Meeting Power Constraints While Mitigating Contention on Clustered Multiprocessor System

Mobile and embedded devices have adopted the clustered multiprocessing architecture paradigm in order to serve the increasing workload demands and control/reduce their power consumption. Cores of the same type are integrated into clusters and cluster-wide voltage/frequency scaling mechanisms help them operate under specific power constraints. In this letter, we propose a methodology that: 1) enhances performance by addressing contention on shared resources of the cluster and 2) pro-actively predicts the maximum frequency under which the cluster should operate in order to meet a user-defined power threshold. Experimental results on an Odroid-XU3 board show that the proposed methodology meets the power thresholds, achieving at the same time high performance. Particularly for 7-W threshold, it achieved 5% performance gain over Linux which at the same time consumed 8-W approximately.