A Type-Aware Workload Prediction Strategy for Non-stationary Cloud Service

Under cloud environment, cloud service providers need to predict non-stationary workloads, and base prediction results to plan and execute fault-tolerance strategies in advance so as to ensure services' ability in fault-tolerance. High prediction precisions in services' workloads improve effectiveness of strategies such that service's fault-tolerance is advanced. In this paper, we propose a hybrid prediction strategy, named as NUP, which dynamically judges workloads' types and relies on their type to self-adaptively switch prediction algorithms. Moreover, it first depends on autocorrelation coefficients and Hurst exponents of workloads to confirm whether the workloads belong to the period or the trend. Then NUP makes use of linear regression and similarities among periods to replace missing data of trend and period workloads separately. After that, NUP adopts linear regression combined with ARMA to predict the trend and SVR for the period. Experiments demonstrate that NUP boosts prediction precisions of both period and trend workloads. Under trend workloads, prediction errors of NUP are only 52% of that from ARMA as well as 48% from Holt-Winters. With regard to period scenario, prediction errors decrease to 25% and 70% of ARMA and Holt-Winters separately.