Dynamic power management for thermal control of many-core real-time systems

Abstract

Many-core systems, processors incorporating numerous cores interconnected by a Network on Chip (NoC), provide the computing power needed by future applications. High power density caused by the steadily shrinking transistor size, which is still following Moore's law, leads to a number of problems such as overheating cores, affecting processor reliability and lifetime. Embedded real-time systems are exposed to a changing ambient temperature and thus need to adapt their configuration in order to keep the individual core temperature below critical values. In our approach a hysteresis controller is implemented on each core, triggering a redistribution of the cores' workload and the transition into an idle state allowing the core to cool down. We propose two approaches, one global and one local approach, to redistribute the tasks and relive overheating cores during runtime. We evaluate the two proposed approaches by comparing them against each other based on simulations.