Improving processor lifespan and energy consumption using DVFS based on ILP monitoring

Cyber-physical systems have a major role to play in improving efficiency and safety in transportation networks, healthcare, energy generation and distribution, among other fields. The thermal management of such systems is important since high operating temperature dramatically reduces chip lifetime; such reduced chip lifetime has obvious cost, reliability and environmental implications. This paper presents a mechanism for thermal management. We use dynamic voltage scaling preferentially applied to highly parallel segments of the workload while continuing to guarantee all hard deadlines. Experimental results show that when the workload contains time-varying levels of instruction-level parallelism (as is quite common), this lightweight algorithm improves chip lifetime over previously proposed methods. This approach is easy to implement, requires no hardware modifications and imposes very little overhead.