Design and Evaluation of a Hierarchical Characterization and Adaptive Prediction Model for Cloud Workloads

Abstract

Workload characterization and subsequent prediction are significant steps in maintaining the elasticity and scalability of resources in Cloud Data Centers. Due to the high variance in cloud workloads, designing a prediction algorithm that models the variations in the workload is a non-trivial task. If the workload predictor is unable to handle the dynamism in the workloads, then the result of the predictor may lead to over-provisioning or under-provisioning of cloud resources. To address this problem, we have created a Super Markov Prediction Model (SMPM) whose behaviour changes as per the change in the workload patterns. As the time progresses, based on the workload pattern SMPM uses different sequence models to predict the future workload. To evaluate the proposed model, we have experimented with Alibaba trace 2018, Google Cluster Trace (GCT), Alibaba trace 2020 and TPC-W workload trace. We have compared SMPM's prediction results with existing state-of-the-art prediction models and empirically verified that the proposed prediction model achieves a better accuracy as quantified using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE).