Research On Flow Scheduling In Data Center Based On Multiple-Level Feedback Queue

With the continuous development of cloud computing and artificial intelligence technologies,data center networks play an important role in providing services and support for various types of applications.Compared with Wide Area Networks,data center networks have higher requirements for flows of different applications.For example,the flows represented by web search services are mostly latency-sensitive small flows,while the flows represented by data replication are mostly throughput-oriented large flows.At present,when the flow information is not known a priori,how to achieve the network performance with low latency,high bandwidth and burst tolerance is one of the difficulties in the research of data center networks.The main technology is to employ the multiple-level feedback queue to simulate the shortest job first scheduling algorithm,allowing small flows to be scheduled first,while ensuring the performance requirements of large flows.However,how to set and adjust the priority thresholds of multiple-level feedback queue and achieve the degradation threshold to match the traffic distribution in the network is a very challenging work.This thesis mainly studies how to solve the traffic scheduling problem when the threshold of multiple-level feedback queues does not match with the traffic distribution in the network.The specific work is as follows:(1)For the mismatch between global traffic distribution and threshold parameters,this paper deploys a priority adjustment algorithm in the switch to detect large flows in the queues and reduce their priority,thereby reducing the queuing delay of small flows.First,the switch calculates the average value of bytes in each priority queue of the finished flows,and uses this average value as the standard for the number of bytes that the flow can forward in each priority queue.Second,the number of bytes forwarded on the queue determines whether the priority matches,and adjusts if they do not match.Finally,simulation experiments verify the effectiveness of this algorithm in mitigating the mismatch between global traffic distribution and threshold parameters.(2)For the mismatch between link traffic distribution and threshold parameters,this paper designs an end-to-end link-level priority adjustment scheme.Mainly by deploying a set of multiple-level feedback queues on the sending host and the receiving host,the multiple-level feedback queue at the sending end adopts global traffic distribution to set the threshold to ensure that small flows have a higher priority overall.This paper leverages a clustering algorithm to design a threshold update algorithm,and use the algorithm to calculate the threshold for the multiple-level feedback queue deployed on the receiving host according to the link traffic distribution.The multiple-level feedback queue can detect flows whose priority does not match the link traffic distribution,and dynamically adjust the priority.