Experimental Energy Profiling of Energy-Critical Embedded Applications

International audienceDespite recent advances that have greatly improved the performance of embedded systems, we still face many challenges with regard to energy consumption in energy-constrained embedded and communication platforms. Optimizing applications for energy consumption remains a challenge and thus is a compelling research direction, both on the practical and theoretical sides. This paper presents a new experimental bench for energy profiling of non-performance-critical embedded and mobile applications and reports preliminary results obtained on two embedded boards. The experiments are driven by an online energy monitoring mechanism using National Instruments' cDAQ and LabVIEW running on a host machine. The host monitors a target device, which runs a set of benchmarks. We describe the experience gained from using and modding two different target boards, namely an Nvidia Jetson TX1 and a TI AM572x evaluation module. In particular, we confirm, and thus further validate, the existence of the Energy/Frequency Convexity Rule for CPU-bound benchmarks. This rule states that there exists an optimal clock frequency that minimizes the CPU energy consumption for non-performance-critical applications. We also show that the gain of frequency scaling is highly dependent on workload characteristics. Any future energy-management approach should take these behavioral traits into consideration. I. INTRODUCTION Continuous CMOS technology scaling (Moore's law) increases the on-chip power density due to the higher transistor integration. As power density increases, many factors like power dissipation, leakage, data activity, and electro-migration contribute to higher on-chip temperatures. The increase in temperature leads to an increase in leakage power, thereby increasing the total energy dissipation and thus forming a part of a vicious circle significantly limitting system performance. The bulk of today's computing does not happen on desktops, laptops, servers, or data centers, but rather on embedded media devices like mobile phones [1]. The embedded computing applications running on those devices demand better energy efficiency and flexibility in operation, while delivering better performance per Watt. At the same time, they cannot compete with application-specific integrated circuits (ASIC) in terms of energy efficiency. Indeed, a well-designed ASIC can achieve an efficiency of 5 pJ/op in a 90-nm CMOS process, whereas a very efficient embedded processor would require about 250 pJ/op. That means the embedded processor may consume about 50 times more energy than a custom designed ASIC [1]. Today's system-on-a-chip (SoC) platforms have a lot of software acting in unison, trying to deliver a seamless use