Delay characterization and performance control of wide-area networks

In packet-switching networks, delay characterization is important to protocol designs, real-time applications and network monitoring and diagnoses. The dynamics of end-to-end delays in wired networks is dominantly determined by queueing delays. Since modern data networks are no longer Markovian queueing networks, conventional queueing analyses and conclusions need a review. Developing new tools to characterize queueing behavior when facing complicated traffic models and new transport control mechanisms is very meaningful to both theory and practice. The work presented in this thesis is one effort in this direction. By characterizing network delays as a multi-structure process, we develop a generic modeling framework for delay phenomena dominated by queueing process. To clearly identify and quantitatively characterize the multi-structure of a delay sequence, a technique called delay jitter deviation-lag function (DLF) analysis is introduced. The information that can be extracted from a path's DLF includes maximum delay jitter deviation, load intensity, traffic burstiness, buffer drainage time, average busy period, and network stability. Relying on the new delay characterization method, we find a prevailing anti-persistence phenomenon in Internet delay sequences, suggesting that the current Internet operates in a stable state. As an application of the DLF analysis, we design a delay boundary prediction algorithm. Preliminary experiments show it can out-perform ARMA-model-based Jacobson's algorithm by significantly improving the prediction error rate. Inspired by the improved understanding on network dynamics revealed by our method, this thesis also proposes a new networking architecture which constitutes a transport control protocol for the end systems and a queue management mechanism for the network gateways. The proposed transport control protocol decouples the flow control and congestion control, and further decouples the sending rate control and the sending burstiness control of the congestion control. The proposed queue management deploys a two-dimensional model to characterize the queue's dynamic behavior. The two dimensions are: queue utilization factor and bursty factor defined in this thesis. Our simulations show that this architecture can conduct micro-scale network performance control with great flexibility to reduce delay jitters.