Research on energy efficient real-time scheduling in multicore processor systems

Abstract

Multicore architecture enables higher throughput with lowered energy consumption than traditional unicore processors. Thus processors with 16 or more cores are expected to be emerged at the market in a decade. The use of hard real-time systems are becoming wider and the functions of them are getting more complex. As a result the multicore systems are expected to be widely used in the real-time systems in the near future. By the way, adopting DVS(dynamic voltage scaling) to the multicore processors have a limitation that all the cores in a processor should operate at a same clock frequency always. This is from the design complexity and manufacturing cost. By the constraint the energy efficiency of a multicore processor is less than that of a multiprocessor system with DVS because when a core requires higher performance than the others all the cores in the processor should operate at the needless high performance. In traditional multiprocessor systems, DVS real-time scheduling is generally done in the way of partitioning the task set into each processors and doing the single processor DVS real-time scheduling for each processors. However, blindly adopting existing DVS algorithms which do not consider the restriction will result in a waste of energy. Based on the partitioning approach this dissertation suggests Dynamic Repartitioning algorithm which redistributes tasks over the cores on the fly to maintain the partitioned task set well-balanced. To assure that the dead-lines are not violated by moving tasks across the cores, the notion of permanent slack and task slack is suggested. A task migration is only allowed when appropriate slack exists in the destination core. The correctness of the suggested method is easily proved. Another major drawback of multicore architecture is relatively high leakage power under low load. The leakage power increases linearly to the number of integrated circuits. Under low load the decreased dynamic power consumption can not offset...