Performance Analysis And Mechanism Enhancement Of Transport Protocols In Data Center Networks

The prosperous development of data center helps various services become reality.However, diverse services have di?erent requirements of network performance in the same network, which applies pressure on the data center network. There exists contradiction between some certain services and traditional network protocol. In many-to-one tra?c pattern, high synchronization of short flows usually makes the congestion in the network become severe at once, thus TCP cannot work normally, which leads to throughput collapse. This phenomenon is called TCP Incast problem. Because of its catastrophic performance, TCP Incast is widely researched both in academia and industry. Among so many solutions to ease TCP Incast problem, DCTCP is the presentative which can postpone the throughput collapse point compared to TCP. It also makes proper tradeo?between throughput and latency. However, DCTCP has terrible problem of self oscillation. The performance is heavily influenced by its unstable queue and server jitter. In this dissertation, we study the two critical problems and the solutions are as follows:(1) Comprehensive Understanding of TCP Incast Problem: We build up an interpretive model which considers all the possible system parameters and mechanism variables related to TCP Incast problem. We expect the qualitative explain which is adaptable to all the circumstances, not restricted to certain environment or chosen mechanism.Based on this model, we find that the window size distribution has close relationship with the throughput, and di?erent mechanisms, algorithms or protocols may have di?erent influences on the window size distribution. The results from simulations and experiments validate that our theoretical analysis can predict and explain the phenomena in the experiments properly, for example, how the parameter adjustments, enhanced mechanisms and new protocols influence TCP Incast problem. We also get some general conclusions from the model, mainly the number of flows at the throughput collapse point is proportional to the bu?er size, and the smaller the average window size and the standard deviation of window sizes, the later throughput will collapse. Besides, the theoretical analysis reveals that most solutions of TCP Incast help to mitigate the throughput collapse, but cannot solve it entirely.(2) Ease the Queue Oscillation: Analysis and Enhancement of DCTCP: Severe queue oscillation is found in the experiment of DCTCP. Based on theoretical analysis,we find that the essential reason of oscillation is the nonlinear component of DCTCP.The single marking threshold of DCTCP at the switch makes the congestion notice arrive too late at the sender side. Thus, we propose DT-DCTCP as the enhancement of DCTCP,which uses two thresholds to share the load of one threshold of DCTCP. It can transfer the congestion information earlier to the sender sides. With the help of describing function and stability criterion, we analyze the stability of DCTCP and DT-DCTCP, and prove the latter is more stable than the former theoretically. DT-DCTCP can maintain smaller queue length, and the queue length is less sensitive to the growing number of concurrent flows. At last, we validate through experiment that DT-DCTCP can postpone the Incast throughput collapse point compared to DCTCP, and it also reduces the tail latency in completion time experiment.