OL_LaEC: An Online Learning Data Scheduler for Multi-site Parallel Downloading

The Multi-site Parallel Downloading (MPD) system employs multipath transport technology to offload traffic from expensive Content Delivery Networks (CDNs) to more cost-effective edge nodes, with the objectives of reducing traffic costs and improving video transmission quality. However, state-of-the-art multipath schedulers suffer from insufficient transmission stability and exhibit limited adaptability to network fluctuations. In this paper, we propose and implement OL_LaEC (Online Learning Loss-aware Earliest Completion), a lightweight online learning algorithm that models packet scheduling as a decision-making process, which is based on the MPQUIC protocol. We have meticulously crafted a reward function tailored to short video scenarios to select the optimal strategy for path scheduling. We implement a MPD system based on MPQUIC and conduct comparisons of OL_LaEC with state-of-the-art algorithms within simulated network environments. Our results demonstrate that the OL_LaEC algorithm achieves a more than 10% improvement in download speed compared to baseline algorithms.