Workload type-aware scheduling on big.LITTLE platforms

Abstract

Optimizing energy efficiency in execution strategies has traditionally been heavily influenced by hardware mechanisms such as frequency scaling and core sleep states. With such facilities, the system can be scaled dynamically and on-demand to trade power dissipation for clock speed or parallelism. Determining the most efficient execution configuration has been described in much related work, but few efforts have been put on including the workload type into the calculation. The type of the workload affects both the performance and the power of the processor, and is especially important when considering heterogeneous systems like the big.LITTLE, since different cores handle the workload with different efficiency. In this paper, we demonstrate the influence of the workload type when choosing an optimal execution strategy on a big.LITTLE platform. We implement schedulers capable of including workload type, and we provide a runtime system capable of executing the schedules on a real-world platform. Results demonstrate that including workload types into the scheduler saves between 7.1% and 31.3% of energy in our best/worst corner case studies, a result that should be considered in future implementations of big.LITTLE schedulers.