The Research And Implementation Of Regional Centralized Global Distributed Network-on-Chip

With the development of semiconductor technology,more and more processor cores are integrated into single chip,the tradional interconnection technology couldn't satisfy the requirements for performance,power,scalablity and reliability etc.Thus,Network-on-Chip is proposed and hopeful to become the interconnection technology in future multicore chip.With the continuous improvement of integration,the tradional Network-on-Chip technology,constrainted by limited on-chip resource,also faces a series of chanllanges in performance,power and efficiency etc.Therefore,the development of high performance,low power,high efficient Network-on-Chip is still a research difficult.In this dissertation,the hierarchical structure,key technology and performance evaluation of Network-on-Chip are analyzed,and the relations between these three aspects are discussed.The research is mainly focued on the topology,routing algorithm,switching technology,congestion control and router microarchitecture,and then make some optimization on the current technology.The major researches and findings are listed as follow:1.Regional centralized global distributed Network-on-Chip architectureBecause of the large overhead and low efficiency in traditional Network-on-Chip architecture,Regional centralized global distributed network architecture(RCGD_Mesh)is proposed by making a trade-off between centralized interconnection architecture and distributed interconnection architecture.2D Mesh topology is employed in RCGD_Mesh as the backbone network.Link allocators are introduced between adjacent rows and column,so that RCGD_Mesh network could be devided into horizontal sub-network and vertical sub-network.Based on the hybrid network topology,RCGD_Mesh proposed a minimal oblivious routing algorithm,which includes three parts: packet injection policy,link allocation algorithm and packet forward policy.The network traffic is distributed into the two sub-networks by packet injection policy,and then the link resources are allocated to each packet based on the regional link usage status.These two measures could effectively balance network workload,improve link utilization,and alleviate network congestion.The experiment results show that RCGD_Mesh could evidently improve the saturation throughput,reduce network latency,and achieve better load balance.2.RCGD_Mesh based dual priority congestion-aware routing algorithmThe relationship between resource contention and network congestion is elaborated and the influence of network congestion to performance is analyzed.The current research and technology for network congestion are discussed.To overcome the drawback of RCGD_Mesh,a dual priority congestion-aware adaptive routing algorithm(DP-CASR)is proposed.To improve link utilization and routing efficiency,based on RCGD_Mesh,the congestion control metric is introduced into DP-CASR algorithm.According to local congestion information,an adaptive packet injection policy is proposed in router.At the same time,router,which takes the hardware cost and efficiency into consideration,employs a deterministic routing policy.Combined with the local congestion information and regional congestion,the link resources are controlled by link allocators.To improve link utilization and alleviate congestion,a dual priority allocation policy is proposed.From the evaluation,it could be found that DP-CASR in 4×4 network achieves maximum throughput in uneven traffic patterns.These facts indicate that DP-CASR could effectively alleviate congestion and avoid network congestion.In 4×4 and 8×8 network,DP-CASR_FULL could improve the network saturation throughput by 68.4% and 114%,respectively.Moreover,through the analysis on the network traffic destirbution and link utilization,DP-CASR could evidently improve the link utilization on the basis of traditional routing algorithm and RCGD_Mesh.3.Shoot through routing scheme for regional centralized global distributed No C architectureThe weakness of traditional routing algorithms and RCGD_Mesh network are discussed.Depending on these shortages,combined with Torus network and the regional centralized global distributed architecture,the Torus-based hybrid network topology(THNo C)is proposed,and then the shoot through routing scheme(STR)is proposed based on the proposed topology.According to different application scenarios,the high-performance shoot through routing algorithm(HPSTR)and low-cost shoot through routing algorithm are proposed.Compared with the conventional routing algorithms,the shoot through routing scheme could evenly distribute the traffic onto the whole network.As a result,STR routing could avoid the congestion region in the centre of the network and improve the link utilization of the edge links.Moreover,STR routing could evidently shorten the routing path and improve routing efficiency.The experiment results show that STR could effectively reduce the non-blocking latency.HPSTR could achieve better network performance,and LCSTR could reduce the hardware overhead with the same performance.In the field of workload distribution,STR routing,which breaks the limitation in traditional routing algorithms,could evenly distributed traffic onto the whole network and achieve better load balance.4.The optimization of the switching technology based on forward-pretreatment methodThe worm-hole switching technology is described and the corresponding disadvantages are also analysed.The relationship between swithcing technology and network performance is discussed.Furthermore,the influence of common optimization techniques to network performance and the trade-offs between different performance indicators are also elaborated.Based on these analyses,the forward-pretreatment based switching technology is proposed,which makes the routing computation and data transmission proceed in parallel.This method could improve the link utilization by avoiding “data stall”,so as to improve network performance.From theoretical analysis,it is shown that forward-pretreatment based switching technology could achieve 100% link utilization.In evaluation,the proposed switching technology is introduced into XY routing and RCGD_Mesh,and then these designs are evaluated compared with traditonal switching technology.The experimental results show that forward-pretreatment based switching technology could effectively improve the network performance and achieve the maximum link utilization.