Methods for Modeling Resource Contention on Simultaneous Multithreading Processors

Simultaneous multithreading (SMT) seeks to improve the computation throughput of a processor core by sharing pri-mary resources such as functional units, issue bandwidth, and caches. SMT designs increase utilization and generally improve overall throughput, but the amount of improvement is highly dependent on competition for shared resources be-tween the scheduled threads. This variability has implica-tions that relate to operating system scheduling, simulation techniques, and fairness. Although these techniques recog-nize the implications of thread interaction, they do little to profile and predict this interaction. The modeling approach presented in this paper uses data collected from performance counters on two different hard-ware implementations of Pentium-4 Hyper-Threading pro-cessors to demonstrate the effects of thread interaction. Techniques are described for fitting linear regression mod-els and recursive partitioning to use the counters to make online predictions of performance (expressed as instruc-tions per cycle); these predictions can be used by the oper-ating system to guide scheduling decisions. A detailed anal-ysis of the effectiveness of each of these techniques is pre-sented. 1