Where Is the Mainframe Blade Server?

Published: June 17, 2008

by Timothy Prickett Morgan

Last week, IBM completed the rollout of the most powerful computer on Earth: the "Roadrunner" hybrid Opteron-Cell supercomputer cluster. This machine was built for Los Alamos Laboratory in New Mexico by the U.S. Department of Energy so it can simulate nuclear explosions and the decay over time of warhead components in the American nuclear arsenal. Roadrunner is also something else other than the first machine to break through the 1 petaflops performance barrier. It is also the most powerful blade server setup in the world.

Blade servers got their start back before the dot-com boom, and the telecom and service provider sector has had standardized blade server form factors, backplanes, interconnections, and other standards because, as IT suppliers know full well, you don't mess with the phone company. Not if you want to make your next quarterly financial results in the telco sector, anyway. Commercial blade servers became available as the dot-com boom was busting and just as companies were running out of money to pay a premium for innovative new server packaging, which blade servers, even the non-standard form factors peddled by manufacturers that want to lock their customers into a certain form factor and management toolset, certainly do offer. And that bad timing has meant that the ramp for blade servers has been a little slower than many industry observers had expected.

Even with a certain amount of compression in the blade server space as IBM and HP wrestled for the corporate market--meaning, companies who don't buy standardized telco blades, which cost a lot of money and which are often lacking in performance features that commercial applications require, such as lots of memory or the fastest processors on the market--there has been a lot of innovation in the blade server space, and the idea has caught on. In the first quarter of 2008, the analysts at Gartner reckoned that blade servers accounted for 10 percent of worldwide shipments of 2.22 million units, and IDC said that blade server sales rocketed up 53.7 percent to hit $1.2 billion, also about 10 percent of total server revenues on the planet in the first quarter. While this is great growth, blade servers would account for maybe a third of all shipments and sales by now if they had followed the same hockey stick curve of rack-mounted servers in the late 1990s.

In any event, there are plenty of blade server options to choose from across the key server vendors--Hewlett-Packard, IBM, Dell, Sun Microsystems, Egenera, Fujitsu-Siemens, Hitachi, and NEC--plus a slew of white box suppliers and tier two server makers. There are Itanium blades, Sparc blades, Power blades, Cell blades, and of course X64 blades. As of this spring, there are even Power6 blades with two or four cores that support the applications created for IBM's AS/400 proprietary minicomputers and their RPG applications and latest i 6.1 operating system. And just this week, HP announced that it has moved the NonStop fault tolerant computing environment and its ServerNet interconnect to a standard HP BL860c Itanium-based blade, which was announced last year. You can also run OpenVMS as well as HP-UX, Linux, and Windows on this same blade.

There's one thing that cannot run on a blade, however: the z/OS mainframe operating system and its applications and databases.

That is, to put it bluntly, short-sighted. IBM, which hung back to get into the blade server racket while the formerly independent Compaq and HP slung blows at each other in late 1999 and throughout 2000, was the blade server market leader in 2005 and 2006 and has created a large ecosystem of blade component suppliers and partners through its Blade.org organization; it has even partnered with Intel to try to push its BladeCenter chassis and interconnect as an industry standard, with a small amount of success. (Meaning some vendors make blades that fit in the chassis, but none of the players mentioned above do.)

IBM should know that customers with mainframes want the power density and simplicity of blades, in terms of wiring and interconnect. The blade form factor does not mean SMP scalability has to be sacrificed. IBM's Opteron blades can snap together to make larger SMP nodes, and so do blades from Appro and Fujitsu-Siemens. IBM has engineered a high-end z10 mainframe product line with the quad-core z6 mainframe chips, for which it should be commended because it has given its largest mainframe customers the headroom they need to grow. But the company did not, for some reason, create a z processor that can scale down to a blade and then be scaled up through SMP electronics. Or, simply used as an entry mainframe for smaller customers and government agencies, which still have modest MIPS requirements despite the behemoths IBM can deliver.

Because there is no mainframe blade, IBM has left itself wide open to competition from the low-end of the server line--just as it did by taking six years to get a blue-blooded, 64-bit Power blade that could run AIX, Linux, and i5/OS (now called "i" for some reason). Sure, IBM finally got a variant of the PowerPC chips used in vintage Apple Macs on the JS20 and JS21 blades, but these chips did not support the full "Amazon" instruction set that Power4, Power5, and Power6 chips do, which meant they could not run OS/400, i5/OS, and i. But despite the bad treatment that OS/400 shops suffered from a short-sighted Big Blue, they eventually got blade servers. Mainframe customers are still waiting, and probably will never see one until IBM goes to a bladed architecture for all of its machines aside from tower servers.

This is not as wild as it sounds. Imagine putting processor sockets and memory on one blade, I/O features on another, and disks and other storage on yet another. Then you plug in the components you need and create a system. You can scale up processing capacity if you need it, or memory, or I/O, depending on your workload needs, and you still get the tight integration and simplified packaging of a blade architecture. And, provided server makers standardize on memory and I/O features, it might even mean that components can be reused as some components change.

The engineering work for a mainframe blade cannot be very daunting, or at least not any more than was the case for Roadrunner. IBM custom built the "tri-blade" interconnection blade used in the Roadrunner super, which lashes one two-socket LS21 Opteron blade with two QS22 Cell blades, which each have two Cell processors and which are used to boost the number-crunching performance of the Opterons. The triblade uses PCI-Express links to hook the QS22s to the LS21, and then has an InfiniBand interface to link to the outside network and storage. IBM could use a dual-core z9 mainframe engine, or a half-dud z6 engine, to make such a blade. The z6 chip already has a SMP hub chip on it, so presumably engineers could figure out how to gang up blades to make larger and larger SMPs in the System z line. (Each z6 core runs at 4.4 GHz and delivers about 920 MIPS.)

Ironically, HP or NEC could get a mainframe blade out the door first, thanks to their partnerships with Platform Solutions and its mainframe emulation environment.

RELATED STORIES

IBM Launches 64-Way z10 Enterprise Class Mainframe Behemoth

Thinking Inside the Box

A Mainframe Renaissance

Motherboarding

IBM Readies Quad-Core z6 Chip for Mainframe Iron

The IBM Mainframe Base: Alive and Kicking

The X Factor: One Socket to Rule Them All

                     Post this story to del.icio.us
               Post this story to Digg
    Post this story to Slashdot