| Abstract |
As network-aware applications such as large data analytics conduct application-layer traffic adaptation, they can benefit from better network visibility to better orchestrate their data flows. As a result, the ability to predict the available bandwidth for a set of flows has become a key component for designing adaptive, distributed data analytics. In this short paper, we present a novel, on-demand network abstraction service that provides this ability. Handling not only non-reactive flows but also reactive flows such as flows managed by TCP congestion control algorithms, we introduce a set of novel techniques, including equivalent linear transformations, and model-driven throughput prediction with fast factor learning to address a set of challenges including privacy, scalability, convergence and unknown system parameters. We design a system called Prophet and leverage the emerging technologies of software defined networking (SDN) to realize it. Evaluations demonstrate that our system achieves significant accuracy in a wide range of settings. |