Research On Key Technologies Of Data-driven Intelligent Routing

With the development of network technology,the scale of the network is becoming larger and larger,while the types of network services and applications are also expanding.The demand for massive data transmission tasks is no longer limited to accessibility,rather diverse requirements of transmission qualities,which poses a huge challenge to the the network services.As a critical part of the network architecture,routing is one of the core technologies for data switching,which provides forwarding policy support for data transmission.However,the traditional "best-effort" scheme struggles to cope with the new demands and challenges brought by the flourishing of networks.Thus,in this dissertation,intra-domain routing,inter-domain routing,and routing entry managements are investigated in detail.Investigations are focused on four aspects,including the intra-domain multi-demand routing,the acceleration of routing decision model,the awareness of intra-domain state for inter-domain routing,and the route entry management of lightweight Ternary Content Addressable Memory(TCAM).This dissertation aims to systematically provide technical support and solutions for network transmission from the routing involved fields,by targeting the implementation of flexible and efficient routing policies and route entry management.Specifically,the four main research contents and contributions are as follows:1.A reinforcement learning-based intra-domain multi-demand intelligent routing scheme was proposed,which can be generically applied to cumulative and bottleneck types of metrics.The multi-demand routing algorithm based on model fusion can efficiently cope with changes of transmission types.The experimental results based on real network topologies indicated that the proposed scheme can reduce the flow completion time by 53.1%compared to the traditional link-state-based routing protocols on the premise of meeting bandwidth requirements,which could be better to support the development of the future network.2.A decision model reconvergence acceleration scheme was proposed,which aims to alleviate the impact of topological dynamics of networks on the convergence of reinforcement learning-based routing models.The acceleration scheme involved two components:1)the non-tightly coupled neural network structure is highly scalable to dynamic changes in transmission requirements and network topology;2)the acceleration module based on the federation learning framework and digital twin network technology can further improve the reconvergence efficiency of the decision model by more additional state interactions.The experimental results conducted on the real network topologies showed that the proposed scheme can reduce the reconvergence time over the state-of-the-art algorithm by about 22.3%.3.A privacy-preserving intra-domain state-aware inter-domain routing protocol is proposed,which intends to bridge the information barriers between each domains that makes existing distributed inter-domain protocols fails to provide performance-guaranteed routing decisions.In consideration of data privacy and network security,it is impossible to share intra-domain network state data among domains.In this context,the protocol exploits,topology abstraction,random number confusion and homomorphic encryption,three strategies to desensitize the private intra-domain data to provide multi-metric path performance evaluations for inter-domain routing.According to the experimental results based on five real inter-domain topologies,the proposed scheme can reduce flow completion time by 35.0%on average compared to Border Gateway Protocol(BGP)and can also provide more flexible routing policies for multi-demand inter-domain transmission tasks.4.A lightweight TCAM-based routing entries management scheme was proposed,which is used to relieve the massive demand on TCAM resources from the explosive growth of routing entries.The AI-based prediction module provides hot entries with guaranteed accuracy and timeliness,which will be efficiently inserted into TCAM in accordance with the block-based entry constraint of the entry update module.The experimental results conducted on real backbone network traffic demonstrated that the proposed scheme can achieve comparable route entry lookup capabilities with the existing schemes by only using 1/8 of TCAM on-chip resources.