Instantaneous IPC based instruction fetch policy for SMT processors

Simultaneous Multithreaded Processors improve the instruction throughput by allowing fetching and executing instructions from several running threads simultaneously in each clock cycle. In this paper, first, the authors introduce simply several instruction fetch policies of SMT processors, and compare their performance on improving IPC of single running workload. Next, an ideal fetch model of instruction fetch policy is given, and a realistic policy named IPCBFP, based on the ideal model, is proposed and analyzed. This policy fetches instructions for a thread according to its instantaneous IPC value and its current instruction number in the instruction queue. Simulation results show that IPCBFP policy can improve the performances of the workloads dramatically. In two-thread and four-thread mix workload experiments, the speedups are 17% and 8% on average respectively. In addition, the sizes and the conflict rates of IQ on average in the experiments with the policy are small than that with ICOUNT. 2.8 policy, which is the best fetch policy up to date to our knowledge. And the authors' policy also has some advantages on the degradation of cache miss rates and TLB miss rates.