Original URL: http://www.theregister.co.uk/2010/10/19/unisys_xeon_mainframes/

Unisys lights up Xeon-based mainframes

Secure partitions and integrated specialty engines

By Timothy Prickett Morgan

Posted in Servers, 19th October 2010 12:55 GMT

Mainframe maker Unisys has made no secret of the fact that its long-term goal is to stop making its own mainframe engines and run both of its MCP and OS 2200 operating system platforms atop machines using Intel's Xeon processors.

That day has not yet arrived, but with the delivery of the Xeon 7500 processors this year and a revamped midrange ClearPath mainframe lineup launching today, Unisys is one step closer to getting out of the chip designing racket.

Unisys is launching four machines today, the ClearPath Libra 4100 running the MCP platform (Burroughs) and three ClearPath Dorado 4100s supporting the OS 2200 platform (Sperry). There might only be one Libra 4100 machine, but this will be the first of the Unisys mainframes to get a homegrown virtualization hypervisor called secure partitioning, or sPar for short, that Unisys has been working on for more than two years.

El Reg told you about it here a little more than two years ago, and it was not entirely clear why Unisys wanted to create its own hypervisor.

But the answer, at least for now, is not to compete against VMware, Microsoft, and Microsoft in the x64-based hypervisor racket, but rather to create a hypervisor that is controlled by MCP and OS 2200 but which allows Windows or Linux workloads, such as Java application servers or cryptographic accelerators, to be moved into sPars on Xeon-based mainframes. This is a much cleaner solution that using network-connected, outboard specialty engines as Unisys has been doing to a number of years.

The basic hardware platform behind the ClearPath 4100 machines is four-socket server using a six-core Xeon X7542 from Intel. This chip has HyperThreading turned off and offers the highest clock speed in the Xeon 7500 lineup. Bill Maclean, vice president of ClearPath portfolio management at Unisys, says the company went with the six-core variant to boost the single thread performance of the MCP and OS 2200 workloads.

Unisys has a partnership with NEC to create high-end Xeon MP servers, with its "Monster Xeon" machine announced two years ago being the first such product. Unisys has resold x64-based servers made by Sun Microsystems and Dell as well as making its own, and Maclean was not in a mood to divulge who the supplier was behind this ClearPath 4100 box, but did say it came from a third party supplier and that Unisys can - and does - change partners as technology and market conditions dictate. (Wouldn't it be funny if the new box was made by Acer, with Gateway having been eaten by Acer and the ex-Gateway executive who sold the company to Acer now running Unisys? But I think it is a Dell box because it is using an iDrac6 enterprise remote management service processor, and specifically I think it is a Dell PowerEdge R910.)

What is important to Unisys customers is that the ClearPath 4100 machines offer more oomph on single-threaded and batch workloads than the current ClearPath 4000 series boxes, which were launched a little more than a year ago. Unisys cannot yet replace its largest ClearPath 700 series mainframes, which have proprietary CMOS processors designed by Unisys and fabbed by IBM and which were updated about a year and a half ago, with Xeon-based iron.

Maclean says because of issues relating to single-threaded and I/O subsystem performance and how this relates to middleware emulation software crafted by Unisys to make its mainframe workloads run atop a Windows or Linux kernel, this process will take years. Like two to three years for MCP workloads to scale as far on Xeon-based machines as they can on the successors to the CMOS Libra 700 boxes and a little bit longer on the Dorado OS 2200 machines. There's a lot of tuning Unisys needs to do to map instruction lookahead and do pipeline management in this emulation firmware to future Intel chips to squeeze out the performance mainframe shops expect.

The Libra 4100 mainframe has 24 cores across its four sockets, and half of these cores are dedicated to running the MCP operating system on up to six emulated MCP central processing modules. (Two cores per CPM.) The MCP half of the machine has 64 GB of mirrored physical memory (128 GB total), and the box has ten PCI-Express slots (four x8 and six x4 slots) for peripherals.

The full-bore emulated MCP engine performance on the Libra 4100 server is 300 MIPS, which compares quite favorably to the Libra 4000 it replaces, which had an emulated engine performance of 200 MIPS, and the single-image system capacity of the dozen cores on the MCP half of the Libra 4100 comes in at 1,750 MIPS, compared to 800 MIPS for the Libra 4000. That's a 50 per cent improvement in single-engine performance and more than a factor of two more of aggregate MIPS in the box. But just so you see how far Unisys has to go to get off its own CMOS engines, a Libra 700 engine that was announced a year and a half ago is rated at 550 MIPS and the system spans an aggregate of 5,700 MIPS running non-emulated MCP.

There's still a dozen Xeon X7542 processors still not accounted for yet, and these are used to run the specialty engine workloads that Unisys has been hosting on outboard servers. These remaining Xeon cores are carved up into three secure partitions, just like the MCP instance has running atop of the emulation firmware; two of these sPars run the JProcessor Java application server (an implementation of Red Hat's JBoss) and the Unisys ePortal, which is used to link COBOL applications to handhelds and iPads, among other things.

Unisys will eventually offer specialty engines that support cryptographic co-processing and IBM's WebSphere-MQ message queuing middleware, as it does on outboard specialty engines today. The sPar-based specialty engines link to the MCP partitions through the MCP memory queues, creating what amounts to a memory-speed virtual LAN link between the virtual boxes. All of the software for the specialty engines (including the stripped down Linux or Windows kernels) come preloaded and preconfigured on the boxes, and are updated through the MCP patching processes.

The Libra 4100 comes in two flavors. The Libra 4180 is the traditional mainframe that starts out at 50 MIPS of initial performance, while the Libra 4190 is the metered version with utility pricing that scales up and down each month based on usage, which starts out at a lower initial 30 MIPS. The initial price of the Libra 4190, is what most customers will go for, is $550,000.

On the OS 2200 front, the Dorado 4100s are also being launched today, but they will not get sPar capabilities for another 18 months or so because more work needs to be done on the emulated OS 2200 I/O subsystems within the sPars.

The underlying hardware is exactly the same, it is just that the dozen cores that would be allocated to sPars running JProcessor or ePortal specialty engines are dormant. The Dorado 4100 engines have an emulated engine speed of 225 MIPS, compared to 195 MIPS for the engines in the Dorado 4000s they replace (not much of a boost, to be honest), but the aggregate scalability is much better at 1,600 MIPS on the Dorado 4100 compared to 600 MIPS on the Dorado 4000. (But again, the big CMOS-based Dorado 700 has a single-engine speed of 525 MIPS running its OS 2200 workloads and an aggregate of 5,700 MIPS to stretch its COBOL legs out in.) These Dorado 4100 machines have two Dorado I/O expansion modules.

The Dorado 4180 can be set up with one or two OS 2200 partitions and can emulate up to ten OS 2200 instruction processors across those dozen cores it has access to. The machine has the same 64 GB of mirrored memory, sports up to 112 Dorado I/O ports, and comes in a base 50 MIPS configuration. The Dorado 4190 is the metered version of the box with utility pricing, and this metered version starts at $330,000 for a 30 MIPS machine. These two boxes will also come in special high availability configurations (dubbed the Dorado 4180-HA and Dorado 4190-HA) that allow a pair of machines to be linked together in a failover cluster. The Libra 4180-HA and Libra 4190-HA models will not be available until the second quarter of 2011.

Finally, there is a new puppy machine called the Dorado 4150, which is based on the same Xeon iron but which only offers two OS 2200 instruction processors and one Dorado I/O expansion module. Pricing and MIPS for this box was not available at press time.

All of the machines will begin shipping on October 29. ®