A general AI-defined attention network for predicting CDN performance

Abstract It has become an increasingly popular trend to replicate network content across a geographically distributed structure. The Content Delivery Network (CDN) is a prime example. Content providers have been struggling to improve the Quality of Experience (QoE) of its clients via accurate predictions of CDN performance, which is a crucial prerequisite for subsequent policies like resource scheduling. Methods of Artificial Intelligence (AI) like machine learning and deep learning have shown their advantages on performance predictions recently, but it is still challenging to simultaneously ensure accuracy, generality and low overhead. In this paper, we propose a general AI-defined attention network to make predictions on CDN performance, the results of which can improve CDN resource optimal scheduling. First, we formulate this CDN performance prediction problem as a sequence learning problem. And then we design a two-phase deep neural network with attention so that our model can work well for all cache groups in a CDN. Experiment results show that our model outperforms state-of-art models in accuracy while maintaining a relatively low training overhead.