面向可重构众核处理器的子网划分与广播机制研究

发布时间：2018-02-13 04:46

本文关键词： 众核处理器片上网络（NoC）广播 XY路由子网划分　出处：《上海交通大学》2013年硕士论文　论文类型：学位论文

【摘要】：基于片上网络（Networks-on-Chip，NoC）互连结构的众核处理器（Many-core）拥有数目众多的计算资源，具有高性能并行处理能力，是未来微处理器体系结构发展的趋势。随着核数的增加，如何有效利用大量处理器资源，提高系统吞吐率是众核处理器体系结构面临的重要问题。同时，随着处理器核的数量和片上网络复杂度的增加，会产生Cache的“一致性墙”（Coherence Wall）问题，急剧增长的Cache一致性代价限制了Cache性能的提升。对局部若干处理单元进行合并构成虚拟计算群（Virtual ComputingGroup，VCG），能有效提高计算资源的利用率。面向片上网络，采取子网划分策略，子网划分达到了子网间有效的通路隔离，有利于Cache一致性的正确实现，并使面向全网的Cache一致性限定在了子网内，大大降低了Cache一致性带来的代价。本文提出了逻辑子网（LogicSub-network）和物理子网（Physical Sub-network）的概念，逻辑子网保证了子网间有效的逻辑隔离，物理子网保证了子网内通信路由的正确实现，相比较前人研究，，此方案能进行任意子网划分，大大提升了子网划分的自由度。与此同时，对于一些特定的Cache一致性协议，其存在大量广播通信特性，本文实现了一种基于树形的XY路由广播机制，能有效降低片上网络内广播通信带来的功耗、通信延迟开销。本课题基于Gem5多核仿真平台，实现了一套基于子网划分的XY路由广播机制，实验结果表明，对比基于单播的广播机制，片上网络性能得到了极大改善。
[Abstract]:Many-core processor based on on-chip network network on-ChipNOC interconnection architecture has a large number of computing resources and high performance parallel processing capability, which is the development trend of microprocessor architecture in the future. How to effectively utilize a large number of processor resources and improve system throughput is an important problem facing the architecture of multi-core processors. At the same time, with the increase of the number of processor cores and the complexity of on-chip network, The consistency wall of Cache will arise, and the increasing consistency cost of Cache limits the improvement of Cache performance. The virtual computing group virtual computing group (VCGG) is formed by merging some local processing units, which can effectively improve the utilization rate of computing resources. In view of the on-chip network, the strategy of subnet partition is adopted, and the subnet partition achieves the effective path isolation between subnets. It is beneficial to the correct realization of Cache consistency, and makes the Cache consistency for the whole network limited to the subnet, which greatly reduces the cost of Cache consistency. In this paper, the concepts of logical Sub-Network and physical Sub-Network are proposed. Logical subnet ensures effective logical isolation between subnets, physical subnet ensures the correct realization of communication routing in subnet. Compared with previous studies, this scheme can be divided into arbitrary subnets. It greatly improves the degree of freedom of subnet partitioning. At the same time, for some specific Cache conformance protocols, there are a large number of broadcast communication characteristics. In this paper, a tree based XY routing broadcasting mechanism is implemented. It can effectively reduce the power consumption and the communication delay overhead caused by the on-chip broadcast communication. An XY routing broadcasting mechanism based on subnet partition is implemented based on Gem5 multi-core simulation platform. The experimental results show that compared with unicast broadcast mechanism, the performance of on-chip network is greatly improved.
【学位授予单位】：上海交通大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP368.1

【参考文献】