CSR--EHS--Reconciling Operating System and Hardware Thermal Control

点击翻译按钮获取中文摘要

This project develops real-time control of systems with dynamic thermal management (DTM-also called thermal throttling). DTM allows the cooling solution to be designed for less than the worst case, an approach that is becoming necessary with exponentially increasing power densities and shrinking form factors. At runtime, if heat dissipation exceeds the cooling solution's capacity for a sustained period of time and on-chip temperatures approach dangerous levels, DTM places the processor in some form of low-power mode that allows temperatures to fall back to a safe level. The problem is that DTM in current systems is typically engaged autonomously by the hardware when thermal stress is imminent, without any warning. This is incompatible with real-time scheduling. To reconcile runtime thermal management with real-time control, new scheduling algorithms and requisite software and hardware support are being developed in this research. This includes fast estimation of a workload's future thermal load, estimation of the likelihood and scheduling impact of DTM, and new methods that allow the operating system to choose the most convenient time to engage DTM. The proposed techniques are evaluated on a real testbed with realistic workloads.