The Complementary Behavior Of Threads On CMT Architecture

In CMT architecture, multiple threads improve resource utilization rate and enhance performance significantly by sharing and competing for various resources on chip. Resource allocation problem has therefore become the core issue of CMT. Fruitful researches from different angles on CMT architecture have been done by predecessors, however few concern the coordination of multithreading, and little conscious attention is paid to the complementarity among threads. Therefore, there is still inefficient use of resources and inappropriate selection in thread scheduling and it is difficult to do in-depth excavation or take advantage of the multithreaded performance potential.To solve these problems, this thesis commences the study from the actual demand of resources of each thread, and makes a deep analysis of the complementarity by combining the theoretical analysis with large amount of experiments. The details are shown as follows:The previous studies generally focused on special pipeline segment or special resource and neglected the impact of the allocation of resources on performance. The research on storage resource allocation, eg IFQ(Instruction Fetch Queue), shows that improper configuration of resources may have a negative impact on overall performance. And the potential interference among various resources and the influences on performance are analyzed.In the process of obtaining resource requirements, typical segments within each thread are extracted first. The needs of RRF (Rename Register File) are analyzed deeply by control experiments of different configuration, and the thread-resource dependent characteristic curves are given, which lay the foundations for analyzing the multithreaded complementarity.According to the curves above, the idea that shared resources should be allocated according to threads'performance potential is proposed. By making use of individual resource rationally, two ideal scheduling schemes MP(Maximization Performance) to maximize performance and MF(Maximization Fairness) to maximize fairness are given. Finally, based on the deep discussion of how resource allocation scheme acts on the multithreading system, the conception of multithreaded complementarity is concreted, and the evaluations of the complementarity at different system targets are quantified as a tool for evaluating and excavating the complementarity.