On the effect of large-scale deployment of parallel downloading

Abstract

In the presence of multiple mirror servers that can serve the same information, clients can improve their performance when downloading large files by concurrently retrieving different parts of the file from different servers. In previous work, experiments with a single client have shown the potential for parallel downloading schemes to improve the performance perceived by this single client. We consider the question of what happens when parallel downloading is widely adopted within the Internet. To that end we start with an experimental evaluation of various parallel downloading schemes and their performance from the perspective of a single client. We use these results as a benchmark for a set of simulations in which we investigate the performance of parallel downloading when performed by multiple clients. We find that, because parallel downloading incurs additional overhead as a result of its use of more network connections, large-scale use of this technique can actually lead to overall degradation of the performance experienced by clients. Further, we uncover a "fairness" tradeoff in that any performance improvement experienced by clients performing parallel downloading comes at the expense of clients who simply go to a single server to retrieve files.