A Priori Power Estimation of Linear Solvers on Multi-Core Processors

Dimitar Lukarski and Tobias Skoglund

September 2013

Abstract:

High-performance computing (HPC) centres simulate complex scientific models which provide vital understanding of our world. In the recent years, power efficiency has become a critical aspect in the new HPC facilities because of high energy consumption costs. In this work, we present our study on power consumption of linear solvers on modern multi-core CPUs which are widely used in many scientific applications. We focus on both dense and sparse linear solvers - parallel direct solvers for the dense, and parallel iterative methods for the sparse problems.

We provide answers to the questions - what is the energy efficiency using multi-core parallel algorithms for linear systems and do we save energy using more cores? Furthermore, we propose a methodology for estimating total power consumption. Based on benchmarks which achieve high accuracy we estimate total power usage of the whole solution phase on multi-core CPUs.

Available as PDF (869 kB, no cover)

Download BibTeX entry.