Phase-Based Dynamic Instruction Window Optimization for Embedded Systems

Abstract

Even though much previous work explores adapting instruction queue (IQ) and reorder buffer (ROB) sizes to application requirements, traditional IQ/ROB optimizations may be prohibitive for resource-constrained embedded systems, due to the hardware/execution time overheads. We propose low overhead, phase-based instruction window optimization to dynamically vary IQ and ROB sizes for different execution phases based on the applications' variable execution characteristics. Results show that our methodology reduces both the average execution time and energy consumption by 23%, compared to a base system with fixed IQ/ROB sizes.