SC13 Denver, CO

The International Conference for High Performance Computing, Networking, Storage and Analysis

Enhancing Learning-Based Autotuning with Composite and Diagnostic Feature Vectors.

Authors: Saami Rahman (Texas State University), Richard Hay (Texas State University), Mario A. Gutierrez (Texas State University), Apan Qasem (Texas State University)

Abstract: The use of machine learning techniques has emerged as a promising strategy for autotuning. Central to the success of such learned heuristics is the construction of a feature vector that accurately captures program behavior. Although the salient features of certain domain specific kernels are well understood, at least in principle, automatically deriving suitable features for general numerical applications remains a significant challenge. This poster presents a learning-based autotuning system that introduces two new classes of features that encapsulate key architecture-sensitive performance characteristics for a broad range of scientific applications. The first class of features are high-level, composite and derived from compiler models for data locality and parallelization. The second class of features is a series of synthesized and normalized HW performance counter values that diagnose causes of program inefficiencies. Preliminary experimental results show that the enhanced feature vectors increase prediction accuracy by as much as 23% for several learning algorithms.

Poster: pdf
Two-page extended abstract: pdf

Poster Index