Design And Implementation Of Switch Shared Buffer Management Mechanism And Simulation System

Nowadays,data center switches usually use on-chip shared memory.A reasonable buffer management strategy is of great significance to ensure fair buffer sharing between ports and reduce packet loss.In this thesis,we first study the switch shared buffer management.It is found that the traditional buffer management strategy can lead to unfair buffer allocation between ports in some scenarios,and even lead to unnecessary packet loss of some burst streams.This is because the traditional buffer management strategy only uses the same buffer management standard for all ports according to the aggregation information of the current buffer(for example,Dynamic Threshold(DT)strategy only uses the buffer management according to the idle buffer size),and ignores the differences and interactions between ports with different traffic patterns.In order to solve this problem,this thesis proposes a switch shared buffer management based on reinforcement learning(SBM-DRL),which uses neural network to learn buffer management strategy efficiently,and realizes switch shared buffer management adaptively in different network environments.In order to facilitate the algorithm researchers to train the model and evaluate the performance of various cache management strategies,this thesis designs and implements NS3 with shared buffer management(NS3-SBM)simulation platform based on NS3 network event simulator.According to the requirements of the subject,this thesis analyzes the requirements,defines the functional requirements and non functional requirements of the simulation platform;then carries out the outline design,gives the overall architecture of the simulation platform;on this basis,completes the detailed design and implementation of the platform.Finally,the simulation platform is fully tested.