A Framework for Performing Statistical Testing of Distributed Systems

A major obstacle facing developers of large-scale distributed applications is deployment of the software under real-world conditions. While the product may hold up under lab conditions, it can fail horribly when faced with actual users using the service over the Internet. Part of the problem lies in the fact that there is very little support for software developers to examine the true behavioral nature of their software at scale. In order to determine this nature, it is important to be able reproduce a large number of different, realistic environmental conditions repeatedly and consistently. Without this ability, it becomes very difficult to argue that the observed behavior is truly representative of the system. Put in a different perspective, while characteristics like average throughput, response delay, and time between failures can be computed based on the lab results, it is unclear whether these measurements do indeed capture the true statistical properties of the system, as opposed to a momentary, rare blip. This work presents a prototype implementation of a framework that allows automated testing, including repeatable experiments (as required by the tenets of the scientific method), sweeping through experiment parameters, and supporting the ability to engage in closed-loop testing. The intention of this prototype effort is to build on top of existing testbed capabilities, such as those provided by Emulab. The results of using this prototype framework to test an industry-held distributed system, and to research collaborative worm propagation mitigation strategies, are shown.