Research On Network Traffic Management Performance Optimization Technology For Datacenters

Modern datacenters provide infrastructure services for mass applications,such as web search and online recommendation systems,etc.As the common carrier for all application data transmission,the datacenter internal network needs to provide high bandwidth,low latency and other guarantee of quality of service(Qo S).Due to the increasingly strict re-quirement of applications on cloud service performance,the guarantee of Qo S in datacenter networks faces great challenges.Besides,the rapid growth of computing power of data-center servers and data access capabilities of storage devices has gradually led to datacenter networks becoming the performance bottleneck in data processing.Therefore,facing the ever-increasing application requirements and constantly updating hardware and software fa-cilities,it is of great significance to study efficient network traffic management technology in datacenters to effectively guarantee the service quality of datacenter applications.Many ex-isting research studies are conducted in network congestion control,traffic priority schedul-ing and network load balancing.However,current solutions still face many problems,such as high difficulty in deployment for high-performance solutions,difficulty in providing per-formance guarantee for application workloads with different traffic burst characteristics and the problem that load balancing interferes with host-side flow rate control,etc.Aiming at the problems in the complex and variable network environment of datacenter,this paper pro-poses two load balancing schemes and one congestion control scheme to effectively improve the transmission performance of the datacenter network.Aiming at the problem of inaccurate congestion detection in the host-based network load balancing schemes,this paper proposes a latency-based load balancing(LLB)scheme based on accurate delay feedback,which optimizes the load balancing decision to select the appropriate transmission path to improve network transmission performance while ensuring easy deployment.Host-based load balancing solutions generally have the advantage of easy deployment because they do not require modification for hardware switches,but the Explicit Congestion Notification(ECN)used by existing solutions and the one-way delay monitored at the application layer cannot accurately estimate the path congestion degree.They cause difficulty in selecting the optimal data transmission path when scheduling flows,which im-pairs transmission performance.LLB uses host-based hardware network card technology or efficient data IO processing technology residing in software layer to accurately capture the network one-way delay to estimate the path congestion degree,thereby improving the accuracy of congestion detection? LLB also actively updates delay metrics between nodes by using other normal flows in the network to accelerate the feedback of path congestion information.Experiments show that LLB can increase the transmission performance by up to 47% compared to the typical solution CLOVE-ECN based on the ECN congestion feed-back? and under relatively smooth application loads,LLB can still improve the transmission performance by 2-8% compared to the scheme CONGA that needs to modify the hardware switch.Aiming at the problem that network load balancing has difficulty in providing adaptive performance guarantee for application workloads with different burst intensities,this paper proposes a load balancing scheme Int Flow,which is adaptive to traffic bursts.By fully exploiting the data plane programmable switch feature,It integrates host-side congestion monitoring and switch-side traffic burst capture to adapt to burst intensity changes from upper-layer application traffic,thereby providing better data transmission performance.The main idea behind Int Flow is to monitor the congestion feedback carried by ACK at the host-side congestion control protocol layer to assess the congestion status of flow and write the status information to the packet header? and then to parse the congestion information in the packet header in the programmable switch to sense the congestion status of the current flow,which is used to combined with the captured traffic burst feature to make proactive and careful load balancing decisions.Experiments show that Int Flow can improve network transmission performance by up to 45% and 30%,respectively,compared to the current representative load balancing scheme CONGA and Hermes.Aiming at the problem of performance loss caused by network load balancing inter-fering with host-side flow rate control,this paper proposes a routing change aware con-gestion control scheme ACK-correction,which provides more accurate flow rate control to improve network transmission performance while ensuring ease of deployment.When the load balancing scheme frequently schedules flows,which causes ACKs reflecting different path congestion conditions to be mixed together,the ACK belonging the path before the flow switching may be used incorrectly to adjust the transmission rate(i.e.,the flow rate)after the flow switches path.This causes chaos in flow rate control and affects transmis-sion performance.ACK-correction records the transmission path number of flows by using the fixed field of the packet header based on the overlay network protocol,and uses ACKs for feedback? ACK-correction compares the path number carried by the ACK with the cur-rent transmission path number of the flow to determine whether the ACK is outdated at the sender,and passes the comparison result to the upper network protocol stack through a re-served bit in the TCP header? a new congestion control protocol MDCTCP is designed in the host virtual machine,which makes the suitable flow rate adjustment based on the con-gestion feedback signal and the above-mentioned reserved bit to avoid the above-mentioned chaos in flow rate control to optimize transmission performance.Experiments show that compared with the typical congestion control protocol DCTCP,while ensuring easy deploy-ment,ACK-correction improves the overall transmission performance of all flows by 15%?ACK-correction can also significantly reduce the average completion time and tail latency of delay-sensitive short flows,which can achieve 22% and 27% performance optimization respectively,greatly improving the network transmission performance for delay-sensitive applications? in addition,in the face of the dynamic network topology in the datacenter,ACK-correction also provides stable and high transmission performance.