U.S. Defense Department Chooses Los Alamos to Help Design Supercomputers
Los Alamos National Laboratory scientists will try to predict how the supercomputers of the future will perform under a three-year, $4.2-million grant recently awarded by the research arm of the U.S. Department of Defense.The Defense Advanced Research Projects Agency’s High Productivity Computing Systems program will fund performance analysis and modeling, system architecture analysis, software tool creation and networking evaluations. The research will take place in Los Alamos’ Performance and Architectures Laboratory of the Computer and Computational Sciences Division, in partnership with IBM.
Los Alamos was selected for the project because of its expertise in advanced supercomputer architectures, including pioneering work in performance modeling, which computer system designers can use to explore how different design options will affect performance. Los Alamos performance models permit analysis of general system characteristics as well as dynamic analysis based on information gathered during the operation of a specific system, and allow computer architects and manufacturers to predict how specific applications will perform on various computational systems.
DARPA’s HPCS program seeks to develop by 2008 the optimal computer architecture for the next generation of machines, those capable of sustained performance greater than one quadrillion operations per second, or one petaflop, or 10 to the 15th power.
The Performance and Architectures Laboratory’s expertise in coaxing more speed out of the world’s most powerful computers is based on its pioneering work in developing and applying novel methodologies that accurately compare, analyze and predict performance of computers from different manufacturers, running different operating systems. In addition, PAL’s research goes the extra step from analysis to proposing design approaches that optimize performance for applications of interest to the Laboratory. The ability to analyze current or proposed computer architectures specifically for applications of interest to Los Alamos before a computer is even built is a unique feature of the PAL team’s work.
“We’re excited about the enormous potential of performance modeling and the opportunity to apply a variety of methodologies that we developed to help in the design these future supercomputers,” said Adolfy Hoisie, who heads the PAL. “Our performance models have become the preferred, predictive design tool for computer system architects who need to run ‘what-if’ scenarios or answer ‘experiment-ahead’ questions.”
Members of the PAL team recently captured the Best Paper Award at the international Supercomputing 2003 Conference for “The Case of the Missing Supercomputer Performance: Achieving Optimal Performance on the 8,192 Processors of ASCI Q.” Using novel techniques encompassing analytical models, custom micro-benchmarks, full applications and simulators, the team identified the source of a serious but previously undetected performance problem on Los Alamos’ Q machine, the second-most-powerful computer in the world. They eliminated the problem and demonstrated how to significantly improve application and system performance in an analysis immediately applicable to other large, cluster-based supercomputers.