Application-Based Tuning of Mathematical Kernels on Petascale Systems

Petascale supercomputers offer the computational power to make significant progress in DOE¿s most complex electromagnetic problems. Existing simulation tools have to be carefully tuned to take advantage of the high degree of concurrency offered by these systems. This project will optimize a plasma physics modeling tool widely used within DOE's scientific community for leadership-class supercomputing systems. Good performance on largest scale systems is achieved by addressing algorithm scalability, single processor performance and code portability. By separating the physics algorithms form the numerical kernel routines this project will be able to take advantage of highly tuned numerical libraries, thus getting good performance while maintaining a high level of portability. The Phase I project completed a detailed performance analysis for an electromagnetic field solver by using profiling tools and comparison with reference implementations. It determined the optimal memory layout for this algorithm and developed a performance model that can predict the performance on large processor counts. It also demonstrated that numerical libraries can lead to high-performance implementations at low code complexity. The Phase II project will first resolve the scalability bottlenecks discovered in the Phase I and then focus on the performance of the charged particle model. It will use and enhance a code transformation system to implement optimizations in a portable way. Finally, it will address scalability of particle problems by using a load balancing library. Commercial Applications and other Benefits as described by the awardee: This project will lead to a highly optimized commercial plasma physics code that fully utilizes leadership-class supercomputing systems.