 |
Inside this Issue |
||||
| Quantifying Quasar | |||||
Solving 30 trillion arithmetic calculations in one second is fast. That’s the speed at which the newest Cray Inc. supercomputer at the Arctic Region Supercomputing Center can subtract, add, divide or multiply numbers. But it’s still not fast enough to do the kind of number crunching needed to create the high-resolution, real-world computer models that multi-disciplinary scientists want. Multi-disciplinary research is rooted in the belief that new discoveries come at the nexus where scientific disciplines intersect. Ecologists, biologists, oceanographers, volcanologists, physicists, computer scientists and mathematicians collectively bring to bear the tools needed to address problems beyond those from a single discipline. An essential tool comes in the form of tons of data distinct to particular areas of scientific expertise. Processing all that data requires power beyond the capabilities of conventional computers. Computer technology has advanced to the point where petaflop speed, a quadrillion calculations a second, is now possible. At the leading edge of the high performance computing (HPC) industry, different technologies are being introduced at a rapid pace, but are not yet fully understood in detail. At ARSC, a new computer named Quasar is providing an open research platform to study next generation technology, helping to quantify how hybrid processors and systems can be used to accelerate computationally demanding applications. “Quasar is a lot of machine for not a lot of money,” says John Mitchell, an HPC systems analyst at ARSC. Mitchell has been putting together the various parts and pieces that make up Quasar since the summer of 2008, when IBM rolled out its BladeCenter QS22 technology. Quasar is actually a cluster of IBM QS22 blades that fit into small frame about the size of a filing cabinet. A blade is much like a computer node, where specialized processors, memory and networks all live together on a single blade of electronic components. There are a dozen QS22 blades in Quasar. Each blade has two Cell Broadband Engine (BE) processors, specifically the IBM PowerXCell 8i processor. And according to next-generation technology experts at ARSC, these Cell BEs are vastly improved from those found in Playstations and elsewhere. Quasar fits in a single IBM BladeCenter H chasis, yet has a theoretical single precision peak performance of more than 5 teraflops, meaning the machine could potentially perform five trillion calculations a second. “Hybrid computing architecture allows us to have faster, smaller, less expensive and more energy efficient machines,” Mitchell said. According to Chief Scientist Greg Newby, who has taken the lead on analyzing next generation technologies for high performance computing at ARSC, Quasar uses the Torque batch system, a Lustre file system and other features that are typically found on much larger HPC machines. “Our current research focus with Quasar is trying to determine the extent to which 5 teraflops of theoretical performance can be achieved in practice,” Newby said. “To do this, you need a formal methodology using a common set of metrics and common experiments over a number of representative tools,” he said. Hybrid computer architectures are highly desirable because of their speed, size, relatively low purchase price and small energy footprint. But the technology has advanced so quickly, it has outpaced the ability of humans to create instructions the computer can understand so that it can work at peak performance. Anton Kulchitsky is an HPC specialist at ARSC studying solar wind and the interplanetary magnetic field. The interaction of the energy-charged solar wind with Earth’s magnetic field can cause disruption of satellite services and ground-based communications. It takes a lot of calculations and physical modeling to simulate the processes. “In general, most existing approaches for HPC do not work well on Quasar,” Kulchitsky said. “It took many days to write programs and test them on Quasar.” According to Newby, most application developers are willing to give up some performance and chip utilization in exchange for productivity. “But it’s hard to know in a general sense how much performance and utilization is being lost in favor of ease-of-use,” he said. Using Quasar, and quantifiable testing measures, researchers hope to determine, among other things, the extent to which the Cell processors offer speedups over traditional multicore microprocessors for real world applications. “What we’re doing is testing the suitability of these new types of technologies for work on computationally intensive campaigns like climate modeling,” Newby said. Think about it in terms of shopping for a new car. “Driving a superfast vehicle isn’t for everyone,” he said. It takes a certain amount of knowledge, skill and training to keep a formula-one race car on the track and ahead of the pack. Faster is only better if you know how to operate the machine for peak performance. “Sure, we want to get there fast, but we also want to get there in the most efficient manner possible.”
|
|||||
| Challenges Index | |||||
