With the advancement of rich media generating devices, the proliferation of live Content Providers (CP), and the availability of convenient internet access, crowdsourced live streaming services have witnessed unexpected growth. To ensure a better Quality of Experience (QoE), higher availability, and lower costs, large live streaming CPs are migrating their services to geo-distributed cloud infrastructure. However, because of the dynamics of live broadcasting and the wide geo-distribution of viewers and broadcasters, it is still challenging to satisfy all requests with reasonable resources. To overcome this challenge, we introduce in this paper a prediction driven approach that estimates the potential number of viewers near different cloud sites at the instant of broadcasting. This online and instant prediction of distributed popularity distinguishes our work from previous efforts that provision constant resources or alter their allocation as the popularity of the content changes. Based on the derived predictions, we formulate an Integer-Linear Program (ILP) to proactively and dynamically choose the right data center to allocate exact resources and serve potential viewers, while minimizing the perceived delays. As the optimization is not adequate for online serving, we propose a real-time approach based on Reinforcement Learning (RL), namely RL-OPRA, which adaptively learns to optimize the allocation and serving decisions by interacting with the network environment. Extensive simulation and comparison with the ILP have shown that our RL-based approach is able to present optimal results compared to heuristic-based approaches.

随着富媒体生成设备的进步，实时内容提供商（CP）的普及以及便利的Internet访问的可用性，众包的实时流媒体服务见证了意想不到的增长。为了确保更好的体验质量（QoE），更高的可用性和更低的成本，大型实时流媒体CP正在将其服务迁移到地理分布的云基础架构中。然而，由于现场广播的动态性以及观众和广播公司的广泛地域分布，用合理的资源来满足所有要求仍然是挑战。为了克服这一挑战，我们在本文中介绍了一种预测驱动的方法，该方法可以估计广播瞬间不同云站点附近的潜在观众数量。这种对分布式受欢迎程度的在线和即时预测将我们的工作与以前的工作区分开来，后者是在内容受欢迎程度发生变化时提供固定资源或更改其分配。基于得出的预测，我们制定了整数线性程序（ILP），以主动，动态地选择合适的数据中心来分配确切的资源并为潜在的观众提供服务，同时最大程度地减少延迟。由于优化不足以用于在线服务，因此我们提出了一种基于强化学习（RL）的实时方法，即RL-OPRA，该方法通过与网络环境进行交互来自适应地学习以优化分配和服务决策。