Research And Implementation Of Container Network Monitoring Technology

Container technology has been widely adopted in network scenarios like cloud data centers due to its advantages of close-to-physical-machine performance,light weight,and high resource utilization.As a result,a large number of container networks have been established,which gives rise to great importance to perform effective monitoring on the container networks.To monitor latency and to automatically manage network monitoring tasks are critical parts of operating and maintaining container networks.On the one hand,latency is an essential metric for interpreting and improving network performance.However,most of the existing approaches for monitoring latency lack consideration of the system overhead and the negative impacts on networks,and hence it will cause more problems in the container networks whose performance is more vulnerable.On the other hand,in the face of the container networks with large topology scales and complicated deployment environments,the traditional methods are not adequate enough to manage the network monitoring tasks,and network monitoring is in urgent need for automated management to reduce the workload and the possibility of errors.It can be seen that the monitoring of container network is under difficulties and challenges.In this thesis,the methods to monitor latency and to manage network monitoring tasks automatically in container networks are focused.The main contributions are as follows:First of all,active monitoring and intrusive monitoring have many drawbacks,for example,they are deficient in elucidating the real performance of application flows,interfere with the regular operation of networks,and take up too much CPU time.To address the problem,this thesis proposes a passive monitoring technology for latency aiming at the container network,which is able to realize the latency monitoring of each data packet within the application flow without injecting monitoring probes into the network or modifying the content of data packets.For the problem that the latency metric is difficult to calculate owing to the lack of unique identification information of data packets,this thesis designs a new data packet identification method based on the packet forwarding principle of the container network,which considers the buffer address of container network packets as a part of the identification information.On this basis,deeply combining the kernel programmability of e BPF with the packet forwarding path of the container network,this thesis proposes an efficient mechanism for collecting packet information in the container network,which is needed by the latency calculation.Experiments show that the technology can effectively monitor the packet latency without affecting the normal operation of the network,and reduce the CPU overhead by 40%-60% compared with the existing similar technologies.Secondly,the traditional network monitoring methods rely heavily on manual operations,which results in the lack of real-time and dynamic.To solve this problem,this thesis proposes a top-down and automated management system for the container network monitoring,which has a simple method of defining monitoring events and can automatically convert high-level monitoring intentions into low-level deployment operations on container networks.Aiming at the problem that there is not enough data of network model to drive the automatic execution when managing the network monitoring tasks,this thesis designs a dynamic network model to accurately describe the real-time state of container networks.On this basis,this thesis proposes a master-slave heterogeneous control framework which not only can achieve the translation and transformation of the upper-level monitoring intent to the underlying application deployment strategy,but also can adapt to the scale expansion of container networks.The analysis of two specific use cases demonstrates the functionality and practicability of the system,and shows that the system can automatically monitor the container network in real time.