A Dynamic Resource Allocation Optimization for SMT Processors

Abstract

The threads on simultaneous multithreading (SMT) processors compete for the common resources rather than share them, and meanwhile they take on changing program phases. It is a challenge to meet the changing resource requirements of the threads by making resource allocation according to limited pipeline information. To avoid the stagnation of resource allocation optimization in the local suboptimal resource distributions and perform quick capture on the new optimal resource distributions, this paper proposes a dynamic resource allocation optimization (DRAO) for SMT processors, which optimizes the resource allocation targeting only the throughput performance. An optimization model with consideration of both throughput and fairness performance is designed for DRAO to steer the exploitation on historical allocation and the exploration on new allocation solutions. Results from experiment show that DRAO obtains potential improvement of fairness performance as well as good improvement of throughput performance. These are attributed to the well-balanced exploitation and exploration activities.