Predicting Short-Term Variations in End-to-End Cloud Data Transfer Throughput Using Neural Networks

Predicting the data transfer throughput of cloud networks plays an important role in several resource optimization applications, such as auto-scaling, replica selection, and load balancing. However, constant short-term variations in cloud networks make the prediction of end-to-end data transfer throughput a very challenging task. The parameters that affect the throughput can be categorized into three different areas: end-system characteristics (e.g., disk I/O bandwidth, CPU utilization), network characteristics (e.g., network bandwidth, latency, background traffic, bandwidth shaping mechanisms), and dataset characteristics (e.g., average file size, dataset size). Although there are promising results in the literature using neural networks, the datasets are collected from network layer devices and memory-to-memory data transfers where end-system and dataset characteristics are not considered as part of the problem. Also, very few studies use multivariate time series data collected from cloud networks, and the variables differ from study to study. In this project, we collected multivariate time series data from Amazon Web Services (AWS) by conducting intra- and inter-region transfers between storage systems and compute resources using monitoring services. This dataset is unique in the sense that end-system metrics in addition to network throughput are collected from both source and destination systems. Different average file size, instance type, and regionality parameters provide various settings, making the dataset applicable to various types of prediction models. Our multivariate neural network models predict the one-step-ahead network throughput with similar to 3.7% and disk throughput with similar to 6.1% error rate, outperforming multivariate models with least-correlated variables and univariate models empowered by transfer learning.