Developing Deep RNN-Based Client Resource Utilization Prediction Model for an Improved Cloud Resource Provisioning

Abstract

One of the key problems facing cloud computing is to estimate accurately the use of resources for future requirements. Cloud computing resource provisioning requires an adaptive and reliable forecast of cloud workloads. The consumption of cloud resources is continuously evolving, which makes it difficult to provide reliable forecasts for predicting algorithms. However, the prevalent approaches cannot effectively forecast high dimensional and highly variable cloud workloads. This leads to a waste of resources and thus, failure to meet SLAs. This motivates the research, which aims to predict the cloud resource utilization for better provisioning of cloud resources using a Deep RNN-Based architecture to improve the accuracy in the prediction of the workload. Since the RNN is sequential data analysis, the problem of workload prediction has recently not been tackled. RNN, however, frequently performs poorly on LTM dependencies in learning and it also does not reliably forecast workloads. To overcome these important challenges, stacked LSTM RNNs have been proposed for predicting CPU utilization, memory, and disk I/O due to the ability to maintain information and accurately predict time series issues and this makes it a promising candidate to forecast cloud resource utilization more accurately compared to traditional approaches. First, the method uses multivariate analysis to derive the sequential and contextual characteristics of the historical workload data via the LSTM network from the original high-dimensional workload data to compare between metrics and increase the applicability of the model. Next, the model integrates the different callback API, hyperparameters (optimizers, metrics, and losses), and layers (Regularization, Normalization, recurrent, and activation layers) to achieve the adaptive and accurate prediction for highly variable workloads. The experiment carried out on real-world Alibaba, Google, and Intel Netbatch workload trace to demonstrate that the proposed method achieves superior prediction accuracy and state-of-the-art performance compared to other workload prediction methods for high-dimensional and highly variable real-world cloud workloads. The maximum mean squared prediction error of the proposed model is reduced up to 0.000403.