Temperature-aware Task Partitioning for Real-Time Scheduling in Embedded Systems

Both power and heat density of on-chip systems are in- creasing exponentially with Moore's Law. High temperature negatively affects reliability as well the costs of cooling and packaging. In this paper, we propose task partitioning as an effective way to reduce the peak temperature in embedded systems running either a set of periodic heterogeneous tasks with common period or periodic heterogeneous tasks with individual period. For task sets with common period, experimental results show that our task partitioning algorithms is able to reduce the peak temperature by as much as 5.8°C as compared to algorithms that only use task sequencing. For task sets with individual period, EDF scheduling with task partitioning can also lower the peak temperature, as compared to simple EDF scheduling, by as much as 6°C. Our analysis indicates that the numbers of additional context switches (overhead) is less than 2 per task, which is tolerable in many practical scenarios.