Machine-Learned Classifiers for Protocol Selection on a Shared Network.

Knowledge about the state of a data network can be used to achieve high performance. For example, knowledge about the protocols in use by background traffic might influence which protocol to choose for a new foreground data transfer. Unfortunately, global knowledge can be difficult to obtain in a dynamic distributed system like a wide-area network (WAN). Therefore, we introduce and evaluate a machine-learning (ML) approach to network performance, called optimization through protocol selection (OPS). Using local round-trip time (RTT) time-series data, a classifier predicts the mix of protocols in current use. Then, a decision process selects the best protocol to use for the new foreground transfer, so as to maximize throughput while maintaining fairness. We show that a protocol oracle would choose TCP-BBR for the new foreground traffic if TCP-BBR is already in use in the background, for proper throughput. Similarly, the protocol oracle would choose TCP-CUBIC for the new foreground traffic if only TCP-CUBIC is in use in the background, for fairness. Empirically, our k-nearest-neighbour (K-NN) classifier, utilizing dynamic time warping (DTW) measure, results in a protocol decision accuracy of 0.80 for k = 1. The OPS approach's throughput is 4 times higher than that achieved with a suboptimal protocol choice. Furthermore, the OPS approach has a Jain fairness index of 0.96 to 0.99, as compared to a Jain fairness of 0.60 to 0.62, if a suboptimal protocol is selected.