Research On Optimization And Analysis Of Network Congestion And Unfairness In Datacenter Network

With the increasing demand of the applications, the data center network loadbecomes heavier. This leads to the frequent occurrence of congestion, and it is easyto form a throughput collapse phenomenon called incast. Traffic characteristics ofdata center applications will result in an inequity phenonmenon called outcast.Through in-depth understanding several key parameters of the incastphenomenon, we find thatRTO minvalue plays a crucial role. With the increasingnumber of the sender, the degree of network congestion increase greatly. This willlead to packets loss and timeouts. Serious congestion makes it impossible for thesenders to receive three redundant ACK. This means that senders cannot enter fastrecovery but to wait for the timeout, which affects the network throughput. Amongthe incast solving algorithms, the transport layer algorithm is one of the favorablealgorithms. This is because such algorithms need little changes on infrastructure andeasy to implement. Therefore, in this paper, by designing an ACK reply changingrate based transport layer protocol, we solve the incast problem. The protocoladjusts the current congestion window by calculating the changing rate of ACK andthe estimation of the theoretical maximum congestion window.There is a kind of unfairness phenonmenon in the datacenter network. It issame with what was found by P. Prakash. This phenomenon shows that thethroughput of a flow with smaller RTT is less than the throughput of a flow withlarge RTT. This is completely on the contrary to the classic TCP algorithm. Afterverifying the widely existence of the outcast, the author gives an explanation basedon port blacking theory. However, after our carefully analysis, we give a substantialcause of the outcast phenomenon. That is because the unfair distribution of the flowswith different RTT on the physical link and the differences on congestion windowsize when finishing the current block transmission. According to the reason, Webuild an quantitative model. In the last part of this paper, we design a windownotification based protocol to address the the outcast phenomenon. By using themean value of the congestion window measurement, we unified the congestionwindow size of flows with different RTT, which improves the throughput of theflows with small RTT.By designing three essential and reasonable experiments on the ns-2simulator,we prove that the explanation of outcast phenomenon is correct. Furthermore, weverify the effectiveness of the two solving algorithms.