Research On Congestion Control Strategies In Data Center Networks

Driven by the demand for data communication,the network between data centers has become more and more important,and congestion control has also become an important issue in the network.To maintain the stability of the data transmission process,researchers have done a series of research on congestion control.Traditional TCP congestion control is typically based on packet loss and delay-based congestion control.With the development of network cards and memory chips,many applications have high throughput and delay constraints,and it is no longer possible to simply use packet loss and delay as a signal of congestion.Finding the right balance between utilizing capacity efficiently and keeping latency low is therefore an important issue in congestion control.In 2016,Google proposed a rate-based congestion control algorithm named BBR,which is known as the creation of a new era of congestion control algorithms.Create a network path model by measuring the available bottleneck bandwidth and estimating the bandwidth-delay product(BDP)of network bottleneck links with the minimum round-trip time,aiming to keep buffer utilization low while maximizing throughput.According to the implementation of BBR in the actual network,BBR has better advantages in terms of throughput than the traditional congestion control algorithm in TCP,and can significantly improve the utilization of bandwidth.With the research on the BBR algorithm,it is found that in the case of network fluctuations,BBR is used in the bandwidth detection stage to select the maximum sample value during the detection process.In network fluctuations,the maximum sample value is often higher than the actual bandwidth,resulting in overestimation,resulting in link Problems like overloading,high queuing delays,etc.According to the problem that BBR bandwidth is easy to overestimate,this paper proposes a new filter.In addition,aiming at the fact that BBR cannot timely distinguish whether the packet loss is due to network congestion or random packet loss,which causes low channel utilization,a congestion signal identification mechanism based on RTT is proposed.(1)Since BBR uses a simple maximum filter(MAX filter),it leads to overestimation during detection.Firstly,through mathematical analysis and simulation experiments,the situation that the sending rate of the asynchronous detection mechanism of the BBR congestion control algorithm does not match the bottleneck bandwidth is analyzed and verified.Secondly,aiming at this kind of behavior,this paper regards the sudden change in bandwidth caused by network fluctuations as a step event and uses the Tobit Kalman filter to correct the characteristics of saturated data in the estimation process and the adaptive Kalman filter to track the change of channel capacity and process the estimation process.The characteristics of the unknown noise in the medium,combine the two,replace the BBR-ATKF algorithm of the MAX filter,and model the capacity as a Gauss-Markov process.At the same time,it is proposed to regard the bandwidth mutation caused by network fluctuations as a step event,and dynamically adjust the mean value of the filter in real-time to respond to network mutations or even network interruptions promptly.Therefore,the obtained channel capacity estimation value is closer to the actual value,and the huge potential fluctuation problem of RTT caused by BBR overestimation is improved.(2)In terms of congestion signal identification,BBR cannot timely distinguish whether packet loss is due to network congestion or random packet loss,and cannot detect congestion in a targeted manner,resulting in low utilization of network links,packet loss,and retransmission in the bottleneck buffer and queuing delays.This paper constructs a congestion queue model based on the dynamic characteristics of the congestion window,deduces the boundary of the model,and obtains the maximum RTT boundary value without triggering congestion as a congestion signal,to respond to congestion more quickly and make full use of the channel resources while reducing buffer queuing delays.Finally,through the test and analysis of simulation software,it is concluded that compared with the standard BBR,the proposed algorithm reduces the retransmission by about half,the packet loss rate is reduced from 32% to 3.3%,and the average RTT is reduced by 40%,which has better transmission performance.