conjunction with the Profusion chipset on its “Saber” OCPRF100 eight-way motherboards. Intel says that a voltage regulator in Pentium III Xeons with 512 KB and 1 MB of L2 cache memory goes nuts when the machines are running at peak power and causes them to crash. The bug is definitely in the Pentium III chips, not the Profusion chipset, and affects only Saber motherboards designed by Intel. Compaq and Hitachi, which make their own Profusion motherboards, are not affected by the bug, nor is the Saber motherboard when it is equipped with Pentium III Xeons with the full 2 MB of L2 cache memory. This is so because the 2-MB design is more recent. Intel hoped to have a workaround by mid- October.
IBM 800-MHZ I-Star PowerPC IBM 560-MHz I-Star PowerPC IBM 450-MHz Pulsar PowerPC IBM 637-MHz Turbo Opera G6 IBM 340-MHz Northstar PowerPC Intel 550-MHz Pentium III Xeon IBM 262-MHz Northstar PowerPC
0.00 0.50 1.00 1.50 2.00 2.50 3.00 3.50
Relative Uniprocessor Performance
INUTE
M
P
ER
C
APACITY
, T P
O
NLINE
AS/400 I-Star (560 MHz)
RANSACTIONS ROCESSING
S/390 G6 (637 MHz)
Compaq Profusion (550 MHz)
AS/400 Northstar (262 MHz)
Source: Vendor, Midrange Computing estimates
1 2 3 4 5 6 7 8 9 1011 121314 151617181920 21 2223 24
N
UMBER OF
P
ROCESSORS IN
E
ACH
S
ERVER
Figure 2: Intel Profusion servers offer more scalability than AS/400e Northstars on a processor-to-processor basis, but next year’s I-Stars will dwarf IBM’s G6 mainframes.
LATEST COMMENTS
MC Press Online