Performance of Predictive Resource Allocation in Homogeneous and Heterogeneous Networks

Predictive resource allocation can boost throughput remarkably by exploiting residual resource in cellular networks. In this paper, we investigate the performance of predictive resource allocation for non-real-time service such as video on demand in terms of boosting the maximal arrival rate when the stalling time during video playback of 95% users is less than the expected value of the users. To study the impact of duration of prediction window on the performance, we consider a two-threshold policy that is with low complexity but with performance very close to the optimal solution. We analyze the impact of prediction window duration, residual bandwidth, prediction methods，user association and inter-cell interference on its performance. Simulation results show that the two-threshold policy is robust to prediction errors if proper methods are designed for predicting the required information. The throughput gain of this policy over non-predictive method increases with the length of prediction window but the growing speed reduces gradually. When the variance of residual bandwidth is larger, the throughput gain of the policy over non-predictive method is greater. In heterogeneous networks, by using a residual-bandwidth based user association method, the two-threshold policy can achieve much better performance than using the maximal-receive-power based user association, and the throughput gain is higher when the traffic load is heavier.