Structured Prediction Method for Small Sample Workload Sequences

Abstract

Accurate workload prediction is the key to realize elastic resource management of cloud platform. Aiming at the problem that a large number of tasks with short running time achieve prediction in the cloud platform, which leads to the lack of training data of the forecasting model, a structured prediction of multivariable workload sequences (SP-MWS) method is proposed. It is based on the characteristics of intrinsic correlation among multiple resources consumed in the running of a single task, and the relationship of multi-dimensional workload sequences is explored to supplement the prediction information of small-scale sequence. Firstly, in order to obtain the related workload types, the maximum information coefficient (MIC) and information entropy are adopted to measure the correlation, and related workload types are selected. Secondly, for the selected multiple related workloads, trace-norm regularization multi-task learning (TNR-MTL) is introduced to construct prediction model to realize structural information mining of related workload sequences and complete prediction of multiple workloads simultaneously. Validated on Google cloud platform’s operational monitoring log dataset, the experimental results show that the proposed method can significantly increase model information; the decision-making basis of the prediction model is interpreted and the contribution of each variable to the prediction model is visualized. Comparative experiments show that, the proposed prediction method is better than the commonly used workload prediction methods in time performance and prediction accuracy.

Keywords

• |cloud computing|elasticity|workload prediction|multivariable prediction|structural information