Congestion Detection and Control in Data Center Networks

Abstract

As mobile devices become increasingly popular, traffic through data centers is increasing rapidly. Cisco has predicted that data center traffic will reach 7.7 Zettabytes in 2017, of which 76% will be generated by the internal exchange of data. Recent research has revealed a severe problem of congestion in data centers that is called TCP Incast. TCP Incast will significantly prolong the time of completion of jobs, degrading the performance of application services. The main reason for TCP Incast is many-to-one communication, which causes the switch buffer to overflow. Hence, in this paper, a method for detecting and controlling congestion of traffic at a data center is proposed. The goal is to detect and solve the problem of congestion. The system is tested in a real data center and demonstrated to detect TCP Incast accurately. Furthermore, application layer solutions for TCP Incast are implemented and the TCP Incast problem is interestingly found to continue to arise occasionally, because application layer solutions have only static environment parameters as their input parameters. The adaptation of the mechanism for detecting and controlling TCP Incast warrants future study.