The highly dynamic and heterogeneous edge computing power demand in cloud edge collaborative systems results in lack of foresight in scheduling of the resources. This study proposes a resource demand prediction model for the purpose of dynamic scheduling optimization of computing power. It built an input sequence that can integrate the heterogeneous features that are provided by the multiple sources, and designed long short-term memory (LSTM) model based on attention mechanism to capture the complex temporal dependencies in long sequences. The training process consists of the use of mean square error loss and Adam optimizer for end-to-end learning. Based on the predicted results, this article further designs a dynamic scheduling algorithm with a target to reduce the delay of tasks and wastes of resources. The proposed prediction model can still maintain the RMSE of 12.89% under the condition of 10 hour prediction duration which shows good prediction stability in the long term; The scheduling scheme based on the above prediction results in the average resource utilization rate of 85.13% and the average task delay time of 124.065 milliseconds in the simulation, which proves that the scheduling scheme can effectively improve the resource utilization efficiency and response performance of the scheduling system through forward-looking information.
Wang et al. (Thu,) studied this question.