Attaining performance fairness in big.LITTLE systems

摘要

In the dark silicon era, achieving energy-efficient execution on highly heterogeneous platforms requires a much tighter integration of hardware and software solutions. To tackle this problem, we propose a framework for energy-aware task management in heterogeneous embedded systems with several multi-core clusters. The proposed framework integrates a set of novel application-aware management mechanisms for efficient resource utilization, frequency scaling and task migration at intra-cluster and inter-cluster levels. These mechanisms rely on a new management concept, which promotes performance fairness among running tasks as a mean for attaining energy savings, while respecting the target performance of parallel applications. With this purpose, the proposed framework integrates several components for highly accurate run-time monitoring and application self-reporting. Experimental results show that the proposed framework allows achieving energy savings of up to 38% in a state-of-the-art embedded platform, while maintaining performance fairness for a set of real-world SPEC CPU2006 and PARSEC benchmarks.