On bursty packet loss model for TCP performance analysis

In this paper, we study the timeout probability of TCP Reno under the bursty packet loss model, which is widely used to represent the loss characteristics of TCP under drop-tail FIFO queues. With a detailed analysis on the three timeout reasons for TCP Reno, we show that the impact of timeout has been underestimated in the existing literature. Surprisingly, we find that this more precise representation of timeout probability does not match the actual performance of TCP under drop-tail FIFO queues. Therefore we conclude that the bursty loss model is incapable of capturing the behavior of drop-tail FIFO queues, and using bursty loss model to analyze TCP performance is flawed. © 2005 IEEE.published_or_final_versio