| The recently emerging 360° video,cloud VR games,autonomous driving and other emerging Internet applications usually require data delivery before the deadline.However,the existing transport layer is too primitive to accomplish that.So that the Deadline-aware Transport Protocol(DTP)was proposed in 2019.The Deadline-aware Transport Protocol(DTP)naturally supports the above functions.It is implemented by the DTP scheduler.The data blocks will be put into the sending buffer first,and the DTP scheduler will determine the transmission according to the deadline and priority of the block.In addition,the scheduler will adjust the sending rate according to the observed network states to maximize the use of network bandwidth.DTP scheduling algorithm mainly consists of two parts: packet selection and congestion control algorithm.This thesis focuses on the DTP congestion control algorithm.DTP congestion control is a mechanism by which DTP adjusts the sending rate on the sender to achieve the most efficient use of network bandwidth.An excellent congestion control algorithm needs to effectively detect network congestion in the presence of noise interference;it can adapt to the dynamic and changeable network environment by adjusting the congestion control strategy in time;it also needs to consider the issues of fairness in the competition of multiple transmission streams.This thesis mainly uses the Deep Reinforcement Learning framework to propose new congestion control algorithms or improve traditional congestion control algorithms.Currently,RL-based congestion control can be divided into three areas:(1)performanceoriented,(2)specific scenarios,and(3)fusion heuristic algorithms.PRLCC is an algorithm that uses DRL(Deep Reinforcement learning)combined with the Probe Bandwidth mechanism.The Agent trained with this algorithm can effectively utilize the network bandwidth,and it is easier to converge during training than other RL algorithms.Regas improves the delay-based congestion mechanism of Vegas and uses a DRL framework to judge congestion instead.Compared with Vegas,Regas utilizes more bandwidth,and more importantly,greatly improves the fairness of Cubic and Vegas. |
