同时多线程结构的2级调度策略  

Improving Two Level Thread Schedule Policy for SMT Architecture

作　　者：王晶[1] 樊晓桠[1] 张盛兵[1] 王海[2] 

机构地区：[1]西北工业大学计算机学院,陕西西安710072 [2]同济大学机械学院热能与环境研究所,上海200092

出　　处：《西北工业大学学报》2007年第3期433-437,共5页Journal of Northwestern Polytechnical University

摘　　要：微处理器的发展已经迈入多线程时代。在同时多线程(SMT)结构中,多个线程共享一个取指单元和指令队列(IQ),如何提高有效取指带宽和IQ利用率是关键问题。论文详细分析了SMT中导致IQ阻塞的因素,以及它们对IPC的影响。基于线程预构思想,将导致IQ阻塞的多个因素分开考虑,提出将线程调度分成预构取指调度和队列取指调度的2级调度。采用基于置信度的线程调度策略用于线程预构,采用基于阈值门限的调度策略用于队列取指。仿真结果表明,2级调度策略与ICOUNT策略相比,错误路径指令可减少18.22%(取指级)、5.93%(执行级),IQ阻塞的概率减少27%(整数队列),IPC提高9.4%。Aim. The full paper starts with a review of relevant papers, Tullsen put forward the ICOUNT strategy for SMT （Simultaneous Multi-Threading） architecture, which is still widely used. But, in our opinion, ICOUNT strategy can be much improved to solve better the well known IQ （Instruction Queue） clog problem. We now present a better two-level thread schedule policy based on the idea of thread pre-construction. In the full paper, we explain our better policy in detail. In this abstract, we just add some pertinent remarks to listing the three topics of explanation. The first topic is： the simulation environment. The second topic is. the analysis of the factors influencing IQ efficiency. In the second topic, we analyze the effect on IPC （Instructions per Cycle） of each factor. The third topic is. our two-level thread schedule policy. The third topic is the most important; among many problems discussed in this topic, we select only two in this abstract. The thread schedule policy based on confidence estimation is used in thread pre-construction stage, where selected instructions are put into fetch queue. We use the policy based on threshold in fetch queue, aiming to reduce the number of long latency instructions. Finally, we obtained simulation results with SMTSIM toolset, shown in two figures in the full paper. These results show preliminarily that, as compared with traditional ICOUNT strategy, our two-level thread schedule policy can： （1） reduce wrong instruction ratio by 18. 22% for fetching and 5.93% for executing; （2） reduce IQ clog by 27% for integer queue; （3） increase IPC by 9.4%.

关 键 词：同时多线程 指令队列 置信度估计 取指调度 线程调度 

分 类 号：TP302.1[自动化与计算机技术—计算机系统结构]