| In this dissertation, we investigate the feasibility, and evaluate the performance, of applying speculative parallelization to the design of protocols that improves network performance over lossy links. Networks with lossy links, such as radio frequency (RF) wireless links, have inherently different characteristics from the traditional wired networks, for which the transport control protocol/Internet protocol (TCP/IP) design assumptions were made. It has been found that this may degrade network performance due to incorrect interpretations of a link corruption as a congestion loss. Although efforts have been put in to develop enhancements, to date, there has been no compelling evidence that any proposals are suitable for field deployment.; We establish a set of basic properties that an enhancement should have. Based on those properties, we propose a cross-layer speculative congestion control system architecture to improve the network performance over lossy links. This system consists of functional blocks at the middleware, transport, and network layers. The congestion control algorithm manager at the middleware layer monitors the global status of the network, and decides what loss treatment algorithm at the transport layer should be selected and executed. At the transport layer, we propose a protocol-level speculation algorithm consisting of a conditional Bernoulli predictor to predict the type of a loss event, and a speculative congestion control algorithm to appropriately adjust the window size for the prediction result. To maximize the speculation accuracy, we develop mathematical models to minimize congestion losses at explicit congestion notification (ECN)-capable random early drop (RED) gateways, by optimally dimensioning the maximum buffer size and the maximum threshold. This is critical to the conditional Bernoulli predictor at the transport layer.; Our simulation verification results have shown that the conditional Bernoulli predictor needs less memory space but produces higher prediction accuracy than several other previous network predictors. Our mathematical models can effectively optimize RED gateway parameters to minimize congestion losses, so that speculation accuracy for the conditional Bernoulli predictor is maximized. The proposed speculative system outperforms several previous TCP improvement algorithms as well as the baseline TCP-Reno. |
