Regenerating TCP Dynamics From Traces Path Characteristics

A deep knowledge of TCP performance on the Internet at large is intimately related to understanding the role of the underlying path characteristics (bottleneck rate, buffer sizes, end-to-end delay, loss rate) on the performance of dynamic TCP flows. The practical way to study the TCP performance is by simulating on a dumbbell topology, Internet statistically equivalent cross-traffic along with simulated TCP, then extracting interest variables such as throughput as the indicator of performance. However, we advocate that much more can be learned about TCP performance by regenerating path characteristic conditions derived from traces, and let TCP flows run loose on top of it. In this study, we investigate several Internet public TCP flows, using LAN-specific traces (Berkeley LBNL traces), and we show how path characteristics can be extracted. We devise a new technique to obtain path capacity by searching longitudinal patterns of packet pairs on each TCP flow present in a trace. In addition, we look on buffer usage patterns based on effective throughput/estimated capacity ratio, as a way to relate RTT variation, packet loss and disclosure buffer sizes. Finally, the natural application is to test advanced TCP proposals on a per-trace scenario basis by regenerating dynamic traffic patterns directly from stored traces, such method is also called source-based traffic pattern. In this trace regeneration testbed, we have as input the TCP trace, then, a post-processing mechanism inferring path characteristics per flow. Finally, emulated virtual machines running in parallel enable the TCP traffic dynamics to run loose.