GOMOS: A GNN-Assisted Network-Media Integrated Optimization Framework for Massive Mobile Live Streaming

In this paper, a network-media integrated optimization framework assisted by a graph neural network (GNN) named GOMOS is proposed, aiming to address the challenge of reducing overall live latency and avoiding rebuffering events in massive mobile live streaming (MMLS) systems. Compared to existing frameworks that optimize latency from either the media or network perspective, GOMOS profoundly considers the mutual impact between them, and jointly optimizes both wired and wireless networks with three steps. Firstly, a bi-directional link traversal algorithm, adopting a customized TCP traversal packet (TTP), is designed to collect and disseminate the real-time link status of wired networks. Secondly, using graph representation to comprehensively consider the characteristics of media and networks, a gated GNN (GGNN) model is utilized to predict the latency of transmitting MMLS flows with wired links. Thirdly, given the predicted wired transmission latency, the resource allocation of wireless networks is formulated as a cross-layer convex optimization problem, ensuring timely completion of a majority of MMLS tasks. Extensive evaluation results with NS-3 show that the overall live latency is kept within an acceptable range, and the probability of rebuffering events is significantly reduced compared to the benchmark schemes.