High-level optimization via automated statistical modeling

  • Authors:

  • Eric A. Brewer

  • Affiliations:

  • University of California at Berkeley

  • Venue:
  • PPOPP '95 Proceedings of the fifth ACM SIGPLAN symposium on Principles and practice of parallel programming
  • Year:
  • 1995

Quantified Score

Hi-index 0.00

Visualization

Abstract

We develop the use of statistical modeling for portable high-level optimizations such as data layout and algorithm selection. We build the models automatically from profiling information, which ensures robust and accurate models that reflect all aspects of the target platform.We use the models to select among several data layouts for an iterative PDE solver and to select among several sorting algorithms. The selection is correct more than 99% of the time on each of four platforms. In the few cases it selects suboptimally, the selected implementation performs nearly as well; that is, it always makes at least a very good choice. Correct selection is platform and workload dependent and can improve performance by nearly a factor of three.We also use the models to optimize parameters of these applications automatically. In all cases, the models predicted the optimal parameter setting, resulting in improvements ranging up to a factor of three.Finally, we use the models to construct portable high-level libraries, which contain multiple implementations and support for automatic selection and parameter optimization of the fastest implementation for the target platform and workload.