Research And Implementation Of Congestion Awareness And Marking Strategies With Micro-Burst In Data Center Environment

In recent years,with the rapid expansion of data center business scale,data center network traffic has evolved from traditional “North-South traffic” to "East-West traffic" inside the center,and the switches used in the data center are gradually changed from single-queue switches to multi-queue switches.The inevitable competition of network bandwidth resources in the data center makes the congestion problem of the data center network become more and more serious,such as link packet loss,network delay,and network jitter,and the data center network throughput is drastically reduced.In order to achieve high throughput and low latency of the network,the data center usually adopts the standard ECN(Explicit Congestion Notification)mechanism.However,with the continuous development of data center networks and the continuous evolution of related technologies,a new traffic pattern called micro-burst has had a huge impact on the performance of the ECN mechanism.The presence of micro-bursts can cause a transient overflow of the switch buffer.The switch instantaneous queue length momentarily exceeds the ECN marking threshold and will generate a large number of mismarkings.These spurious congestion signals will directly cause the sender to reduce the transmission rate,and the network throughput drops sharply.Therefore,it is necessary to improve the congestion awaring and marking with micro-burst traffics by improving the ECN mechanism,which is widely supported by data centers,and ensure high throughput of data center networks.In this paper,the research will be studied in the following three aspects.Firstly,for the problem that the micro-burst causes the switch buffer to overflow instantaneously,this paper adopts the steady-state analysis method to model and analyze the whole process of the micro-burst flow generation and transmissions.The analysis results in a proper ECN threshold lower bound that can absort the micro-burst.Further,in order to ensure the matching of the input rate and the output rate of the switch,based on the ideal GPS(Generalized Processor Sharing)model,the lower bound of the ECN threshold is adjusted to obtain an optimal initial value of the ECN threshold that can be adapted to the micro-burst traffic.The initial ECN threshold could make switches sense the arrival of the micro-burst and have enough room to absorb it.Secondly,in order to adapt to the dynamic data center network,the current queue length of the switch is used as the perceptual parameter,and the sum of all queue buffer utilization ratio is taken as the target variable to establish an adaptive ECN threshold dynamic adjustment optimization model based on the above-mentioned initial ECN threshold.The optimization goal is to maximize the throughput of the switch link.To be able to adapt to the micro-burst traffics,the ECN marking threshold will be adjusted after each round of queue scheduling.At the same time,combined with the double-marking threshold and the queue growth slope,an appropriate marking strategy is determined to distinguish whether it is an spurious congestion caused by the micro-burst.Finally,a congestion awareness and marking system for micro-bursts under the data center environment is designed and implemented.In order to further verify the effectiveness of the congestion awareness and marking mechanism proposed in this paper,the system is based on the Linux kernel network flow control module for secondary development,to realize the initial setting function and dynamic adjustment of the marking threshold for the micro-burst ECN mechanism.In addition,based on the multiple data center environment of Southeast University Cloud Computing Center,the system is deployed to verify the effectiveness of the research results.The experimental results in the actual environment of the Cloud Computing Center of Southeast University show that the congestion awareness and marking mechanism with microburst traffics under the data center environment can effectively reduce the generation of mismarkings and improve the flow completion time.Ultimately our scheme can achieve high throughput of the data center network without hurting latency.