Evolutionary algorithms for compiler-enabled program autotuning

Abstract

PetaBricks [4, 21, 7, 3, 5] is an implicitly parallel programming language which, through the process of autotuning, can automatically optimize programs for fast QoS-aware execution on any hardware. In this thesis we develop and evaluate two PetaBricks autotuners: INCREA and SiblingRivalry. INCREA, based on a novel bottom-up evolutionary algorithm, optimizes programs offline at compile time. SiblingRivalry improves on INCREA by optimizing online during a program's execution, dynamically adapting to changes in hardware and the operating system. Continuous adaptation is achieved through racing, where half of available resources are devoted to always-on learning. We evaluate INCREA and SiblingRivalry on a large number of real-world benchmarks, and show that our autotuners can significantly speed up PetaBricks programs with respect to many non-tuned and mis-tuned baselines. Our results indicate the need for a continuous learning loop that can optimize efficiently by exploiting online knowledge of a program's performance. The results leave open the question of how to solve the online optimization problem on all cores, i.e. without racing.