Large-scale experiments on a cluster

Evaluation of large-scale network systems and applications is usually done in one of three ways: simulations, real deployment on Internet, or on an emulated network testbed such as a cluster. Simulations can study very large systems but often abstract out many practical details, whereas real world tests are often quite small, on the order of a few hundred nodes at most, but have very realistic conditions. Clusters and other dedicated testbeds offer a middle ground between the two: large systems with real application code. They also typically allow configuring the testbed to enable repeatable experiments. In this paper we explore how to run large BitTorrent experiments in a cluster setup. We have chosen BitTorrent because the source code is available and it has been a popular target for research. Our contribution is two-fold. First, we show how to tweak and configure the BitTorrent client to allow for a maximum number of clients to be run on a single machine, without running into any physical limits of the machine. Second, our results show that the behavior of BitTorrent can be very sensitive to the configuration and we re-visit some existing BitTorrent research and consider the implications of our findings on previously published results. As we show in this paper, BitTorrent can change its behavior in subtle ways which are sometimes ignored in published works.