Design of Non-Critical Path Resource Distributor for SMT Processors

Abstract

The performance in simultaneous multithreading (SMT) processors is mainly determined by the distribution of the common resources among the threads. However, resource distribution methods often take cycles to calculate a resource allocation solution. Implementing a many-cycles resource distributor in the critical pipeline path will result in an innegligible impact on SMT processor performance. This work proposes a design of non-critical path resource distributor (NCPRD) for SMT processors, which separates the resource distribution from the critical pipeline path to avoid the clock wastage caused by the computation on the allocation solution. Our limit case study shows that, NCPRD benefits from its asynchronous work mode under both throughput and fairness metric in all type workloads but the memory-intensive workloads, and obtains more gains over the critical-path-involved resource distributor when the computation on the allocation solution spends more cycles.