Approaching a Smart Sharing of Resources in SMT Processors

SMT processors increase performance by executing instructions from several threads simultaneously. These threads use the processor’s resources better by sharing them, but, at the same time, threads are competing for these resources. The way critical resources are distributed among threads determines the final throughput and also the performance of each individual thread. Currently, the processor instruction fetch policy decides each cycle which threads enter the processor to compete for resources. However, these fetch policies only use indirect indicators of how resource allocation is carried out. This may cause resource monopolization by a single thread, or wasted resources when no thread can use them. Both situations can harm the processor performance and occur, for example, after an L2 cache miss. This paper is a first step toward dynamic resource allocation for SMT processors. We show that being conscious about resource demand and directly controlling resource assignment significantly improves performance of SMTs. We introduce for the first time the concept of resource allocation policy in order to provide such a control. Our results show that our resource allocation policy outperforms the best published fetch policies for throughput and fairness, like FLUSH, by 7% on average. In addition, our resource allocation policy does not need to squash instructions from the pipeline, like FLUSH, in order to get this performance improvement. As a result, it reduces dynamic power consumption and hardware complexity.