It's Time for an iSeries BladeCenter
by Timothy Prickett Morgan
I spent a day last week up at Somers, New York, where most of IBM's marketing plans are hatched. Somers, which has four buildings with glass pyramids on top, designed by IM Pei, is also where the top brass of the Systems Group get together to create the products that IBM will make and to hatch the plans to sell them. If you were going to do a hostile takeover of IBM, this, not Armonk, is where you would start.
Mark Shearer, vice president of eServer product marketing, hosted the event, and he brought along a few other vice presidents, as well as two customers and an industry analyst, to talk about how the data center is changing. Shearer said a lot of things, but the main message was that IBM sees the future as a combination of scale-up SMP servers with mainframe-class features and scale-out blade servers, with grid middleware linking the two into a big, flexible infrastructure. Again and again, he said it: grid plus blade plus mainframe. I have a feeling IBM was using journalists as a sample on which to test a new marketing message.
Here's an interesting aside about the event, before I get on my soapbox about IBM's BladeCenter and how there ought to be a way to get an iSeries version of this puppy out the door. One of the customers that IBM brought in was Hewitt Associates--specifically, Perry Cliburn, chief information officer of the company, who flew up just before Isabel hit. Hewitt is a $1 billion global human resources outsourcing and consulting firm. It has a set of CICS and Smalltalk applications running in two mainframe data centers. One particularly CPU-intensive application running on one of the mainframes was a pension projection program that was just rolled out last week. This program ate about 1,000 MIPS of processing capacity, which comes to about $2 million worth of base zSeries 900 or 990 hardware. Hewitt ported that application to a set of BladeCenter servers with 10 2.8-GHz Xeon processors. Because all numerically intensive jobs love high clock speed, hyperthreading, and big L2 caches, this pension projection code ran a lot faster on the BladeCenters than it did on the mainframe, which runs at well under 1 GHz. Here's the kicker: the BladeCenter configuration necessary to do the job only cost one-tenth as much as the mainframe capacity.
To IBM's way of thinking, the BladeCenter is being positioned as an adjunct processor to the mainframe, like the Integrated xSeries Server and its predecessors have been adjuncts to the AS/400 and iSeries for the better part of a decade. IBM would rather sell the BladeCenters--which can pack 14 two-way servers, redundant switches, as well as a whole server network infrastructure into a 7U, rack-mounted chassis--as customers offload their workloads from mainframes, than lose out to a rival, so this approach shows some sense. I was surprised that IBM would even bring it up, though. I can't imagine IBM being happy watching applications move off the mainframe. But Rich Lechner, vice president of sales and marketing for enterprise servers, told me later in the day that he was convinced that Hewitt would very quickly find other uses for that freed-up mainframe capacity, and would even grow its mainframe workloads in the near future.
While the BladeCenter machines are, for now, being peddled as boxes to run Web and other infrastructure workloads, and are popular as the basis of Linux clusters for running supercomputer-style workloads, Jeff Benck, vice president and director of marketing for the blade server line at IBM, said at the Somers shindig that IBM reckoned that more than 90 percent of the IT workloads running in the world would be deployable on scale-out architectures like the BladeCenter by 2006. That one rocked me back on my heels. But think about it. Distributed, clustered versions of DB2 and Oracle actually work. It is significantly less expensive to buy two two-way servers than one four-way server; 32-way servers, like the iSeries 890, are much more expensive than a small iSeries box.
I met Benck at COMMON and didn't yet know he was the vice president in charge of the BladeCenter, but I took the opportunity to tell him that if IBM was going to start shipping, in early 2004, a PowerPC 970-based dual-processor blade for the BladeCenter chassis to run either AIX or Linux, then IBM should go the extra mile and port OS/400 to it. When I met him again in Somers, and knew who he was, I gave him the hard sell on the idea. He said that an OS/400 blade was not in the official plan but that IBM was kicking the idea around. I explained I was not exactly happy about that, and that iSeries shops wouldn't be, either.
The story is not going to end there. It just so happens that Jim Pertzborn, who many of you will remember from his days rolling out line after line of AS/400 servers in the 1990s, is the vice president of eServer products strategy at IBM, and he's the guy who spearheaded the design efforts of the BladeCenter. He also cut in front of me in the lunch line. Bad move for him; good move for me. Once I figured out who he was, I started giving him the hard sell, too. He conceded that there is no reason why IBM could not roll out an OS/400 blade. There are issues with I/O, which is different in OS/400 machines. But there are ways around this, particularly with a SAN-ready box like the BladeCenter.
