Study Of Routing Algorithms And Technology Of Circuits For Software Defined Network-on-Chip

With the development of integrated circuits(IC)techniques,the tendencies of increased integration level,reduced sizes of devices and wires,as well as the growing communication amount have taken over current design for chips,the System-on-Chip(SoC)have been proved as a solution to highly complex chip designs.Nevertheless,with the further increase for the number of on-chip processors and the further decrease for process node,the connectingions of shared bus have brought about bottlenecks for SoC communication.To address this issue,the Network-on-Chip(NoC)transplants computer network into the IC field,allowing varied on-chip computing resources to be connected to independent network by means of routers and packet switching,computing units can in turn communicate with each other in forms of data packages.On account of the software and hardware co-design for NoC,the Software Defined Network-on-Chip(SDNoC)separates the control logic and data forwarding of NoC in recent years,on the basis of computing and communication separation.Research works show that the communication network of NoC possesses advantages of parallelism,scalability and reusability,which helps NoC to overcome the communication bottlenecks for highly integrated chips.Among the critical techniques that influence NoC,the routing algorithm can decide the transmission path for data package on the network.The capabilities for data packages to alleviate congestion and balance network load during communication,and the NoC performance in latency and throughput depend on the flexibility of routing algorithm;thus,efficient routing algorithms build the primary approach to design high-performance NoC.Adaptive routing algorithms provide multiple candidate paths and proper selection strategy for data transmissions,which is more approximate to the aim of light-loaded traversing that can bypass heavy congested regions.In this paper,the Odd-Even(OE)turn model is taken as the example to analyze the adaptiveness and congestion information demand for routing algorithms,with arbiter and distant congestion prediction applied as solutions to higher flexibility;in addition,the influences of deterministic,congestion-oblivious and congestion-aware routings to network performances are studied on SDNoC through software control.The major research works and results are listed as follows:(1)By analyzing the demand for bandwidth ratio of bus-shared SoC,along with the advantages of fixed-priority,round-robin and lottery arbiters as well as the thought of two-level arbitration,a two-level reconfigurable arbiter which can grant every request approximate to target bandwidth ratio is proposed,started from endless-loop avoidance and granting accuracy enhancement.C/C++ simulation results for arbiters show that the granted ratio of two-level reconfigurable arbiter is highly identical to the target bandwidth ratio.(2)According to the conditions that multiple candidate output directions are permitted by the routing function,while selection strategy is required to decide the forwarding path,the balanced network load has become the primary design target for NoC arbiter;thus,fairness instead of bandwidth ratio has become the most important performance demand for NoC arbiter.In this paper,the granting scheme of fixed-priority arbiter is referenced to design a fair arbiter that transfers the fixed priority order into varying priority orders according to the pointer,i.e.the alterable priority arbiter(APA).Simulation result based on System C show that the APA can grant all requests more equally than round-robin arbiter,which is proved to be an arbiter with strong fairness.(3)A congestion-oblivious adaptive routing algorithm OE-FA is designed with fair APA applied as selection strategy for OE turn model.For the nodes permitting multiple candidate directions according to OE turn model,OE-FA router regards these permitted directions as requests,the granting result of APA is used to determine the ultimate output,which improves the number of candidate paths in and achieves more balanced load.Simulation results based on Noxim platform and verification results based on SMIC 0.13?m technique show that the OE-FA routing algorithm can enhance the network performance of OE turn model evidently,with negligible price of hardware cost.(4)By adjusting the application field of fair arbiter from all candidate directions to the candidate outputs sharing the same congestion level,a congestion-aware adaptive routing algorithm CAOE-FA is proposed.The selection strategy of CAOE-FA regards the congestion level on downstream router,i.e.the occupied ratio of input buffer,as the primary basis for output determination;for the candidate directions sharing an equal congestion level,the ultimate output is decided by the granting result of APA.CAOE-FA can provide the equally congested downstream routers with the same opportunity to receive data;as a result,the network load can be balanced.Simulation results based on Noxim platform and verification results based on SMIC 0.13?m technique show that although the process of congestion-awareness can increase the latency and power consumption on a router,the congestion alleviation effect of CAOE-FA is obvious,which reduces the network latency and improves the network throughput obviously with acceptable hardware implementation cost.(5)An adaptive routing algorithm LDCAP that can predict distant congestion is designed.On the network beyond next-hop,distant congestion level can also impact the forwarding pressure of NoC;thus,simplex local congestion information is difficult to describe the load statuses faced by candidate direction comprehensively.In this paper,the proposed LDCAP researches on the set of all possible forwarding paths after the departure of data packages at candidate adjacent router,the number of required hops in corresponding candidate direction to reach the destination is applied to predict its potential distant congestion level.When determining the congestion index described by the LDCAP,both local and distant congestions are connected with proper combinational weight;it can be concluded that the prediction scheme of LDCAP can not only alleviate congestion and unbalanced load,but also omit the propagation network and the complex calculation for distant congestion information.Simulation results based on Noxim platform and verification results based on SMIC 0.13?m technique show that the LDCAP possesses favorable network performance and lower hardware cost.