Research And Implementation Of Data Center Network Traffic Scheduling Mechanism Based On P4

With the rapid development of Internet services such as big data and cloud computing,the scale of data centers is getting larger and larger,and the requirements of service flows are gradually changing from low latency to deterministic latency,and data center networks are evolving towards "deterministic and predictable".However,the current research on deterministic data center networks mainly focuses on the network side,i.e.,host side or central controller,and there is little research on scheduling flows directly inside the network.In addition,there are many burst flows and sporadic flows in data center networks,and the TCP Incast phenomenon is serious,so there is an urgent need for a scheduling mechanism that can adjust the sending rate of flows deployed inside the network.In view of this,this paper designs and implements a P4-based traffic scheduling mechanism for data center networks,which can provide deterministic end-to-end delay upper bound by deploying an asynchronous traffic shaping mechanism on each hop switch inside the data center network to perform flow-by-flow shaping of data flows,avoiding the accumulation of bursts.In addition,for the phenomenon that low-priority traffic is contested by high-priority traffic leading to exponential growth in queuing delay this paper designs a priority-aware load balancing strategy that periodically monitors the network state and other conditions through in-band network telemetry and performs different path selection strategies for flows according to their high and low priority,effectively reducing the backlog of data flows and decreasing the average flow completion time.The specific work of this paper is as follows:First,this paper designs a P4-based traffic scheduling mechanism for data center networks,which is functionally divided into two parts: traffic shaping and load balancing,and structurally divided into two parts: data plane and control plane.The control plane mainly implements the load balancing function,which is accomplished by the telemetry data processing module,the load balancing module and the flow table distribution module in collaboration.The load balancing module includes state identification and path calculation function units,which are responsible for calculating the global average priority and queue occupancy rate based on the telemetry data provided by the telemetry data processing module,and executing different path selection algorithms for the new streams entering the network according to the priority,and the flow table distribution module finally distributes the path calculation results to the data plane through the flow table;the data plane mainly implements traffic shaping and telemetry data collection functions.The data plane mainly implements the traffic shaping and telemetry data collection functions,which are completed by the identification and filtering module,shaping module and output module,and telemetry data collection module in cooperation.After that,this paper implements the above design scheme,in which ONOS is selected as the controller in the control plane and BMv2 software switch in the data plane,focusing on the processing of telemetry data in the control plane and the implementation of algorithms for path selection of flows using the collected network information in the load balancing module;the packet initial definition and packet parsing in the data plane based on P4,the identification filtering module’s matching action table in the data plane,the shaping module using the interleaving shaping algorithm to allocate qualified time for data frames,and the implementation of the outgoing transmission selection algorithm in the output module.Finally,a prototype system is built for functional testing and performance verification,and the implementation of two functional modules,traffic shaping and load balancing,is tested,and the performance of the mechanism designed in this paper is verified around four indicators: flow completion time,effective throughput,queue length and packet forwarding delay.The test results show that the traffic scheduling mechanism proposed in this paper achieves better performance in terms of delay and throughput when the network is heavily loaded,and reduces the long-tail delay of the network,providing a "deterministic and predictable" performance guarantee for the network.