Thermal-Aware Scheduling for Real-Time Applications in Embedded Systems

The emergence of multi-core and many-core architectures in embedded systems has increased the energy and thermal challenges faced by such systems. As the technology for microprocessors moves toward the nanometer scale, power density becomes one of the major constraints to the performance improvement of embedded processors. Realtime embedded signal processing tasks on multi-core systems have a high potential to thermally stress the die and negatively impact processor performance. This potential problem is crudely regulated by Dynamic Thermal Management (DTM) in commercial processors. DTM is typically realized by gauging the chip temperature at run-time to enable dynamical adjustment of the processor voltage and frequency (DVFS); it slows down the processor upon die overheat. However, real-time embedded systems often cannot tolerate the implied performance degradation entailed in DVFS techniques should any job fail to complete within its deadline as a result of processor slowdown. Consequently, DTM alone cannot avoid real-time application failures completely in the event of die overheat.