Research On Data-Driven Adaptive Routing Technology For Network-on-Chips

With the development of many-core architecture,network-on-chip has become a new type of on-chip communication scheme that replaces traditional bus architecture.And the emergence of new applications and protocols has led to an exhibit complicated and time-varying traffic patterns.With the diversified traffic patterns,alleviating congestion and improving network performance have become challenging multi-objective optimization problems.While witnessing most existing heuristic adaptive routing algorithms fail to address the above challenges well,this study aims to explore a new approach of thinking and extracting insights from network behaviors to address the above difficulties.In this paper,we propose a novel data-driven adaptive routing algorithmic framework,which uses reinforcement learning technology to adaptively select routing paths according to network state,and dynamically adjust decision strategies based on feedback generated by the network.It is able to effectively isolate endpoint congestion when facing adversary hot-spot and bursty traffic,and achieve dynamic load-balancing and mitigate network congestion when meeting heavy uniform traffic.Furthermore,we conduct extensive experiments and diversified traffic to evaluate our design.The experimental results show that,the proposed framework can reduce average packet latency and improve network performance under synthetic traffic patterns and high-performance computing(HPC)benchmark.Compared with the Footprint routing algorithm,the average reduction is 17.16%,26.27% and 11.63% of average packet latency under the Uniform,Bitcomp and Transpose traffic patterns,and the average reduction is 9.86% of average packet latency under the HPC benchmark GPCNe T.There are many advantages of adaptive routing,such as load balancing,solving congestion problems and fault tolerance,so it has been widely studied by researchers.However,the adaptive routing algorithm will cause out-of-order packets that belong to the same flow when they arrive at the receiver.This requires setting a reordering buffer at the destination node of the Network-on-Chips(No Cs)to reorder the packets,so that the packets can be received and processed by the applications,but reordering buffer incurs extremely high storage overhead.Therefore,it is a challenge to maintain packet order while multipath transmissions.In this paper,an adaptive proactive traffic limiting scheme is proposed through the analysis of experimental data,which adaptively controls the injection rate of packets according to the network state,so that the packets are queued on the endpoint.At the same time,it is combined with a priority-based virtual channel allocator to help alleviate the out-of-order degree.Finally,we conduct extensive experiments to evaluate our scheme.The experimental results show that the proposed scheme greatly reduces the out-of-order degree of packets without affecting the performance of adaptive routing algorithm under various synthetic loads and hotspot traffic patterns.On average,the out-of-order degree is reduced by an average of 27.16%and 31.3% under the Uniform and Transpose traffic patterns,respectively.And the out-of-order degree is reduced by an average of 29% under the Hotspot traffic pattern.