The first step, said Pertzborn, would be to make the BladeCenter compatible with the Integrated xSeries Adapter card for the iSeries line. This would allow BladeCenter machines to link back to the iSeries through the High Speed Link interconnect, like other outboard xSeries servers from IBM do today. While this will be interesting and useful, particularly for big customers who, like Hewitt, want to offload some numerically intensive applications from their iSeries to the BladeCenter, I keep thinking about the little guy, the core iSeries customer.
I think the only computer that small companies want is something like the BladeCenter. If the AS/400 and then the iSeries have been about the integration of all the core software necessary to run a small or midsized business, the BladeCenter carries that integration out to the hardware and network level. Guild Companies has a half-rack of 1U and 2U Intel-based servers, a storage area network disk array, a bunch of switches, and all the other infrastructure that a small or midsized business has. And it is a mess. The rack system we have is better than anyone would have imagined only five years ago, but it's still less than elegant, with its spaghetti tangle of wires. It is not particularly well-packed, not when you compare it with the BladeCenter. Rather than sell tower servers to small and midsized businesses, IBM might do well by creating a quarter rack and putting in one BladeCenter and one SAN. This would be a great starting system, and would have lots of room to grow.
I keep thinking about those midsized OS/400 shops that have paid hundreds of thousands of dollars to buy monolithic AS/400 and iSeries machines, and heaven knows how much additional money on Windows servers. A bunch of PowerPC 970 blades running OS/400 and DB2/400 Multisystem might yield the same performance for these customers as an iSeries Model 825 or Model 870, and they could put their Windows servers in the same BladeCenter chassis, too.
Last year, I estimated that a two-way PowerPC 970 blade with the processors running at 1.4 GHz would have about 4,500 CPWs of raw computing power; using 2 GHz PPC 970s, that raw performance jumps to 6,400 CPWs. I think these numbers might be a little high. A single 1.3 GHz Power4 processor delivers about 1,500 CPWs of power, and in a 32-way machine, each processor delivers about 1,150 CPWs after the 20 percent overhead of being glued together in a big SMP cluster is taken out. For argument's sake, say that a 2 GHz PPC 970 blade with two processors is rated at 4,600 CPWs. A fully loaded BladeCenter would then have 64,400 aggregate CPWs of power. That's 35 percent more aggregate power (before SMP overhead is taken out) than a 32-way iSeries Model 890 server. Each blade could have partitions, just like iSeries machines do today, and IBM might even be able to make partitions span across blades.
Again, what I am talking about is not really about big customers, but it might have a profound effect on how they use OS/400 in the future. This will be particularly true if the volume economics of the blade market can be brought to bear. Eventually, hundreds of thousands of these blades will ship each year. IBM may be able to sell OS/400 blades more profitably than it does a Model 270 machine.
Another big reason to push for OS/400 blades is that many customers want dense, rack-mounted iSeries machines. The AS/400 started out in racks, and just as the industry moved en force to racks, IBM abandoned them. IBM has not brought back racks for the iSeries, but OS/400 blades for the BladeCenter would make it possible. People want racks.
There would be problems with OS/400 blade servers, of course. For one, the BladeCenter runs on 240-volt power lines; whereas small AS/400 and iSeries servers have always plugged right into normal 120-volt sockets. IBM needs to fix that and to deliver a 120-volt version of the BladeCenter. The other problem would be explaining to customers when they need a real iSeries and when they can use OS/400 blades. The iSeries channel will want to sell the more expensive product, of course, because they will make more money on it in the short run. But a well-thought-out OS/400 blade server strategy, coupled with the ability to integrate OS/400, Windows, Linux, and AIX workloads in a single chassis, and internalizing the server network inside of that chassis, might help make OS/400 server sales a pull, rather than a push. It's something to think about.
If you think this is a good idea, there is only one way it will happen: you have to tell IBM. Send me your e-mails at email@example.com, and I will forward them to the right people at Big Blue. If there isn't grassroots support for this, it isn't going to happen.
Contact the Editors
|Copyright © 1996-2008 Guild Companies, Inc. All Rights Reserved.|