An Online Learning Multi-path Selection Framework for Multi-path Transmission Protocols

In the last decade, we have witnessed a tremendous growth of inter-connectivity among hosts in networks. Many new data transmission protocols have been developed to enable multi-path data transmissions between two hosts. However, the existing multi-path transmission protocol designs are limited as they neglect the stochastic nature of the metrics of the paths, e.g., latency, available bandwidth, and packet loss. Moreover, there are different design requirements in the applications, such as low latency, bandwidth throttling, and low loss rate in data delivery. In this paper, we propose a flexible online learning multi-path selection (OLMPS) framework to select multiple paths by learning the stochastic metrics of the paths and meeting the design requirements of the applications. Specifically, we design a set of novel online learning algorithms in the OLMPS framework for three different applications, maxRTT constrained, bandwidth constrained, and loss rate constrained, multi-path selection, to select paths and satisfy the requirements. We prove that the algorithms can provide theoretical guarantees on both sublinear regret and sublinear violation in our OLMPS framework.