电子与信息学报
電子與信息學報
전자여신식학보
JOURNAL OF ELECTRONICS & INFORMATION TECHNOLOGY
2014年
10期
2536-2540
,共5页
孔峰%韩国栋%沈剑良%简刚
孔峰%韓國棟%瀋劍良%簡剛
공봉%한국동%침검량%간강
片上网络%层次化%流量模式%网络直径
片上網絡%層次化%流量模式%網絡直徑
편상망락%층차화%류량모식%망락직경
Network-on-Chip (NoC)%Hierarchical%Traffic pattern%Network diameter
针对片上网络(NoC)较远距离节点路由跳数较大导致的网络功耗和面积过大问题,该文通过分析Mesh和集中式Mesh(CMesh)结构特性,提出一种基于Mesh的新型层次化CHMesh结构。该结构分两层,底层以Mesh方式互连,并划分为多个路由区域,以保证邻近节点的通信需求,上层以CHMesh方式通过中间节点将底层各个区域进行互连,以降低网络直径。设计了针对性最短路径CHXY路由算法,该算法复杂度低,能够避免死锁。性能分析和仿真实验表明,在非均匀流量模式下,CHMesh结构的吞吐量比传统Mesh和Ref-Mesh分别提高约60%和10%,在较大规模片上网络中更有优势。
針對片上網絡(NoC)較遠距離節點路由跳數較大導緻的網絡功耗和麵積過大問題,該文通過分析Mesh和集中式Mesh(CMesh)結構特性,提齣一種基于Mesh的新型層次化CHMesh結構。該結構分兩層,底層以Mesh方式互連,併劃分為多箇路由區域,以保證鄰近節點的通信需求,上層以CHMesh方式通過中間節點將底層各箇區域進行互連,以降低網絡直徑。設計瞭針對性最短路徑CHXY路由算法,該算法複雜度低,能夠避免死鎖。性能分析和倣真實驗錶明,在非均勻流量模式下,CHMesh結構的吞吐量比傳統Mesh和Ref-Mesh分彆提高約60%和10%,在較大規模片上網絡中更有優勢。
침대편상망락(NoC)교원거리절점로유도수교대도치적망락공모화면적과대문제,해문통과분석Mesh화집중식Mesh(CMesh)결구특성,제출일충기우Mesh적신형층차화CHMesh결구。해결구분량층,저층이Mesh방식호련,병화분위다개로유구역,이보증린근절점적통신수구,상층이CHMesh방식통과중간절점장저층각개구역진행호련,이강저망락직경。설계료침대성최단로경CHXY로유산법,해산법복잡도저,능구피면사쇄。성능분석화방진실험표명,재비균균류량모식하,CHMesh결구적탄토량비전통Mesh화Ref-Mesh분별제고약60%화10%,재교대규모편상망락중경유우세。
As the number of modules in System-on-Chip (SoC) increases, the topology is more likely to suffer from excessive end-to-end hop-counts, causing an increase of power consumption and area overhead. Concerning this issue, a novel Mesh-based Hierarchical topology called CHMesh is proposed, which is divided into two levels. The bottom level is interconnected with Mesh and divided into several regions, so as to guarantee communications of adjacent nodes, and the upper level employs intermediate nodes to promote the interconnection among different bottom routing regions with CMesh, so as to decrease the network diameter. Correspondingly, this article elaborates on a shortest-path CHXY routing algorithm, which has a low complexity and can realize deadlock avoidance. Performance analysis and experimental results demonstrate that, compared with traditional Mesh and Ref-Mesh, the CHMesh can increase the average throughput by about 60% and 10% respectively under non-uniform traffic patterns, presenting more advantages on large-scale NoC.