A variable assignment method by analyzing access pattern for power optimization on a Hybrid memory with volatile/non-volatile RAM

Abstract

In this paper we present an approach to improve energy saving and performance in a STT-MRAM/DRAM hybrid memory architecture. In this approach we analyze the variable read/write of an application for optimal allocation of memory for these variables during runtime. Knowing the variable read/write ratio from the analysis, we allocate the once with high read access ratio to STT-MRAM and variables with high write access to DRAM to improve energy consumed by memory module in embedded system, and improve performance by reducing memory read latency. We present a variable read/write analysis of the very well adopted FFMpeg audio/video library. We exploit the huge memory footprint and computationally intensive characteristics of the FFmpeg application. We also present a modified Qemu emulator which we added a features to enable trapping and logging all cpu memory access operation to specific variables, which enable an accurate analysis. Our experiment result in a Hybrid memory aware FFmpeg application which shows a reduced energy consumption of up to 42%.