Dummy Probe Re-Transmission Algorithm for Available Bandwidth
Several research efforts have recently focused on the burstiness of Internet traffic in short, typically sub-second, time scales. Some traces reveal a rich correlation structure in those scales, while others indicate uncorrelated and almost exponential interarrivals . What makes the Internet traffic bursty in some links and much smoother in others? The answer is probably long and complicated, as burstiness in short scales can be caused by a number of different application, transport, and network mechanisms. In this note, we contribute to the answer of the previous question by identifying one generating factor for traffic burstiness in short scales: high-capacity flows. Such flows are able to inject large amounts of data to the network at a high rate. To identify high-capacity flows in a network trace, we have designed a passive capacity estimation methodology based on packet pairs sent by TCP flows. The methodology has been validated with active capacity measurements, and it can estimate the pre-trace capacity of a flow for about 80% of the TCP bytes in the traces we analyzed. Applying this methodology to Internet traces reveals that, if a trace includes a significant amount of traffic from highcapacity flows, then the trace exhibits strong correlations and burstiness in short time scales.