Energy management via PI control for data parallel applications with throughput constraints

Abstract

This paper presents a new proportional-integral (PI) controller that sets the operating point of computing tiles in a system on chip (SoC). We address data-parallel applications with throughput constraints. The controller settings are investigated for application configurations with different QoS levels and different buffer sizes. The control method is evaluated on a test chip with four tiles executing a realistic HMAX object recognition application. Experimental results suggest that the proposed controller outperforms the state-of-the-art results: it attains, on average, 25% less number of frequency switches and has slightly higher energy savings. The reduction in number of frequency switches is important because it decreases the involved overhead. In addition, the PI controller meets the throughput constraint in cases where other approaches fail.