CAREER: A Framework for Dynamic Self-Tuning of General Purpose Programs

University Of Illinois At Urbana-Champaign, Urbana IL

Investigators

Abstract

CAREER: A Framework for Dynamic Self-Tuning of General Purpose Programs Abstract As computing systems become more complicated, they are exposing an increasingly large number of "knobs" that can be used for tuning. These knobs typically represent a trade-off (e.g., size vs. speed of a local data storage resource) and thus must be set differently for different workloads to achieve optimal performance (or power-performance). While many such knobs have been proposed, there has been little work towards a comprehensive approach to set these knobs automatically. The proposed research is meant to help fill this gap. Specifically, this work proposes a dynamic self-tuning framework for optimizing the compilation of general-purpose programs. The proposed framework draws inspiration from the successes in empirical optimization frameworks like ATLAS, applying their ideas to a new context where optimization cannot be done at install time. By performing the tuning at run time, code can be optimized for specific input data (a necessity for non-numeric programs), but introduces challenges in maintaining low overhead and good performance in the presence on non-stationary workloads.