IBM's "T-Rex" to rule mainframe shops through 2004. (Infrastructure News Review).
In any event, the zSeries 990 mainframes are built on a more scalable symmetric multiprocessing design that will eventually allow the machine to scale to 64 active processors with, according to IBM sources, a single system image of an IBM z/OS operating system spanning 48 processors. The current zSeries 900 "Freeway" servers use 16 of IBM's G7 mainframe processors, which are rated at about 300 MIPS each. The G8 processors in the TRex machines were expected to be in the range of 400 to 450 MIPS each, and all of the analysts are today doing the math to try to reckon this. Based on some relative performance metrics supplied by IBM to business partners, it looks like IBM did a little better than expected, with a machine that uses processors that deliver between 50% and 60% more performance, CPU for CPU. It looks like the G8 is actually delivering somewhere around 465 to 485 MIPS. The performance increase, by the way, directly applies to how Linux runs on the box.
The 64-bit G8 processors have a clock speed of 1.2GHz, compared to the 770MHz clock speed of the fastest 64-bit G7 Turbo processor. The Freeways scaled to 16 processors in a single system image, but the T-Rex boxes will support 32 processors by the end of the year. System I/O bandwidth in the T-Rex servers has been nearly quadrupled, from 24GB/sec in the Freeways to 96GB/sec in the T-Rexes. Main memory has also been quadrupled to 256GB in the T-Rex machine, up from 64GB in the Freeways. IBM said today that more than 1,200 developers have been working for four years on the T-Rex machines, and that it had invested over $1bn in research, development, and advanced manufacturing techniques to bring T-Rex to life.
There are four different models of the T-Rex mainframes (known inside IBM by the machine number 2084) that will be available in 2003. Incidentally, there are only four models being sold, compared to the 42 different versions of the zSeries 900 that are currently being sold as regular mainframes or Linux variants. IBM only plans to add two more models in the near future to the zSeries 990 line, too. IBM's zSeries managers are following the lead of their iSeries brethren and radically simplifying their product line to make manufacturing and selling the machines easier and more profitable. On Demand utility computing means IBM doesn't have to carve out each processor power level as a unique machine.
The zSeries 990 Model A08, which has eight active processors, and Model B16, which has sixteen active processors, will be available June 16, and they will support up to 15 logical partitions and have one channel subsystem. These machines will support all current mainframe operating system versions and releases, and upgrades to them will be available from installed Freeway servers. Prepaid On/Off Capacity on Demand features as well as Customer Initiated Upgrades will be available within the zSeries 990 A08 and B16 machines starting September 15. The former allows companies to buy blocks of processor days for peak workloads, while the latter allows customers to permanently activate processors within the zSeries mainframe central electronics complex when they need them. The zSeries 990 C24 and D32 models have 24 and 32 active processors, respectively, and will be available October 31. These machines have two channel subsystems and will support up to 30 z/OS logical partitions. IBM is expected to roll out a 48-way T-Rex box (presumably the E48) in late 2003 or early 2004, and will debut a 64-way version of the machine (probably the F64) about the same time the 64-way Power5 "Squadron" servers are available. These latter machines will have three- and four-channel subsystems and will support up to 60 logical partitions.
The zSeries 990 architecture puts a bunch of extra processors on each "book," which is what we might call a cell board in other servers. But IBM's multichip module (MCM) packaging for the zSeries machines crams what would be multiple motherboards onto a single piece of ceramic substrate, so the cell board metaphor doesn't really hold up. That MCM is 3.7 inches square and threequarters of an inch high, and it contains all of the processors, SAPs, spares, cache memory, and other guts of the T-Rex server (a total of 16 unique chips). Over 3.2 billion transistors, made using a copper/silicon-on-insulator processor, are crammed onto each MCM, which is made at IBM's Burlington, Vermont chip fab.
In any event, each T-Rex book has 12 processor units. Eight of them can be activated to do real work, either running z/OS or another mainframe operating system or Linux through the VM-based Integrated Linux Facility (ILF). Two of the extra four processors are assigned as System Assist Processors (SAPs), which are used by the channel subsystems, and two of the extras are spares in case a primary active processor fails. Any unactivated processors in a T-Rex book can be activated as a coupling facility (used for Parallel Sysplex clustering) or as SAPs. Each book can support four 8GB memory cards, which suggests that the 64-way will support 256GB of main memory when it ships next year. The T-Rex chassis can support 512 channels.
The design of the T-Rex machines is one that is trying to embody two different and sometimes competing impulses that all server designers face today. Customers want a machine that can scale up, allowing them to throw hardware (meaning more processors, memory, I/O, and so forth) at an existing database or application so they don't have to re-write their code. But some customers also want such a large machine to scale out horizontally with virtualization technologies like logical partitions so they can consolidate their physical server sprawl of hundreds of machines down to a virtual server sprawl inside one actual machine (or at least fewer machines). The T-Rex machines, which will eventually support a single z/OS system image that has at least three times the processing power of the largest 16-way Freeway server and which will support hundreds of virtual Linux partitions through a modified version of the z/VM operating system, can meet both of these requirements.
Pricing is anyone's guess on mainframes these days, but IBM is telling partners that it is offering Linux processors with 50% more power at the same price as Integrated Linux Facilities cost on the zSeries 900 machines. IBM has said it has more pricing flexibility because of the lower-cost design in the T-Rex box, but stops short of actually giving prices. The company told partners the price/performance on T-Rex machines running Linux would improve by 45%, the price/performance of memory would improve by 45%, and the price/performance of maintenance on the machines would improve by 25%. (This is a weird way to characterize a price cut, I realize.) It's what IBM doesn't say that is sometimes as important as what it does say.
The current zSeries 900 machines, in reasonable base configurations, range in price between $1,800 and $1,900 per MIPS. The fact that IBM did not say price/performance on z/OS workloads would improve can only mean IBM is not looking to cut prices very much, if at all. Say IBM is willing to go to $1,600 per MIPS on a T-Rex box. When you play around with the numbers, a zSeries 990 A08 should have about 2,988 MIPS and cost about $4.7m, compared to a zSeries 900 2C8, which has 1,940 MIPS and costs about $3.2m. A 16-way zSeries 990 B16 should have about 4,884 MIPS and cost about $7.8m, compared to a 16-way zSeries 900 216, which is rated at about 3,192 MIPS and which costs about $5.3m. Prices for processors running Linux will be considerably lower; these prices assume customers want to run z/OS, OS/390, or another IBM-supplied mainframe operating system.
IBM said today that a 32-way T-Rex would have up to 9,000 MIPS of aggregate processing power, almost three times the processor power of 16-way Freeway machine. It's hard to say where a 64-way T-Rex will end up on the MIPS scale, but somewhere in the 13,000 MIPS range seems possible, if IBM can get the same kind of SMP scalability out of the T-Rex that big Unix servers are demonstrating. This is a lot of processing capacity for mainframe customers to digest, and it may be one of the reasons that there is speculation that the "Copernicus" G9 series of mainframe servers have been put on hold or pushed out beyond 2006. There simply may be no reason for a more scalable mainframe, provided IBM can boost the clock speed, memory capacities, and I/O bandwidth in the T-Rex chassis for a few years.
|Printer friendly Cite/link Email Feedback|
|Date:||May 28, 2003|
|Previous Article:||IBM thinks Dynamic for policy-based utility provisioning. (Infrastructure News Review).|
|Next Article:||Intel says some "McKinley" Itanium 2 Chips have bugs. (Infrastructure News Review).|