The IBM mainframe may not have a lot of direct competition when it comes to z/OS-based batch and transactional work, but the story is different when it comes to Linux. There's plenty of competition among Linux platforms, and Big Blue can't ignore the pressure that Moore's Law brings to bear.
So IBM is slashing the prices it charges for its System z mainframe engines that are set up to exclusively run Linux, the so-called Integrated Facility for Linux speciality engines. These are sold at much lower prices than standard mainframe engines, just like other speciality engines designed to accelerate WebSphere middleware (zAAPs) and DB2 database routines (zIIPs).
The price that IBM charges for mainframe engines depends on which variety you are talking about. Firstly, there are the high-end System z Enterprise Class (EC) servers, which cram up to 16 of IBM's four-core, 4.4 GHz z6 mainframe engines into a single system image, spanning 64 processor cores and nearly 30,000 MIPS of processing capacity.
Alternatively there is the System z Business Class, a midrange box that uses three-core z6 chips that presumably run at a lower clock speed, and that have from one to five engines for z/OS workloads and up to ten IFLs for running Linux images atop of the z/VM hypervisor. Last fall, when the System z BC machines were launched, IBM cut the price of IFLs by 40 per cent, from $125,000 per engine to $75,000. Because the System z BC machines were aimed at smaller shops, and used engines with less oomph, IBM charged only $90,000 for IFLs on these smaller boxes.
Now, IBM has cut the price of IFLs on the System BC machines to $47,500 a pop, a 47.2 per cent price cut that is directly attributable to the increasing core counts and performance for x64 servers and the increasing sophistication of hypervisors on x64 iron, which allows many Linux images to be crammed onto a single machine, much as the combination of z/VM and IFLs allows dozens to hundreds to thousands of Linux images to be put on a System z mainframe.
"The price change is in part because of the increased performance with the latest Nehalem EP Xeons," says Karl Freund, vice president of System z strategy and marketing at IBM, adding that the price change on the BC mainframe boxes is also aimed at blunting the announcement by Intel of eight-core "Nehalem EX" Xeon 7500s at the end of this year.
The Nehalem processors, with their QuickPath Interconnect, are able to support many more virtual machines, thanks to better virtualization electronics and much higher memory bandwidth. While expensive, mainframes offer established virtualization, as well as high memory and I/O bandwidth. This is a must for machines that have to run at somewhere north of 95 per cent utilization to make them economically feasible.
Because IBM is not ready to revamp its System z mainframes with eight-core processors (but probably will sometime around the launch of eight-core Power7 processors next year), it has little choice but to compete with better and faster and more capacious x64 iron from both Intel and Advanced Micro Devices on price.
Because IBM doesn't provide pricing on mainframe systems (including processors, memory, storage, and other necessary peripherals) or for its z/VM hypervisor, it is tough to reckon how to compare the IFLs inside the System z EC and BC machines. But you can isolate the per-engine costs easily enough and work backwards to see how many x64 cores the mainframe engine prices imply they can cover when running Linux workloads. Let's just talk about the five-engine System z 2098-z05, the top end BC box, just to get a feel for it.
This machine offers five z6 engines running at what I would guess to be 3.2 GHz, based on the 2,760 MIPS rating for the box, and according to mainframe watcher Technology News, has a list price of $2.55 million when configured to run z/OS. A base box with hardly no MIPS and memory costs $97,500, so activating the five engines in the box as IFLs would add another $237,500. Then you have to take off another $11,796 for the base 26 MIPS of z/OS capacity that was in the entry configuration.
When you do the math, that works out to $323,204 for a five-processor Linux machine with no memory, no disk, and no systems software, including z/VM or Linux. Because there is a big disparity between the cost of Linux on the mainframe and Linux on x64 boxes - excuse me, I forgot to speak perfectly GNUbie there: the disparity is between Linux support costs, since Linux is not an operating system and even if it were, it is free - let's add Linux to this barebones mainframe.
Given the discounts that Novell has cooked up and its 80 per cent market share on the mainframe, let's slap some SUSE Linux Enterprise Server 11 on the System z BC box. It costs $10,200 per engine, with discounts for a one-year standard support contract for SLES 11 on mainframes, so that adds another $51,000. (At list price, this would cost $75,000.)
If customers want to pre-pay for five years of support, they can get a support contract for $37,499, or $7,499 per engine per year. That's a big price drop, so let's be generous and do the comparison over five years. The bare System z machine with Linux and five years of support costs $510,699.
Let's now look at how many Xeon 5500 boxes with Linux and support that gets you, not including memory, disk, or hypervisor costs, as fairly as we can, given the lack of information about mainframe pricing. We'll take IBM's own System x servers, and use the System x3550 M2 rack server, the cheapest of IBM's Nehalem EP boxes. With two of Intel's four-core, top-end Xeon X5570 processors, the x3550 M2 costs $6,681. Main memory for this x64 box costs $109 per GB, and there is no way IBM is not charging a lot more for main memory on mainframes.
Now, toss on SLES 11, which is priced per machine, not per processor core, on x64 machinery. It costs $799 per box per year, or $3,995 per box for five years of support. So that comes to $10,676 for a bare-bones x64 Linux image.
What these system prices imply is that a fully-engined System z BC with five 3.2 GHz engines should be able to support the same number of Linux images as 48 of those System x3550 M2 rack servers, which have a total of 384 Nehalem cores running at 2.93 GHz.
I am not, of course, suggesting that the System z BC mainframe can support that number of images. As efficient and wonderful as z/VM may be, VMware's ESX Server 4.0 hypervisor is no slouch. It seems very reasonable that at the very least, an x64 box should be able to support one virtual machine per core with acceptable response times running Linux workloads. So that would work out to 384 Linux slices on those System x racks.
The question mainframe shops have to ask IBM is whether a five-processor System z BC can support something akin to this number of VMs running Linux concurrently - and then have Big Blue run benchmarks to prove it. I have my doubts, and until IBM provides some benchmark results that prove otherwise, so will Linux shops.
Traditional mainframe shops, and IT managers keen on maintaining their empires and taking away some power from the x64 server crowd, will no doubt talk about the operational benefits and savings that come from consolidating Linux workloads onto mainframes. These benefits are even harder to qualify and quantify, but that is how mainframe iron and other proprietary machines - and indeed Unix boxes - continue to sell against cheaper x64 iron. ®