Workload Type-Aware Scheduling on big.LITTLE Platforms

Abstract Optimizing energy efficiency in execution strategies has tra-ditionally been heavily influenced by hardware mechanisms such as fre-quency scaling and core sleep states. With such facilities, the systemcan be scaled dynamically and on-demand to trade power dissipationfor clock speed or parallelism. Determining the most efficient executionconfiguration has been described in much related work, but few effortshave been put on including the workload type into the calculation. Thetype of the workload affects both the performance and the power of theprocessor, and is especially important when considering heterogeneoussystems like the big.LITTLE, since different cores handle the workloadwith different efficiency. In this paper, we demonstrate the influence ofthe workload type when choosing an optimal execution strategy on abig.LITTLE platform. We implement schedulers capable of includingworkload type, and we provide a runtime system capable of executing theschedules on a real-world platform. Results demonstrate that includingworkload types into the scheduler saves between 7.1% and 31.3% of energyin our best/worst corner case studies, a result that should be consideredin future implementations of big.LITTLE schedulers.