Autotuning Algorithmic Choice for Input Sensitivity

Abstract

Empirical autotuning is increasingly being used in many domains to achieve optimized performance in a variety of different execution environments. A daunting challenge faced by such autotuners is input sensitivity, where the best autotuned configuration may vary with different input sets. In this paper, we propose a two level solution that: first, clusters to find input sets that are similar in input feature space; then, uses an evolutionary autotuner to build an optimized program for each of these clusters; and, finally, builds an adaptive overhead aware classifier which assigns each input to a specific input optimized program. Our approach addresses the complex trade-off between using expensive features, to accurately characterize an input, and cheaper features, which can be computed with less overhead. Experimental results show that by adapting to different inputs one can obtain up to a 3x speedup over using a single configuration for all inputs.