Dynamic temperature-aware task scheduling based on sliding window model for MPSoCs

As the power density of modern chips increases drastically, chips are prone to overheating. Thermal hot spots increase cooling costs, negatively impact reliability and degrade performance. A valid task scheduling can reduce chip's average temperature and temperature variations. We propose a dynamic temperature-aware task scheduling policy based on sliding window model. This scheduling policy calculates the probability of task allocation for each core according to current and historical temperatures of the core, and then the one with the maximal probability is chosen to execute the ready task. If multiple cores have the same probability, the scheduler gives priority to the core that has the minimal average temperature of neighbor units. The experimental results show that this scheduling policy can reduce hot spots, decrease spatial and temporal temperature variations of all units, and thus achieve a relatively lower average temperature and more balanced temperature distribution.