Workload And QoS Prediction For Data Analysis Services On Cloud

Data analysis services on cloud,DASC,utilizes cloud computing to fulfill their high demand on computing resources,and hence become a new cost-effective approach to implement data analysis.As a result,they receive broad concern in scientific computing,finance and other domains.The workload of a DASC consists of the service request rate and values of business parameters in the requests,which will change dynamically during the running of DASC and have significant impact on the QoS of it.However,none of current research works on QoS prediction for software on cloud considers the QoS influence brought by the change of business parameter values,and thus makes the prediction result unsatisfactory.To address this issue,a two phase DASC QoS prediction approach,TPQPA,is devised.In phase one,the service request rate and the values of business parameters in the requests of the workload are predicted.In phase two,the predicted workload is supplied as an input to predict the QoS of DASC to increase the prediction accuracy.First,a workload prediction approach is proposed as phase one of TPQPA to predict the service request rate and business parameter values in the requests of the workload at a designated time in an indicated day.In the approach,statistics analysis and machine learning is conducted on the basis of everyday workload information to get different time blocks(such as peak time,valley time,etc)within a day.All time blocks within a day comprises a daily workload,which is further clustered into several daily workload patterns.Then given an indicated date,a daily workload pattern is predicted with a SVM model.With the restriction of the pattern,the service request rate is predicted with K-Means and the business parameters are predicted by a fitted distribution function of business parameter values.Then,a workload based DASC QoS prediction method is presented as phase two of TPQPA.In the method,the QoS metrics of QoS-critical components of DASC are predicted with SVM based on the workload prediction and computing capacity of the environment.The QoS metrics of non-QoS-critical components of DASC are obtained by testing.Combining the two parts together,the QoS prediction of the whole DASC is available according to its deployment structure.At last,a case study about transaction data analysis application in finance domain is carrired out to demonstrate the application of TPQPA.Furthermore,a set of comparison experiments are conducted between TPQPA and other QoS prediction methods,and the result shows that TPQPA improves the accuracy of QoS prediction for DASC considerably.