Run-time Cache Configuration for the LEON-3 Embedded Processor

Abstract

Cache parameters such as size and associativity are fixed at manufacturing time which are often not tuned for the specific characteristics of each application code. The net result is excessive energy consumption and lower performance. This paper explores the benefits of the use of a reconfigurable data cache in terms of capacity and associativity in a LEON-3 embedded system. We present real energy and execution time results for a set of graph-based and numerical algorithms. For a combined application of these algorithms, the results reveal an aggregate energy savings of 7% and a execution time penalty of just 1% over the best fixed-associativity cache architecture with the same capacity. We further explore the performance of a dynamic cache way shutdown adaptive algorithm and evaluate its performance and energy benefits in the context of the SLAM-EKF position estimation robotics algorithm.