基于数据流块的空间指令调度方法

发布时间：2018-11-26 15:31

【摘要】：分簇超标量处理器将硬件资源分区来避免大的单体部件导致的功耗与周期惩罚,动态多核处理器融合多个物理核的硬件资源提供适应程序需求的计算能力,这些结构合理使用空间分布的硬件资源实现高能效的计算.空间分区结构中指令负载不均衡和跨区操作数传递延迟等问题可导致性能惩罚,需要有效的指令调度方法将计算在分区间进行分布.提出了基于数据流块(data-flow block,DFB)的空间指令调度方法.DFB是动态构建、缓存并重用的一个或数个顺序执行的指令基本块的调度模式.DFB调度算法建模动态指令流中的数据流约束和硬件资源定义的调度空间,然后根据指令量化的相对关键性完成调度决策.介绍了DFB调度的微结构框架和算法.通过对分区数、分区间延迟和调度窗口容量等与调度方法密切相关的微结构参数的实验,证明了DFB调度的性能和稳定性优于负载均衡调度和基于依赖的调度.最后举例证明结合一种数据流块缓存实现的DFB调度达到的调度效果接近理想化的DFB调度.
[Abstract]:Cluster superscalar processors partition hardware resources to avoid power consumption and periodic penalty caused by large individual components. Dynamic multicore processors integrate hardware resources with multiple physical cores to provide computing power to meet program requirements. These structures use spatially distributed hardware resources to achieve high energy efficiency. The problems of instruction load imbalance and transmissibility delay in spatial partition structure can lead to performance penalty. It needs an effective instruction scheduling method to distribute the calculation in the sub-interval. A spatial instruction scheduling method based on data flow block (data-flow block,DFB) is proposed. DFB is dynamically constructed. Scheduling pattern of one or more sequential execution blocks of instructions cached and reused. DFB scheduling algorithm models data flow constraints and scheduling spaces defined by hardware resources in a dynamic instruction stream. Then the scheduling decision is completed according to the relative key of instruction quantization. This paper introduces the framework and algorithm of DFB scheduling. The performance and stability of DFB scheduling is proved to be superior to that of load-balanced scheduling and dependent-based scheduling through experiments on the number of partitions, interval delay and scheduling window capacity, which are closely related to the scheduling method. Finally, an example is given to prove that the DFB scheduling implemented with a data stream block cache achieves a scheduling effect close to the idealized DFB scheduling.
【作者单位】：计算机体系结构国家重点实验室(中国科学院计算技术研究所);中国科学院大学计算机与控制学院;杭州电子科技大学信息与控制研究所;
【基金】：国家重点研发计划项目(2016YFB0200501) 国家自然科学基金项目(61332009,61521092,61671196,61327902) 数学工程与先进计算国家重点实验室开放基金项目(2016A04) 北京市科委科技计划专项项目(Z15010101009)~~
【分类号】：TP332

【相似文献】