The Four Hundred

IBM may not be one of the first companies to enter the nascent blade server market, but it has undoubtedly found a lot of prospective tire kickers, early adopters, and enthusiasts for the blade server concept. The company says that it has shipped 5,000 blade servers in the 10 weeks that it has been shipping product, and that it is talking up its future plans for the BladeCenter platform, including four-way 32-bit Xeon MP and two-way 64-bit Power and Itanium blades.

The question I have had since I first caught wind of the future Power-based blade servers, which will fit in the existing BladeCenter chassis, is why not make a blade server that supports OS/400-based blades? (See "BladeCenters Can Be IBM's Fifth Kind of eServer," from the September 30, 2002, edition of this newsletter.) When I was designing those hypothetical iDeal iSeries machines back in November, just after IBM announced that it was working on a volume, uniprocessor version of the Power4 processor called the PowerPC 970. As it turns out, this is the processor that IBM will be using in the two-way Power-based blade (oddly enough, not a dual-core Power4 or Power4+ chip, which is much more expensive to make) that is a future product from IBM. And as I learned doing the iDeal iSeries story, the PowerPC 970 processor supports the full PowerPC and Power instruction sets, including all of the special instructions that are required to support OS/400 and its unique single-level storage architecture. The PowerPC 970 processor is expected to be available in the range of 1 GHz to 2 GHz, which translates into about 1,800 to 3,600 raw CPWs of power depending on the cache structure. A two-way PowerPC 970 server running OS/400 with no governors on it using the 1.8 GHz PowerPC processor that IBM is expected to debut in the BladeCenters would be rated at about 5,800 CPWs of power by my reckoning. This translates into about 55,000 transactions per minute on the TPC-C transaction processing test. This is roughly equivalent to a new iSeries Model 825 with five of its six 1.1GHz Power4 processors activated, to give you some perspective.

So why an iSeries blade inside the BladeCenter chassis? Let's examine the market. IBM's own executives in charge of the blade server business at Big Blue say that they expect that blade server shipments are going to explode this year--more than many industry analysts expect, and particularly as enterprises look to blade designs to pack more computing in less space and to offer an ease of administration and redeployment of blade servers that is not possible across heterogeneous and incompatible server platforms. We'll have to see. There are as many skeptics about the current crop of blade server designs as there are tire kickers. What can honestly be said is that many of the ideas encompassed in blade servers are ones that resonate with IT managers. The question is whether the blade server vendors can match the promises they make about provisioning and administration with blade servers with products that live up to those promises. Only then can blade server shipments grow to become a significant slice of overall server shipments worldwide.

BladeServers are interesting because they internalize a hybrid network of incompatible servers. IBM's current blade servers support Windows and Linux on Intel-based processors, but the current line could just as easily support NetWare, SCO Unix, FreeBSD, or any other Intel-compatible operating system. The Power-based blades will, in theory, be able to support Linux, IBM's AIX Unix variant, and even its OS/400 midrange operating system, should customers want it. Intel-based blades could also support emulated versions of the MVS operating system, which have been available for years, as well as Solaris on 32-bit Intel blades and HP-UX on Itanium-based blades. The ability for IBM to create a single box to support all of these environments, which its customers most certainly use, could be a serious boon for the blade server business and give IBM an edge over its rivals and customers a single platform on which to manage their disparate servers. There's a lot of potential here.

Back to the BladeCenter iron for a minute, and how IBM will change it. The BladeCenter is a 7U form-factor chassis that can house up to 14 two-way server blades, yielding a total of 168 processors in a standard 42U rack. This is twice the level of density as a rack of two-way 1U form factor servers. IBM has been shipping two-way H20 blades based on the "Prestonia" Pentium 4 Xeon DP processor since early December. Big Blue has been shipping 2 GHz and 2.4 GHz processors in these blades, and it expects to start shipping blades with the 2.6 GHz and 2.8 GHz Xeon DP chips in the next few weeks.

But that is not all that IBM plans to do with the BladeCenter chassis. The company is working with Intel under the blade server development alliance we told you about last September to create an Itanium-based two-way blade. This blade, it turns out, will be based on the current 1 GHz "McKinley" Itanium 2 processor, not the 1.5 GHz "Madison" processor due in mid-2003. This blade will take up the space of two Xeon DP blades, so only seven two-way blades will fit in a BladeCenter chassis and only 84 processors will fit in a 42U rack. Also coming out of the partnership with Intel is a four-way blade based on the "Gallatin" Pentium 4 Xeon MP processor. This four-way blade will eat two blade slots in the BladeCenter chassis as well.

Sources also tell us that IBM will indeed ship a two-way blade server based on the future 1.8 GHz PowerPC 970 processor--something many have been speculating about for months. The PowerPC 970 is a single-core variant of the dual-core Power4 processor that IBM is making for its own use in Unix and Linux workstations and servers, as well as in desktop and server machines for Apple. Like the Itanium 2 blade server, it will eat up the space of two Xeon MP blades.

And, finally, sources at IBM suggest that the company may go radical with its blade designs and build blades from its "Summit" chipset. The Summit chipset is a hybrid SMP/NUMA clustering design that allows from one to four processor boards to be clustered together to create machines that scale up to 16 processors in a single-system image. Summit can scale up to 32 or even 64 processors, if IBM gets clever. IBM's blade server designers are pondering the possibilities of creating SMP servers by using variants of the Summit chipset and its external scalability ports, which connect processor cell boards together to create a single-system image, to glue blades in a chassis together into an SMP server. If designed properly, customers could create SMPs based on the components of their choosing and on the fly, reconfiguring as necessary. This would represent a different way of virtualizing server hardware components that is different from logical partitioning.

This latter development is intriguing, and it would be particularly interesting if it could be extended to Power-based blades, too. My sources at IBM suggest that this would not be too easy, since the Summit chipset is tightly coupled to IA-32 and IA-64 processors from Intel. But IBM Rochester has smart engineers, and making a chipset that would allow customers to take a rack of Power-based blade servers and wire them up, on the fly, as SMP servers ranging in size from four-way up to 32-way machines, for instance, is something to think about.

One last note: Here's some more perspective on the IBM blade server shipment numbers. While 5,000 blades in 10 weeks is a great number at the beginning of a market, IT organizations consume millions of servers a year. HP has only sold 15,000 of its "QuickBlade" ProLiant BL blades in a year, and it is by far the dominant supplier so far, although IBM's ship rates have more or less reached the level that HP has set. Blades are not a volume market, so IBM has time to work OS/400 blade servers into its plans. The kind of integration and ease of management that blade servers are aiming to bring to the market will resonate strongly with OS/400 shops.