A Bandwidth Prediction Method Based on Hybrid LSTM for Content Delivery Network

Abstract

Content Delivery Network (CDN) can store workflow content by deploying edge service nodes and realizing network optimization with reasonable scheduling strategies. Manual selection of appropriate service clusters is coupled with low timeliness and a high economic burden. Therefore, the timing prediction of bandwidth remains a persistent demand to choose service clusters that can take on a safe workload. In this study, we proposes a model which can predict the load level in the future period to optimize the content delivery network. This model uses a machine learning method (K-means) for data optimization, which discarded 3114 of 185572 pieces of data that impact subsequent prediction models. Afterward, the model uses a 4-layers Long Short-Term Memory Network(LSTM) to predict the aggregated temporal data. The model named BK-LSTM considers the execution time and accuracy, eventually learning the real-time bandwidth demand pattern of 654 servers in the specific cluster. Experiments show that our BK-LSTM model has a mean absolute percentage error(MAPE) metric of about 15.2% on the test set, demonstrating this model’s ability to predict bandwidth workload well.