Advancing the State of Auto-tuning with Programming Languages

Abstract

In the realm of computer science, auto-tuning refers to techniques for software performance optimization. The focus of traditional auto-tuning research is to identify novel performance parameters to expand the optimization space for a given target software/platform combination, and improve the automated search within this optimization space. This makes high-performance computing (HPC) a prime candidate for auto-tuning research, as it sits at the nexus of architectural diversity and performance criticality. However, the major successes for HPC auto-tuning to date involve tailoring memory access patterns to specific cache hierarchies. While important, this is just a small piece of the overall performance portability puzzle.

I argue that auto-tuning has room to expand and optimize a richer set of HPC application tuning parameters through the combination of novel non-intrusive programming language idioms and advanced lightweight online search techniques. I support my argument through four contributions to the field. This dissertation describes two techniques for expanding auto-tuning optimization spaces, and two techniques for distributing the auto-tuning search for parallel efficiency.