A Swarm-Inspired Resource Distribution for SMT Processors

The performance in Simultaneous Multi-Threading (SMT) processors is mainly determined by the distribution of the common resources among the threads. However, the threads exhibit dynamically complicated behavior while they compete for resources at runtime. It is a challenge to meet the changing resource requirements of the threads. This work proposes a Swarm-inspired Resource Distribution (SRD) policy to address the dynamic optimization problem of resource distribution for SMT processors, which uses the runtime performance to guide the generating of trial distributions. A computational model is established by adaptation of swarm intelligence to direct the social exploitation and self exploration activities of the trial distributions in the dynamic optimization environment. Results from simulation show that, benefiting from the good cooperation between SRD's social exploitation on historical experience and self exploration of new solutions, SRD obtains satisfying improvements of both throughput and fairness performance, especially in complicated SMT environment.