Research On System Modeling And Topology For Hybrid Wireless Network-on-Chip

Nework-on-Chip(NoC)has emerged as an interconnection and communication architecture for complex System-on-chip(So C)paradigm.However,the performance limitations such as high latency and power consumption due to planar multihop wired links hinder seriously the further performance enhancement of multicore So C with the increasing complexity and scale.Recently,the hybrid wireless NoC architecture based on wired/wireless links becomes the mainstream research of current wireless NoC.The system model is the basis of study on performance of hybrid wireless NOC,and the topology structure as one of the key research issues has a decisive role for communication performance of hybrid wireless NoC.The system model and high-performance topology for hybrid wireless NoC are researched deeply in this thesis.We mainly solve the model establishment of intra-chip wireless communication system for hybrid wireless NoC in physical and systematic level,the automatic detection and dynamic bandwidth allocation mechanism based on congestion measurement to hot wireless links for the design of regular universal topology,the wireless link placement and channel assignment,the method of power-interference co-optimization(PICO)for irregular application-specific topology,and the medium access control(MAC)mechanism with low latency for hybrid wireless NoC with large system size and so on.Our work will guide the research on the performance evaluation and design of high-performance topology for hybrid wireless NoC.The major contributions and innovations of this thesis are as follows.To solve the issue of system modeling for intra-chip wireless communication system in hybrid wireless NoC,the models for path loss and latency based on the previous multi-path signal propagation by using the ray tracing method in physical level of hybrid wireless NoC are established firstly.Meanwhile,the statistical analysis models for power consumption and delay are established by using the statistical analysis method in system level,and then we establish the statistical analysis models for delay and power consumption based on the previous physical models.The theoretical analysis and experimental results based on simulation show that the received signal is mainly consisted of the surface wave and guided wave propagated by silicon substrate for the millimeter-wave on-chip antennas operating.The signal from guided wave is negligible gradually with the increasing communication distance.For high resistivity silicon substrate,our model for path loss has an average error of 10.08% compared with the existing model by measured and curve-fitting method.Especially,the average error is smaller than 8.20% between the two models while the communication distance is 5~30mm.Meanwhile,the path loss could be reduced and the coherence bandwidth is close to 100 GHz by using high resistivity silicon substrate.If the optical on-chip antenna operating at THz frequency is integrated to silicon substrate,the signal is mainly from the reflected wave via the reflection of vacuum-package surfaces.Since the reflection coefficient is assumed to be-1 while used on chip,the traditional two-ray model will cause a big error for the path loss.If the permittivity of all the surfaces of vacuum package is decreased,it is beneficial to the reduction of path loss and the increasing of coherence bandwidth.For the universal topology design with unknown traffic characteristics,to solve the problem of adaptive adjustment of the intra-chip topology according to the different communication traffic models,a virtual Torus-based adaptive wireless NoC(VT-AWi NoC)structure is proposed,and thus the overall performance of regular universal topology for hybrid wireless NoC has been improved significantly.The automatic detection and dynamic bandwidth allocation mechanism to hot wireless link based on the sensing parameter for link congestion measurement is adopted.Moreover,the dynamic allocation circuit for transmitter is designed to realize adaptive adjustment of the intra-chip topology and link bandwidth as different traffic patterns.We compare the performance of our proposed VT-AWi NoC to other hybrid wireless NoC topologies.Experimental results show that the proposed topology obtains a delay improvement of 16.52% to 23.27% under random traffic pattern,an energy saving per packet of 39.19% under hotspot traffic pattern,a delay improvement of 17.20% to 21.68% and energy saving per packet of 23.49% on real application such benchmark as FFT(Fast Fourier Transform).To solve the problems of efficient irregular topology generation,the interaction between power consumption and interference,and the limited wireless resources for application-specific design in hybrid wireless NoC,we propose a method of irregular application-specific topology generation based on PICO.The significant performance gains of hybrid wireless NoC are achieved by the frequency reuse.After analyzing the potential factors of high power consumption and interference,the method was implemented to obtain the collaborative optimization between power consumption and network interference by the combination of traffic load under certain signal-to-interference plus noise ratio(SINR).The cycle-accurate simulation results demonstrate that although the performance gains of generated topology using PICO is not very obvious for the small-scale benchmarks,it is significant by the frequency reuse for the larger scale benchmarks.For the benchmark as 12*12 FFT,the topology generated by PICO has 23.56% more throughput and an energy saving per packet of 7.61% than the topology generated by MOWI-E.Moreover,the performance gains are more significant with the increasing IP core size.Finally,to address the two issues including communication authorization of parallel links and unreachable wireless token for the MAC communication mechanism in irregular application-specification topology,a centralized MAC mechanism based on two-level wireless token is proposed in this thesis.Therefore,the problems of channel contention and data transmission collision for the irregular application-specific topology based on frequency reuse are solved.In the centralized MAC mechanism,a channel centralized controller unit(CCCU)placed at the center of the topology is in charge of maintaining several lookup tables for external token control word,therefore,the channel contention is solved by broadcast authorization.The system-level simulation experiments for the centralized MAC mechanism above are carried out in the irregular application-specific topology generated by preceding chapter.The experimental results show that the centralized MAC mechanism could obtain a saturation throughput improvement of 6.14%~12.58%,and a delay improvement of 8.45%~12.76% compared with the centralized MAC mechanism on 8*8 and 12*12 FFT.When compared with the distributed MAC mechanism,the centralized MAC mechanism could obtain a smaller token passing delay and returning period.Furthermore,the centralized MAC mechanism is beneficial to the improvement of throughput and delay.