Providing Fairness in Heterogeneous Multicores with a Predictive, Adaptive Scheduler

Multicore applications contend for resources -- especially memory bandwidth -- reducing both quality-of-service and overall system performance. Contention-aware schedulers have been proposed to provide fairness and predictable behavior through thread-level scheduling. Prior approaches have two drawbacks, however. First, many introduce overhead that reduces overall performance. Second, the emergence of heterogeneous multicores has made handling contention and providing fairness much more difficult as the scheduler must now account for both application interference and the performance effects of different core types. This paper proposes augmenting existing contention-aware approaches with predictive and adaptive components to providefair memory access and performance improvements on heterogeneous multicores. The predictive component's closed-loop approach anticipates how different processes will perform with different core types, while the adaptive component dynamically tunes key scheduling parameters to the current workload. We implement and evaluate this approach on a real Linux/x86system with a variety of memory and compute intensive benchmarks. We find that adding prediction improves fairness and performance by 38% and 4% (respectively) compared to aprior state-of-the-art contention-aware approach. The additionof adaptation allows users to select for fairness or performance optimization, providing an additional 24% improvement infairness or a 9% improvement in performance beyond the predictive approach.