Home > Research > Publications & Outputs > Intelligent time-series forecasting framework f...


Text available via DOI:

View graph of relations

Intelligent time-series forecasting framework for non-linear dynamic workload and resource prediction in cloud

Research output: Contribution to Journal/MagazineJournal articlepeer-review

Article number109653
<mark>Journal publication date</mark>30/04/2023
<mark>Journal</mark>Computer Networks
Publication StatusPublished
Early online date28/02/23
<mark>Original language</mark>English


The industrial revolution 4.0 (I4.0), internet of things developments, and the expansion of online web services have caused exponential growth and deployment in the number of cloud data centers(CDC). Cloud computing is a paradigm that enables tenants to use storage and computing resources in the pay-per-use model. Cloud service providers maximize their profits by distributing the tenant's demands to the reserved storage and computing servers, minimizing the reservation cost with the satisfaction of the tenant's quality of service level agreement. Workload prediction and resource management are fundamental and critical problems due to cloud-distributed infrastructure and nonlinear dynamic workload conditions. Conventional prediction techniques in a cloud environment provide the one-dimensional output. Existing solutions mostly forecast resources, such as CPU and memory usage, each as a single output. However, the one-dimensional output in the form of resource provision and usage is not able to capture the relationship of application requirements of multiple resources such as CPU, memory, CPU cores, Disk, and network, which result in inaccurate prediction results and limited information. Efficient resource management requires predicting multiple resource parameters using multivariate state variables for efficient resource allocation. This study proposes an intelligent computing framework based on multivariate time-series bidirectional long short-term memory (BiLSTM) forecasting for predicting cloud virtual machine resources. We consider multi-dimensional resources such as CPU provisioned and usage, memory provisioned and usage, CPU cores, Disk write and read throughput, and network receives and transmit throughput. We investigate several deep learning techniques the proposed multivariate BiLSTM, LSTM, stacked BiLSTM, stacked BiLSTM with LSTM, and BiLSTM auto-encoder. Furthermore, we evaluate the effectiveness of the proposed framework on two real workload traces: Bitbrains traces fastStorage and Rnd. The performance metrics used to evaluate forecasting accuracy are the root mean square error, mean absolute error and mean absolute percentage error. Furthermore, we observe the training-testing data size and the historical window size variation effects on these models.