| One-way delay trend detection has been used to estimate end-to-end available bandwidth in recent research. This dissertation analyzes various error sources that may distort the measured delay trend and propose corresponding counter measurements. Instead of trying to find out the absolute available bandwidth, we focus on testing whether a path has available bandwidth at a specified target value. The result of this test can help a rate-based congestion control algorithm to adjust the sending rate. A new delay trend model is proposed to accommodate variable packet sizes used in a probe. We investigate the effect of trend detection algorithm, packet size, packet size distribution, clock measurement error, probe size, and context switch etc. on the performance of available bandwidth estimation through extensive simulations and Internet experiments. Based on the delay trend model, we propose a bandwidth inference congestion (BIC) control protocol for multimedia applications using a cumulative layered multicast transmission. BIC uses one-way delay trend detection technique to infer the path's available bandwidth. BIC is purely rate-based, which does not use the round trip time (RTT) information. A BIC receiver only conducts a layer-join when the end-to-end available bandwidth can support the cumulative rate of the next layer. The BIC sender uses a multi-layer token bucket model, which accommodates variable-size packets, to control the sending rate for each layer. Each receiver decides to join or leave a layer independently. BIC is very simple since the sender and all receivers are independent without interaction. BIC has good stability, especially when the available bandwidth is smooth, since no join experiment is used. BIC is also highly scalable to a large number of heterogeneous receivers with different bandwidths and delays. Being RTT-independent, BIC sessions have different fairness characteristics compared with TCP and other TCP-like protocols, which are RTT-biased. We evaluate BIC for a large variety of scenarios, especially the fairness issues in details, by simulations and Internet experiments. |
