PRR Is Not Enough

We study the effects of wireless channel burstiness on TCP. We measure TCP throughput over single-hop link traces from MIT’s 802.11b Roofnet and Intel Berkeley’s 802.15.4 Mirage testbeds. We observe that links with the same packet reception ratio (PRR) have throughput variations of up to 320%. We find that differences in throughput are accompanied by differences in link burstiness. Using the GilbertElliott model, we compute the parameter μ as a measure of burstiness. We show that for sufficiently large data traces, a simple second-order fit over both PRR and μ lowers the estimation error of TCP throughput by 50%. The estimation error reduces by 60%-99% for a more general fit over PRR and μ. We find that while μ has good deductive quality, the corresponding Gilbert-Elliott model is not accurate for simulation: TCP throughput values from empirical links and their corresponding simulated links based on the Gilbert-Elliott model can differ by up to 80%. These results help develop a better understanding of the underlying causes for the wide variations seen wireless network performance.