Auto-tuning on the macro scale : high level algorithmic auto-tuning for scientific applications

Abstract

In this thesis, we describe a new classification of auto-tuning methodologies spanning from low-level optimizations to high-level algorithmic tuning. This classification spectrum of auto-tuning methods encompasses the space of tuning parameters from low-level optimizations (such as block sizes, iteration ordering, vectorization, etc.) to high-level algorithmic choices (such as whether to use an iterative solver or a direct solver). We present and analyze four novel auto-tuning systems that incorporate several techniques that fall along a spectrum from the low-level to the high-level: i) a multiplatform, auto-tuning parallel code generation framework for generalized stencil loops, ii) an auto-tunable algorithm for solving dense triangular systems, iii) an auto-tunable multigrid solver for sparse linear systems, and iv) tuned statistical regression techniques for fine-tuning wind forecasts and resource estimations to assist in the integration of wind resources into the electrical grid. We also include a project assessment report for a wind turbine installation for the City of Cambridge to highlight an area of application (wind prediction and resource assessment) where these computational auto-tuning techniques could prove useful in the future.