Fix the code. Don’t tweak the hardware: A newcompiler approach to Voltage-Frequency scaling

Abstract

Traditional compiler approaches to optimize power effi-ciency aim to adjust voltage and frequency at runtime tomatch the code characteristics to the hardware (e.g., run-ning memory-bound phases at a lower frequency). However,such approaches are constrained by three factors: (i) voltage-frequency transitions are too slow to be applied at instructiongranularity, (ii) larger code regions are seldom unequivocallymemory- or compute-bound, and, (iii) the available voltagescaling range for future technologies is rapidly shrinking.These factors necessitate new approaches to address power-efficiency at the code-generation level. This paper proposesone such approach to automatically generate power-efficientcode using a decoupled access/execute (DAE) model.In DAE a program is split into tasks, where each task con-sists of two sufficiently coarse-grained phases to enable ef-fective Dynamic Voltage Frequency Scaling (DVFS): (i) theaccess-phase for data prefetch (heavily memory-bound), and(ii) theexecute-phase that performs the actual computation(heavily compute-bound). Our contribution is to provide acompiler methodology to automatically generate the access-phases for a task-based programming system. Our approachis capable of handling both affine (through a polyhedral anal-ysis) and non-affine codes (through optimized task skele-tons). Our evaluation shows that the automatically gener-ated versions improve EDP by 25% on average comparedto a coupled execution, without any performance degrada-tion, and surpasses the EDP savings of the correspondinghand-crafted tasks by 5%