Research On Congestion Control Scheme For Certain Applications

Powerful computing and massive training data drive the rapid development of technologies based on machine learning for image classification,speech recognition,and driverless driving.As machine learning models continue to grow,the need for storage and computing needs to be addressed through distributed machine learning systems.The data-parallelism commonly used in large-scale distributed machine learning will generate a typical many-toone traffic.The rapidly increasing data which need to be synchronized require higher bandwidth of the network.Network communication has become an important bottleneck for machine learning applications in distributed systems.The traditional congestion control strategy does not consider the communication mode and traffic distribution characteristics of distributed machine learning applications.The coarse-grained control mechanism makes the network unable to respond flexibly to network,resulting in network congestion and affecting the training of distributed machine learning.Based on the research ideas of current congestion control strategies,this thesis analyzes the related strategies in detail from two aspects: single-path transmission and multi-path transmission.Based on these two ideas,this thesis proposes corresponding solutions to solve the flow completion time optimization problem based on single-path transmission and Transmission Control Protocol(TCP)incast problem based on multi-path transmission.Current control strategy has problems,like the granularity of current control strategy is coarse,the predictability of subsequent traffic is poor,and the convergence speed is slow.This thesis designs a delay-quantized congestion control strategy.It is based on single-path transmission.By quantifying the predictable queuing delay and adding the trend factor,the sender can immediately obtain fine-grained link information and precisely control the send window with the adjustment of the transmission rate.The status information uses a fast feedback mechanism to generate custom data packets and send them back to the source,reducing the retention time.The simulation results show that the strategy can effectively improve network throughput by 20% under certain conditions,reduce the average stream completion time by 50%,and reduce the tail latency of mice flows.In order to solve the disadvantage of multi-path transmission mechanism on TCP incast problem while preserving the advantages of multi-path transmission network,this thesis designs a sub-flow adaptive congestion control strategy based on multi-path transmission.Through fine-grained sense of congestion information and sub-flow number adaptation mechanism,the scheme can dynamically adjust the number of available sub-flows and select lighter congestion paths for transmission according to the sub-flow path congestion status.This scheme can enhance the network's tolerance for TCP incast problems without reducing network utilization.The simulation results show that the scheme can effectively solve the TCP incast problem under multipath transmission,and the performance is equivalent to the congestion control strategy based on single path transmission.Under certain conditions,the scheme has higher network utilization and lower tail latency.