Research On Workload Structural Prediction Algorithm For Elastic Resource Management In The Cloud

Cloud computing is a computing paradigm with pay-as-you-go pattern.Elastic resource management is a key means to provide services on demand for the cloud platform.With the increasing requirement of cloud elastic service which focuses on the rapid variation of resource demand,it is of great significance to make a reasonable prediction of a large number of complex workloads for dynamic resource management to satisfy elastic service in the large-scale cloud platforms.Aiming at the workload prediction problem,this paper constructs three kinds of workload joint prediction models based on structured information,which includes the workload sequences among similar tasks and the multi-dimensional workload sequences in a single task.And the proposed workload prediction models are verified on Google cluster trace.The main works are as follows:(1)Aiming at the problem of the simultaneous arrival of a large number of tasks and the diversities of workload variation patterns in the cloud platform,a joint prediction method based on structured information among workloads is proposed.It can obtain valuable trends information from workload sequences to realize simultaneous and effective prediction for multiple workloads.First,in order to explore the intrinsic relationship information among different tasks with the periodic and irregular variation trends,two clustering strategies based on sequence and model are proposed respectively.Second,for the clustered tasks,trace-norm regularization multi-task learning(TNR-MTL)model with structural output characteristic is established to realize collaborative prediction for multiple tasks.Experiments results show that the effectiveness of the proposed clustering strategies is verified from prediction accuracy and task characteristics by analyzing the clustering process;both the computational time and prediction accuracy outperform other methods for workloads with different variation patterns.(2)Aiming at the problem that temporal correlation among workload sequences is not fully utilized in the workload prediction,a joint prediction method of multi-workload sequences based on temporal correlation is proposed.It can describe the variation trend of workloads,and improve the prediction accuracy by using the temporal characteristic of workload and spatial correlation among workloads.First,long short-term memory(LSTM)is adopted to extract the temporal feature of workload sequences,and the representation of sequences feature of the original space is transformed into the temporal feature space of workloads.Second,hierarchical clustering algorithm is utilized to obtain workload sequences clusters.Third,for the obtained each sequence cluster,the TNR-MTL model is introduced to construct the prediction model that achieves joint prediction of multiple workload sequences,which can capture the shared domain knowledge among multiple workload sequences.The experiment results demonstrate that the temporal feature clustering can effectively extract and utilize the global temporal feature of workload sequences,reduce the noise of the original sequences and get the workload sequences clusters with similar characteristics.The proposed method outperforms the existing methods in terms of prediction accuracy.(3)Aiming at the problem that a large number of tasks with short running time achieve predict in the cloud platform,a structured prediction of multivariable workload sequences method is proposed.It is based on the characteristics of intrinsic correlation among multiple resources consumed in the running time in a single task,and explores the relationship of multi-dimensional workload sequences to realize the prediction of small-scale sequence.First,in order to obtain the related workload types,the maximum information coefficient(MIC)and information entropy are adopted to measure the correlation,and select related workload types.Second,for the selected multiple related workloads,the TNR-MTL model is introduced to construct the prediction model to realize the structural information mining of related workload sequences and complete the prediction of multiple workloads simultaneously.The experiment 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;the proposed prediction method is better than the wide prediction methods in time performance and prediction accuracy.