Research On Light-Weight Fine-Grained In-Band Network-Wide Telemetry

In recent years,with the vigorous development of the Internet industry,the scale of network traffic and infrastructure is experiencing an explosive growth.The complex network topology has put a heavy burden on the operation and maintenance task.How to measure the fine-grained network state to assist the automation management task is an important issue.With the proposal of protocol oblivious forwarding and corresponding programming language,it is possible to monitor device-internal states only through data plane.In band network telemetry(INT)technology has came into being.INT provides the interface between data packet and data plane device,which can collect fine-grained device-internal states without the intervention of the control plane.Although the equipment status collected by INT brings the hope of improving the visibility of network status,it is still faced with the problems of high telemetry overhead of data plane,difficult to realize uninterrupted telemetry,high bandwidth overhead of telemetry information uploading through the southbound interface,and high overhead of the controller CPU for processing the INT information.In this paper,a series of researches have been conducted to address the above problems,and the research results are as follows:1.We designed INT-path,a source routing-based centralized nonredundant in-band network telemetry architecture,which can monitor the entire network with a theoretical minimum path number,and greatly reduce the overhead of dataplane bandwidth and controller CPU for telemetry information analysis.In order to eliminate redundant detection and alleviate the overhead of telemetry on data plane bandwidth,this paper introduces the DFS-based and Euler trailbased path planning policy to generate non-overlapped probing paths,which significantly reduce the telemetry overhead on the dataplane.INT-path can "encode" the network-wide traffic status into a series of"bitmap images",transforming network troubleshooting into pattern recognition problems,which can well serve the intelligent operation and maintenance task.2.We propose INT-label,a probeless In-band Network-Wide Telemetry architecture via distributed labelling,which could reduce the telemetry overhead of dataplane bandwidth.In order to seamlessly adapt to occasional topology changes,this paper realizes non-redundant distributed telemetry based on "probeless" architecture,that achieves 7*24 telemetry.In addition,in order to reduce the overhead of controller CPU for telemetry information analyzing,we proposes the probabilistic labelling algorithm based on the INT number carried by packets.Furthermore,to avoid telemetry resolution degradation due to loss of labelled packets,we also design a feedback mechanism to adaptively change the instant label frequency,which can still maintain high telemetry coverage when the network state is poor.3.We propose INT-filter,a prediction-based southbound redundant telemetry information filtering architecture,that reduce the bandwidth costs of southbound to upload telemetry information.Telemetry information not only occupies the transmission bandwidth of the dataplane,but also needs to be uploaded to the controller through the southbound interface for further analysis and processing.To solve the problem that telemetry information occupies a large amount of bandwidth of the southbound interface and affects other high priority tasks,we propose polynomial fitting-based telemetry information prediction system.Based on the INT-path,INT-filter replaces unnecessary upload with prediction,further reducing the overhead of network wide telemetry.In addition,we design an integration mechanism that conducts predictions using multiple methods simultaneously,which could significantly reduce the bandwidth overhead of INT telemetry data collection.