Efficient buffering and scheduling for a single-chip crosspoint-queued switch
Proceedings of the eighth ACM/IEEE symposium on Architectures for networking and communications systems
| Hi-index | 0.00 |
The real-world impact of network losses on the performance of Transmission Control Protocol (TCP), the dominant transport protocol used for Internet data transfer, is not well understood. A detailed understanding of this impact and the efficiency of TCP in dealing with losses would prove useful for optimizing TCP design. Past work in this area is limited in its accuracy, depth of analysis, and scale. In this dissertation, we make three main contributions to address these issues: (i) design a methodology for in-depth and accurate passive analysis of TCP traces, (ii) systematically evaluate the impact of design parameters associated with TCP loss detection/recovery mechanisms on its performance, and (iii) systematically evaluate the ability of Delay Based Congestion Estimators (DBCEs) to predict losses and help avoid them.We develop a passive analysis tool, TCPdebug, that accurately tracks TCP sender state for many prominent OSes (Windows, Linux, Solaris, and FreeBSD/MacOS) and accurately classifies segments that appear out-of-sequence in a TCP trace. This tool has been extensively validated using controlled lab experiments as well as against real Internet connections. Its accuracy exceeds 99%, which is double the accuracy of current loss classification tools. Using TCPdebug, we analyze traces of more than 2.8 million Internet connections to study the efficiency of current TCP loss detection/recovery mechanisms. Using models to capture the impact of configuration of these mechanisms on the durations of TCP connections, we find that the recommended as well as widely implemented configurations for these mechanisms are fairly sub-optimal. Our analysis suggests that the durations of up to 40% of Internet connections can be reduced by more than 10% by reconfiguring prominent TCP stacks.Finally, we investigate the ability of several popular Delay Based Connection Estimators (DBCEs) to predict (and help avoid) losses using estimates of network queuing delay. We find that aggressive predictors work much better than conservative predictors. We also study the impact of connection characteristics—such as packet loss rate, flight size, and throughput—on the performance of a DBCE. We find that high-throughput connections benefit the most from any DBCE. This indicates that DBCEs hold significant promise for future high-speed networks.