Cache-Aware SPM Allocation Algorithms for Performance and Energy Optimization on Hybrid SPM-Cache Architecture

Abstract

Recently, hybrid SPM-cache (HSC) architectures that combine caches and scratch-pad memories (SPMs) have been increasingly used in commercial processors and research prototypes. However, most SPM allocation algorithms do not make use of the cache in parallel, so they cannot harness the full potential of the HSC architecture. In this paper, we explore six SPM allocation algorithms, including four heuristic-based algorithms and two optimal algorithms based on model checking. Our experimental results show that the greedy stack distance based allocation for energy (GSDA-E) can also reduce the energy by as much or by nearly as much as the optimal results attained by the optimal stack distance based allocation for energy (OSDA-E), while also achieving performance close to those of the OSDA and GSDA.