POUGHKEEPSIE, N.Y.-In the Warner Bros. cartoons, the Road Runner dashed across a vast desert landscape continually outmaneuvering its archrival, Wile E. Coyote, thanks to incredible speed and quick thinking.
IBM’s Roadrunner supercomputer will also find a home in a desert landscape-Los Alamos National Laboratory in New Mexico-and will also rely on speed and quick thinking, or in this case calculations, to keep ahead of its competition. The massive system is poised to become one of the world’s fastest high-performance computers, according to IBM officials.
At IBM’s facility here in upstate New York, engineers are putting the final touches on Roadrunner before it’s shipped to the Department of Energy’s facility in Los Alamos in August. Donald Grice, IBM’s chief engineer for the Roadrunner project, said the $100 million machine is likely to offer a sustained performance of 1 petaflop-1 quadrillion calculations per second-a goal that several other HPC makers such as Cray, Sun Microsystems and SGI are working toward.
In the past two weeks, SGI and Intel have announced plans to create a supercomputer at NASA that will break the petaflop mark. Meanwhile, Sun is working toward its own system that will reach the petaflop plateau.
What looks to make Roadrunner different from other supercomputers-including IBM’s own Blue Gene systems, which hold several spots in the upper echelon of the Top 500 supercomputer list-is its use of the Cell processors that IBM co-developed with Sony and Toshiba for use in game consoles like PlayStation.
What IBM has developed is a heterogeneous system that utilizes standard x86 Opteron processors from Advanced Micro Devices as well as Cell chips for use as accelerators within the Roadrunner system.
This type of internal design should help developers create the scientific applications that will be used within the Roadrunner system. Developers will be able to write applications that take advantage of the Cell processors as accelerators for the heavy computational parts of the application, while the x86 processors can handle the standard computing.
“You get to the hard part and you accelerate it,” Grice told eWEEK. “It takes a little work, but you do it once and you use it over and over again. They do a little bit of work to put routines down there and it does all the work. Conceptually, it’s not very different from branching down into assembly language, so you have to find your hard part; you move that down to the Cell. Then you code the application for the Cell, just like you would for assembly language … and then you move on.”
This supercomputer design helps overcome some of the difficulties associated with the trend of emphasizing increases in processors’ density and number of cores rather than clock frequency. It also allows IBM to build a machine that uses less electricity, while increasing its ability to scale to a petaflop and beyond.
Anatomy of a Supercomputer
Gordon Haff, an analyst with Illuminata, said IBM is following a trend within HPC (high-performance computing) toward hybrid, heterogeneous offerings that use a combination of chips to accelerate different tasks.
“We have been in this path toward very general-purpose computing for a while now, and you are starting to see a lot of people talk about ways to basically combine different types of processing elements,” Haff said. “The really hard part here is the software, because we don’t have very good programming models to handle heterogeneous processing, but the supercomputer, high-performance computing folks, particularly at the extreme high end, are much more tolerant of programming difficulty given the type of performance they need.”
According to IDC, the market for HPC servers grew 15.5 percent in 2007, to $11.6 billion. The research company said it expects the market to reach $15 billion by 2011.
Analysts and vendors say technology that is introduced in HPC often trickles down into the enterprise. For example, Grice said financial firms are always anxious to get hold of supercomputing technology, which can be used to speed up transactions and data-intensive calculations.
In addition to incorporating Cell into the system, Grice and his colleagues also turned to standard IBM parts and other commodity components. Roadrunner will use AMD processors and IBM blades, and the system will run Red Hat’s Fedora Linux operating system exclusively.
At the heart of Roadrunner is what IBM calls a triblade, a node made up of two BladeCenter QS22 blades that contain four Cell processors and an LS21 blade with two Opteron chips. IBM engineers then used a PCI Express cable to connect each of the Opteron cores to a corresponding Cell processor.
With Roadrunner and new QS22 blades that IBM brought out May 13, the company is using a variation on the Cell processor called PowerXCell 8i, which allows for “native” double-precision computing, doubling the blade’s capacity to process data. This is essential to scientific applications.
When Roadrunner is reassembled in Los Alamos later in 2008, it will comprise a series of what IBM calls “connected units,” with each containing a Voltaire InfiniBand switch. Each unit is made up of 16 racks and contains 180 of the specially made triblades.
Altogether, the installation at Los Alamos will have 18 of these connected units, which should offer a sustained petaflop of performance.
With Roadrunner, Grice said IBM is following a path to increasing the power of supercomputers that the industry has followed during the last 30 years. Grice estimated that the performance of the top machine on the Top 500 list grows by 1,000 every 10 years.
“There is no limit to the demand of what scientists and the financial people will want to do,” Grice said.