ConfigReco: Network Configuration Recommendation With Graph Neural Networks
Zhenbei Guo, Fuliang Li, Jiaxing Shen, Tangzheng Xie, Shan Jiang, Xingwei Wang
Abstract
Configuration synthesis is a fundamental technology in the context of self-driving networks, aimed at mitigating network outages by intelligently and automatically generating configurations that align with network intents. However, existing tools often fall short in meeting the practical requirements of network operators, particularly in terms of generality and scalability. Moreover, these tools disregard manual configuration which remains the primary method employed for daily network management. To address these challenges, this paper introduces ConfigReco, a novel, versatile, and scalable configuration recommendation tool tailored for manual configuration. ConfigReco facilitates the automatic generation of configuration templates based on the network operator’s intent. First, ConfigReco leverages existing configurations as input and models them using a knowledge graph. Second, graph neural networks are employed by ConfigReco to estimate the significance of nodes within the configuration knowledge graph. Lastly, configuration recommendations are made by ConfigReco based on the computed importance scores. A prototype system has been implemented to substantiate the effectiveness of ConfigReco, and its performance has been evaluated using real-world configurations. The experimental results demonstrate that ConfigReco achieves a coverage rate of 93.35% while concurrently maintaining a redundancy rate of 23.07% within a configuration knowledge graph comprising 890,464 edges and 40,885 nodes. Furthermore, ConfigReco exhibits high scalability, enabling its applicability to arbitrary datasets, while simultaneously providing efficient recommendations within a response time of 1 second.