Network Health and e-Science in Public Clouds

Commercial cloud providers are increasingly offering high performance and GPU-enabled resources capable of facilitating e-Science applications. However, the limitations of a public cloud´s internal network performance are well documented and can lead to the decision to use dedicated infrastructure over cloud resources for scientific applications. This paper explores the potential for improvement in the performance of e-Science applications on public clouds through the examination of the network in more detail. We introduce health indicators and evaluate various tomographic techniques for their ability to infer information regarding the network connection between instances. We also propose and formulate a set of health markers and health metrics to efficiently assess the network over time in order to make informed deployment decisions. Finally, we evaluate our work through a real-world medical image reconstruction application.