Dragonfly: In-Flight CCA Identification

Dean Carmel, Isaac Keslassy

Abstract: We introduce the Dragonfly system, which is designed to classify on the fly the congestion control algorithm of any flow that crosses a given router, starting at any time, and quickly reach a reasonable accuracy. To do so, we discuss the unique challenges of real-time congestion control classification. We explain how the number of bytes of the flow within the shared router queue contains an intrinsic memory that significantly helps real-time classification. However, we show that this number of bytes is not straightforward to compute in real time, and introduce ways to do so. We further design an eBPF-based scalable traffic-collection system that helps dynamically filter specific flows at high rates. Finally, we evaluate our Dragonfly system using a variety of platforms, and show that it clearly outperforms state-of-the-art algorithms.