Performance Prediction Model In Distributed Environment

Inthepast20years,distributedcomputinghasbeendevelopedrapidlyand used widely to satisfy the demands for computing power. On the otherhand, it brings us a lot of problems on the management of the computa-tional power. In order to take advantage of this tremendous computationalpower, it is quite important to provide an efective and efcient schedul-ing program for the tasks and resources. The scheduling program wouldwork better if the load of host in the distributed system can be predictedaccurately.This paper introduces a novel two-layer feedback ensemble modelwiththeaimofincreasingthepredictionaccuracyandstability. Ourmodelconsists of two parts, predictor optimization module and predictor ensem-ble module. First with the predictor optimization module the performanceof constituent predictors can be improved continually which helps get bet-ter ensemble result. And then in predictor ensemble module the resultsgenerated by the constituent predictors are ensemble to obtain better per-formance than any of them. The predictor optimization module helps toget better results in ensemble module, in turn the ensemble module willafect the optimization module.First, we apply this ensemble prediction model to single-dimensionprediction. The evaluation of the approach is conducted on diferent kinds of data set and shows a satisfactory performance compared with the con-stituent predictors. Then after introducing several multi-dimension pre-diction model including some machine learning models, we optimize thesupport vector regression prediction model. We build an multi-dimensionensemblepredictionmodelusingtheoptimizedSVRpredictionmodelandother traditional models. A series of experiment are conducted on severaldata set of CPU load and memory usage and a satisfactory performance isachieved to prove the efectiveness of the ensemble prediction model.