Verification Of Intelligent Cache Management Mechanism Of High-speed Switch Platform Design And Implementation

Due to the rapid development of networks in recent years,it is difficult for traditional networks to cope with upper-layer applications.In order to optimize the traditional network,improving the performance of infrastructure is particularly critical.At the same time,machine learning has developed rapidly and has achieved many good results.Many algorithm researchers have begun to try to use machine learning to improve the performance of network equipment,such as queue scheduling,cache management,and congestion control.The survey shows that there is a lot of room for improvement in the allocation mechanism of shared-memory switch.If machine learning can be applied to shared cache allocation,it will definitely bring efficiency improvements.However,the current simulation environment is difficult to reflect the changes of the real network and cannot easily access the machine learning model.There is a problem that it is difficult for algorithm researchers to train and test the performance of algorithm models.In this context,algorithm researchers need an efficient machine learning training and evaluation verification platform.This thesis focuses on solving the problem that algorithm researchers are difficult to train and test the algorithm model.Based on the cache management in the shared cache switch,this thesis uses DPDK(Data Plane Development Kit)technology to develop a verification platform for the switch intelligent cache management mechanism.The functional modules of this system include the implementation of the switch data plane,the generation of custom test traffic,the baseline reference scheme,model training,model decision communication,model performance visualization and model training data storage.The purpose is to provide a training and evaluating environment for algorithm researchers,who can complete training and evaluating by configuring and running scripts,reducing the additional burden on algorithm researchers.Compared with the simulation system implemented by NS-3(Network Simulator 3),this system is based on real network traffic and is superior to the simulation system implemented by NS-3 in terms of efficiency and data authenticity.This thesis conducts a detailed test on the system,and the results show that the efficiency of training the model on this system is about ten times that of the simulation system implemented on NS-3.This system can facilitate the training and evaluating of algorithm researchers through configurable experimental scripts and visual scripts,and solves the problem that machine learning is difficult to deploy to shared cache allocation.