Characteristics of Network Delays in Wide Area File Transfers

In this paper we present an analysis of over 236,000 file transfers between 10 widely distributed Internet hosts. The goal of this work is to broaden the understanding of how network path and congestion properties contribute to delays in TCP file transfers. The first part of our analysis investigates how end-to-end path properties (eg. physical distance, router hops, autonomous system hops, or bottleneck bandwidth) relate to file transfer latency and variability. We evaluate end-to-end path properties as a predictor of file transfer latency, and employ dimensionality-reducing techniques to identify clustering in path space. We find that expected transfer latency can be effectively predicted by a number of path properties and that the relationship between paths and latency is strongly linear with some intense outliers. The second part of our analysis employs critical path techniques to break down the network component of file transfer latency into three categories: propagation, queuing and loss. We compare the contribution of each of these components to delays along particular paths, and their effect on variability of total delay. We find that propagation delay is the dominant aspect of expected total delay for most paths and that queuing and loss are substantial effects typically for a minority of paths. On these paths, queuing contributes most significantly to periodicity in total delays while loss contributes most significantly to variability in total delay.