Features:

INTERVIEW WITH JEREMY KEPNER, MIT LINCOLN LABORATORY
By Alan Beck, Editor-in-Chief, HPCwire

HPCwire: The introduction to your SC2003 panel notes that "the DARPA High Productivity Computing Systems is focused on providing a new generation of economically viable high productivity computing systems for the national security and industrial user community in the 2007-2010 timeframe." How will DARPA go about accomplishing this goal? What technological innovations, now on the horizon, will contribute to it?

JEREMY KEPNER: First, I think you will see very innovative hardware designs from the three vendors which bridge the gap between todays vectors machines and the future promise of Quantum computing. Second, and equally important, I think you will see development of novel program environments that allow ordinary users to effectively 10,000+ CPU systems. Finally, you will see a new approach to evaluating HPC systems that takes into more diverse factors such as programming cost; this is where the HPCS Productivity Team I lead comes in to play.

HPC: It is further noted that with respect to these computing resources "(t)he goal is to provide systems that double in productivity (or value) every 18 months." Your panel will debate the issue of productivity measurement; how do you feel such benchmarking efforts should be directed?

JK: I think we will see a new class of benchmarks emerging out of HPCS (e.g. the new HPCchallenge benchmarks) that are potentially representative of a broader range of real applications. In addition, I think there will be some very novel work in the area of benchmarking how hard (or easy) a particular HPC system is to use.

HPC: Do you believe there can or should be a universal standard for measuring HPC productivity across divergent applications? Please elaborate on why or why not.

JK: One of the things we see emerging is the need for a framework that clearly separates the system specific and application specific parts of HPC evaluation. It is our goal that users should be able to develop a standard methodology for obtaining parameters describing HPC systems and parameters describing HPC applications. We also plan to develop models that will combine these parameters so users can gauge how good a particular systems is for their application.

HPC: Has the increasing importance of commercial HPC applications changed the way productivity should be evaluated? Why or why not?

JK: I think it allows us to draw on new evaluation ideas that are more mainstream in the commercial sector. For example, determining the cost of HPC software. In addition, we are very much aware that today bioinformatics is an important emerging area.

HPC: In your opinion, will Grid-based systems ever truly draw even in productivity with proprietary, vector-based machines? Why or why not?

JK: It is clearly application specific. The government will probably always have applications that are best suited to both classes of systems.

HPC: Is there anything else you feel our readers should understand about these issues?

JK: The HPCS Productivity Team is very interested in working with the HPC community to move forward together on this very tough problem of what is the best way to measure HPC systems. We hope the community will take a serious look at what we are doing and we are open to ideas from across the community.