CaR-PLive: Cloud-assisted reinforcement learning based P2P live video streaming: a hybrid approach

摘要

In recent years, live video streaming has become one of the most popular and prevalent applications of the Internet. The Peer-to-Peer (P2P) and Content Delivery Network (CDN) are popular approaches to stream video contents. These approaches respectively have faced some drastic challenges such as obtaining the desired Quality of Service (QoS) level and minimizing economic cost. The cloud computing infrastructures can reveal proper solutions to these problems. The P2P systems can eliminate their bandwidth shortage by renting resources from the cloud environment. This paper depicts CaR-PLive as a hybrid cloud-assisted P2P live streaming system. CaR-PLive uses video servers such as Amazon EC2 from cloud to stream video contents and rents Cloud Storage Services (CSSs) such as Amazon S3 to assist P2P live streaming system to reach the desired playback continuity. In CaR-PLive, we proposed two stages (sub-windows) sliding window for buffer management that a sub-window belongs to the P2P system and another one belongs to CSS. The objective of CAR-PLive is to optimize the size of sub-windows to minimize the overall rental cost of CSS restricted to a desired QoS level. We formulate this problem as an optimization problem and model it with Markov Decision Process (MDP) and then propose a reinforcement learning based algorithm to solve this problem. Finally, we evaluate the performance of CaR-PLive by performing extensive simulations and experiments with realistic settings. Simulation results demonstrate that CaR-PLive efficiently mitigates overall CSS billing cost in different system configurations and provides desired playback continuity in different system settings.