A cluster workload forecasting strategy using a higher order statistics based ARMA model for IaaS cloud services. (23rd March 2022)

• Record Type: Journal Article
• Authors: Amekraz, Zohra; Hadi, Moulay Youssef
• Is Part Of: International journal of networking and virtual organisations. Volume 26:Number 1/2 (2022)
• Page Start: 3
• Page End: 22
• Publication Date: 2022-03-23
• Subjects: IaaS cloud services -- workload prediction -- cluster workload -- autoregressive moving average -- ARMA -- higher order statistics -- HOS; Computer networks -- Periodicals; Information networks -- Periodicals; Virtual corporations -- Periodicals; Electronic commerce -- Periodicals; Management information systems -- Periodicals; Management -- Data processing -- Periodicals; 004.65
• Languages: English
• ISSNs: 1470-9503

Abstract

With cloud services becoming more popular among internet users, cloud providers are facing a challenge in allocating resources according to users' demand instantly due to the delay caused by the virtual machines' start up time. This problem can be solved using proactive allocation techniques that predict the workload in advance and make scaling decisions ahead of time. In this paper, we present an adaptive workload prediction method based on higher order statistics (HOS) and autoregressive moving average (ARMA) model. We use HOS to make a Gaussianity checking test of the cloud workload and decide the suitable identification method of the ARMA model to be used for forecasting. We evaluate our proposal with two real traces extracted from cluster workloads. The results show that the proposed method has an average of 34% higher accuracy than the baseline ARMA model and presents a low forecasting overhead (< 2 s).