HPA: An Opportunistic Approach to Embedded Energy Efficiency

TL;DR

HPA is a dynamic, transparent optimization method that enables generic applications to efficiently utilize heterogeneous computing units, leading to faster processing and reduced energy consumption on embedded platforms.

Contribution

This paper introduces HPA, a novel on-the-fly optimization scheme that automatically exploits heterogeneous computing resources for energy-efficient embedded processing.

Findings

HPA improves processing speed on NVIDIA Jetson TK1.

HPA significantly reduces total energy consumption.

HPA effectively leverages heterogeneous resources for energy efficiency.

Abstract

Reducing energy consumption is a challenge that is faced on a daily basis by teams from the High-Performance Computing as well as the Embedded domain. This issue is mostly attacked from an hardware perspective, by devising architectures that put energy efficiency as a primary target, often at the cost of processing power. Lately, computing platforms have become more and more heterogeneous, but the exploitation of these additional capabilities is so complex from the application developer's perspective that they are left unused most of the time, resulting therefore in a supplemental waste of energy rather than in faster processing times. In this paper we present a transparent, on-the-fly optimization scheme that allows a generic application to automatically exploit the available computing units to partition its computational load. We have called our approach Heterogeneous Platform Accelerator (HPA). The idea is to use profiling to automatically select a computing-intensive candidate for acceleration, and then distribute the computations to the different units by off-loading blocks of code to them. Using an NVIDIA Jetson TK1 board, we demonstrate that not only HPA results in faster processing speed, but also in a considerable reduction in the total energy absorbed.