| Guaranteeing network reliability is very important for data center fat tree networks,where detection and location of gray faults plays an important role.Thanks to the emergence and development of data plane programmable technology,many telemetry schemes have been proposed to diagnose data center network faults.However,they are insufficient in detecting and locating gray faults.In addition,telemetry solutions usually need to verify their feasibility and working effects in network simulation platform.However,existing network simulation platforms do not natively support telemetry verification.As an open source,data plane high-level programming language,P4（Programming Protocol-Independent Packet Processors）has attracted widespread attention from academia and industry since its birth.P4 is designed for programming reconfigurable networks,which also defines the telemetry standard.Therefore,this thesis proposes a P4-based gray fault detection and location technology for data center fat tree network to achieve high accuracy with low overhead and designs a network simulation platform for P4 telemetry verification.The main work of this thesis is as follows:（1）We analyze existing fault detection mechanisms in data center network and conclude that existing solutions are insufficient in accuracy and bandwidth overhead in detecting gray faults.To improve detection efficiency,we design a new detection and location technology for gray faults in fat tree network and implement it in network simulation platform.First of all,we describe the overall structure of the proposed technology.We then give a detail description on the working process of detection and localization.After defining and analyzing the problem of path planning for fat tree network,we implement a P4-based switch that supports traditional Layer 2 forwarding,source routing forwarding and INT information injection.The function of sending and receiving detection packets and the controller program are also implemented.Compared with the existing gray fault detection and localization schemes,the detection and location technology proposed in this thesis can reduce the bandwidth overhead by two exponential levels in one detection cycle.In addition,the experimental results show that the detection and location technology proposed in this thesis has higher fault localization accuracy under different workloads of background traffic.（2）We analyze the deficiencies of existing network simulation platforms to verify the telemetry solutions in the scenarios of supporting large-scale network topology distributed deployment and interconnection and intercommunication between P4-based custom network nodes.Based on this,we design and implement a network simulation platform for P4 telemetry verification,which can contribute to verifying P4-based telemetry solutions.We adopts the idea of centralized control and distributed service to divide the servers that carry the network simulation into master nodes and working nodes;the master node itself does not create a virtual network,but is only responsible for managing the working nodes,and provides interface functions such as topology deployment for users;the working nodes are based on the sub-topology issued by the master node.information to create the corresponding virtual network.We call the topology deployment API of the designed platform through the JSON of the large-scale topology,which proves that the platform can well support the distributed deployment of large-scale networks.We illustrate the platform’s ability to support telemetry scheme validation experiments in a network containing P4 custom network nodes by deploying a topology containing BMv2 switches. |
