Design And Implementation Of Congestion Control Algorithm For Data Center Network Based On Online Learning

With the explosive development of mobile Internet technology,large-scale traffic aggregation centers have become the source of various network services.The data center network serves as a convergence point and a switching center for network traffic,providing services for various network applications.At the same time,due to the rapid development of related technologies such as cloud computing and artificial intelligence,the data center plays an increasingly important role as a powerful parallel computing and distributed storage resource in Internet.The performance of the data center network has become one of the bottlenecks in network development,which greatly restricts the development of high-quality network services and limits the computing and storage resources of artificial intelligence technology.Among them,the control algorithm is the key to improve the performance of the data center network.Therefore,the research of data center network congestion control algorithm based on TCP protocol has also become one of the hot issues in data center network research.Congestion control architecture for datacenter networks suffers from bad performance.In order to address the issue,there are a surge of approaches raised to congestion control in both academia and industry.However,that past approaches still prone to Non-convergent send rate and unfair bandwidth allocation in bursty and incast traffic patterns.This paper first proposes a congestion feedback mechanism based on active detection,which can detect the transmission status of the network in advance and plan the transmission flow of the bottleneck link in advance,so as to avoid network congestion.That is to say,the mechanism can detect the transmission capacity of the network in advance,arrange the appropriate data flow to avoid the transmission of network congestion in advance.This mechanism is mainly to accurately calculate the amount of data packets sent by detecting packets,so as to avoid the congestion of data packets as much as possible and provide a very low packet loss rate and transmission delay for the network.Although this method will sacrifice part of the bandwidth,but the subsequent experiments show that this method can ultimately improve the performance of congestion control algorithm.Then,we proposed a novel adaptive learning congestion control,named ALCC,based on online learning theory and credit packet feedback mechanism to achieve higher utilization of bandwidth,fast convergence and zero data loss for datacenter.It's noting that ALCC does not have to satisfy certain assumption,and can automatically learn the parameters from actual datacenter networks.ALCC requires only sender-side changes and is friendliness towards traditional DCTCP and AIMD.Extensive experiments show our algorithm higher utilization of bandwidth,more stable convergence and lower data loss.