Rethinking TCP throughput and latency modeling

TCP throughput and latency models are useful tools to characterize the TCP performance. The canonical throughput model [2], while useful, has some limitations since it does not consider how packet loss rate changes over time. This approach leads to poor predictions for short flows. We present a new modeling approach that characterizes the throughput and latency models by: (i) discovering the relationship between the packet loss rate and the congestion window size, and (ii) incorporating the starting congestion window and the number of parallel connections. Experimental results show that our models significantly improve modeling accuracy.