ITR/NGS: Automatic Performance Tuning for Large Scale Scientific Applications

University Of Tennessee Knoxville, Knoxville TN

Investigators

Abstract

Simulation has become a critical component of the scientific method, with the resulting demands for computational power continually pushing the limits of available hardware. Scientific applications need to be tuned to acheive high efficiency, but hand tuning by application scientists cannot keep pace with the ever changing features of the underlying hardware. It is not uncommon for applications that involve a large amount of communication or memory operations relative to computation to run at under 10% of the peak performance of a machine. The goal of the project is to address the widening gap between peak performance of computers and attained performance of real applications by developing several innovative approaches to address such challenges. Namely: design new automatic tuning techniques for dense linear algebra, sparse linear algebra, and inter processor communication kernels; target emerging architectures of importance to the high performance computing community, including commodity processors with SIMD extensions, clusters, vector processors, and highly parallel machines; develop an intermediate representation and compilation model for these operations that will allow tuning of new operations to be incorporated into compilers and programming systems; deliver these capabilities to users by integrating them into languages (Matlab, UPC, Titanium); and explore the impact of new kernels on higher level algorithm design. The project will deliver tuned kernels to users through standard libraries (BLAS, LAPACK, MPI, ScaLAPACK, PETSc) as well as parallel languages designed to make high end machines more accessible (UPC and Titanium), and will evaluate the impact of our work on large scale scientific applications that use these systems.