Department of Computer Engineering
MULTIVARIATE CLOUD WORKLOAD PREDICTION WITH DEEP LEARNING
(Supervisor: Prof. Dr. İbrahim Körpeoğlu)
Computer Engineering Department
Over the years, cloud computing has undergone remarkable changes considering the perspectives of both service providers and customers. Maximizing the utility of data centers to increase profits while providing a high quality of service appears to be a challenge in data center management. Therefore, an effective resource allocation and scaling policy is required while managing data centers. Although there are various factors that affect resource scaling such as the number of users and the current status of the system, provisioning of resources such as central processing unit (CPU), storage (disk space), networking (bandwidth) and memory (RAM) has become a significant aspect in cloud computing. In order to acquire a proactive resource scaling policy and improve resource utilization, it is necessary to scale data center resources prior to the arrival of the workloads. To achieve this goal, various techniques such as time series and machine learning models have been proposed. However, effective prediction of workloads with high dimensionality has become a major challenge for these models. We introduce and analyze deep learning-based models for multivariate cloud workload prediction. First, a Long Short Term Memory (LSTM) network is constructed to handle time-dependency of workloads. Secondly, 1D Convolutional Neural Network is introduced to process parallel sequences. Lastly, LSTM autoencoder is proposed to use the compressed representation of workloads for future predictions. These deep models are compared to a baseline model, Vector Auto Regression, with respect to their overall effectiveness and efficiency. Workload traces from Alibaba cloud data centers are used for training and testing purposes.
DATE: 05 April 2021, Monday @ 15:30