Examining Energy Efficiency of Vectorization Techniques Using a Gaussian Elimination

Abstract

Modern computer environments are limited by energy and power constraints during the execution of programs. These limits can be due to power lines, budgeting, ecology, battery life or many other reasons. To bypass these limits, hardware and software development strive to reduce the energy and power consumption of the execution of algorithms. This article investigates the capabilities and limitations of vectorization with respect to energy efficiency. Vectorization is a technique to instrument on-chip SIMD execution that increases the performance of programs. The capability of vectorization to reduce energy consumption of programs has to be shown. As applicative algorithm, the well-known Gaussian elimination is vectorized and investigated. Several implementations, including automatic and manual vectorization techniques, have been developed and their execution time and energy consumption have been measured and compared.