Heuristic Scheduling of Streaming Applications for Energy Efficiency on Heterogeneous Multicores

A recent manifestation of the trend towards heterogeneous computing is the introduction of heterogeneous multicore processors into the desktop segment, offering the potential for unprecedented energy efficiency in the execution of parallel applications. Such an application may be launched repeatedly, as it is, e. g., the case with stream processing applications, and at the same time exhibit dynamic behavior, for instance due to varying input quality. The scheduler is then challenged with delivering premium schedules within a short time frame to tap the full potential for energy efficiency as well as react to shifting application behavior. We present a heuristic approach to heterogeneous application scheduling, which also takes into account prevalent hardware limitations with regard to frequency scaling. The scheduling heuristic is capable of accomplishing high energy efficiency of the resulting schedules in conjunction with rapid scheduling. This is demonstrated experimentally in a comparison to two static schedulers based on linear programming, which-contrary to our heuristic-cannot account for dynamic application behavior but serve to provide reference values for energy consumption and scheduling time.