Research On Resource Allocation In Simultaneous Multithreading Processors

Simultaneous Multithreading (SMT) processors improve performance by multiple threads sharing processor resources. The performance of each thread and overall per-formance of the processor are all determined by how to allocate shared resources among the co-scheduled threads, that is, by the resource allocation policy. It is an important project to allocate shared resources efficiently in the research of SMT processors.With SMT processors used widely, besides performance, processor hardware is required to consider fairness and Quality of Service (QoS) of real-time tasks. Current researches on resource allocation policies in SMT processors are mainly concerned with how to improve performance, how to optimize fairness and how to guarantee QoS of real-time tasks. However, for improving performance, there exist low resource utiliza-tion, severe resource under-use and resource abuse in current resource allocation poli-cies. How to improve performance further is still challenging. For optimizing fairness, current researches almost focus on evaluating fairness of previously proposed policies. No policy is presented specially for optimizing fairness. How to implement fairness by efficient recourse allocation is demanding. For guaranteeing QoS of real-time tasks, re-lated works are rare and only one policy allocating resources explicitly is presented. How to guarantee QoS of real-time tasks needs further study. Detailed researches have been done on these three issues. The main contributions are as follows:1. To improve performance further, MFP, a resource allocation policy for SMT proc-essors, is proposed. MFP policy combines the advantages of fetch controlling with those of resource partitioning, and allocates shared resources mainly by fetch con-trolling to allow resources to be competed freely and to increase resource utiliza-tion as a result. Resource partitioning is used to reduce resource under-use and re-source abuse caused by improper resource competition.2. To support fairness better, FAIR, a resource allocation policy for SMT processors, is proposed. FAIR policy allocates shared resources indirectly by fetch controlling, and prioritizes the thread with lower normalized speedup to accelerate it and to make all the co-scheduled threads share resources fairly as a result.3. To guarantee QoS of real-time tasks, CPIT, a resource allocation policy for SMT processors, is proposed. CPIT policy allocates shared resources indirectly by fetch controlling, and duly adjusts the fetch priority of a real-rime task to control its per-formance and to guarantee QoS of a real-time task as a result.4. To optimize fairness at the same time to improve performance, FAIR+, a resource allocation policy for SMT processors, is proposed. FAIR+ policy combines FAIR with any policy purchasing performance as the main optimization goal, making full use of the two policies'advantages to achieve as high performance as possible while optimizing fairness.5. To guarantee QoS of real-time tasks at the same time to improve performance, it is proposed to combine CPIT policy with MFP policy, making full use of the advan-tages of CPIT and MFP to achieve as high performance as possible while guaran-teeing QoS of real-time tasks.Plenty of experiments are made. Results show that all the proposed resource allo-cation policies can achieve their own goals effectively. Our researches make it possible to apply SMT processors to various domains.