Task performance collection and classification method in cloud platforms
Chunyi LIU,Xiao ZHANG,Yuansong QIN,Shangqi LU
DOI: https://doi.org/10.11772/j.issn.1001-9081.2017102790
2018-01-01
Abstract:It is difficult for users to determine the type of cloud hosts on cloud platforms when they are actually using cloud platforms,which results in low utilization of cloud platform resources.In some typical methods to solve the low resource utilization,the placement algorithms are optimized from the perspective of cloud provider,and the user selection will limit the utilization of resources;while in other methods,the collection and prediction of task performance under the cloud phtform in a short time are made,but it will reduce the accuracy of task classification.In order to achieve the goals of improving cloud platform resource utilization and simplifying user operations,a multi-attribute task performance collection tool,named Lbenchmark,was proposed to collect the performance characteristics of task comprehensively,and the load was reduced by more than 50％ compared with Ganglia.Then,with the performance data,a K-Nearest Neighbor (KNN) application performance classification algorithm with the multiple K-Dimension (KD) tree based on configurable weights was proposed.The suitable parameters were selected to establish multiple KNN classifiers with KD tree,and the cross validation method was used to adjust the weight of each attribute in different classifiers.The experimental results show that,compared with the traditional KNN algorithm,the calculation of the proposed algorithm is significantly increased by about 10 times,and its accuracy is averagely improved by about 10％.The proposed algorithm can use data feature mapping to provide resource recommendations to users and cloud providers,improving the overall utilization of cloud platforms.