Loading Prediction Of Host Computer In Distributed System

In the parallel distributed system having several processing elements(PE), the loading of each PE will be unbalancing when the distributed tasksof PE are non-uniform. In order to using resource of each PE efficiently andfeasibly, loading balance is introduced to parallel distributed system. Physicalstructures of each PE are different, but their logic structures are separate anduniform. The running tasks on each PE are equality and distributedcompletely.Technology of loading balance in distributed system can be divided intothese classes: static loading balance and dynamic loading balance, explicitloading balance and implicit loading balance, centralized loading balance anddistributed loading balance.This paper is aimed at loading character of host computer. Taking statusof host computer forecast in domestic and foreign into account, the paperprovides a appropriate forecast algorithm based on time series analysis. Itcompares several different algorithms based on time series analysis with eachother, and offer a feasible algorithm of prediction error. Emphasizing realtime correction and on-line estimate, it greatly improves prediction efficiencyand precision for loading balance. Via simulation verification, the predictionalgorithm offered in the paper is proved to be feasible and efficient. Thealgorithm includes the following parts:1) Choice for model structure: From loading character of host computer, hostcomputer loading is a kind of time series and has a strongly relating character.It means that time sequence model can be adopted to predict loading of hostcomputer. Time sequence model can be used for rapid and efficientprediction.2) Real time correction of model parameter: After confirming a modelstructure, it should adopt a real time algorithm of estimating model parameterfor overcoming the shortcoming that the model is not suitable for complexityloading of host computer and for preparing for the real time prediction of thesubsequent loading. The estimate algorithm is requested to be relativelysimple and rapidly converge. It should have a high precision for estimatingparameter to be using on-line.3) Predicting correction of model error value: For improving predictingprecision, a future error value can be predicted for correcting predicting valuethrough an error analysis of the process that already happened for seeking achanging regulation.4) Prediction of status variable: It is through the Kalman Filtering Algorithmto correct status variable for real time. It can improve the predicting precisionafter correct the later predicting value.5) Model evaluation for predicting on-line: It should offer a feasible andefficient on-line predicting algorithm to resolve the problem that happens inoff-line evaluation.6) Predicting principle of several differential predicting algorithms running inthe same time: The predicting loading status of host computer is divided intotwo parts, smooth running status and adjusting status. In the smooth runningstatus, one predicting algorithm is running and in the same time several otherdifferent predicting algorithms are standby and they are not participating incalculating and analyzing. When the average error value exceeds thethreshold value settled beforehand, the system run in the adjusting status.Firstly, parameters of this algorithm should be adjusted, and in the same timeseveral other differential predicting algorithms are in on-line standby status.If the error ratio can not be reduced or other algorithm is better than thisalgorithm, the better one will be selected.7) For the feasibility of a better verification algorithm, the algorithm isverified. The result proves that the method of several algorithms running inthe same time and taking real-time correction can improve predictingprecision and efficiency. In order to evaluate the system capability and ensure the loading target,the resource of a processor is divided into four parts: (1)CPU, the efficiencyof CPU. (2)I/O, the efficiency of I/O. (3) bandwidth, the efficiency ofbandwidth. (4)cache in memory, the efficiency of cache in memory. Theaccess speed of memory's cache is more rapid than that of magnetic disk.The different task consumes different resource. So when system's capabilityis evaluated, different coefficients are assigned to different resources. Thecoefficients are determined by statistics result of measured values. It canensure the result of predicting algorithm by analyzing different predictingalgorithms of error functions. This predicting algorithm has the following merit: 1) The fewer training data is needed. It means that Least Square Methodwith forgetting factor can be used and accurate parameter estimation can beacquired. 2) This predicting algorithm can adjust parameter value of model inon-line status. It can be suitable not only for demand of off-line predictionbut also for demand of on-line dynamic prediction. 3) The process of making a multi-algorithm model of time seriesanalysis is simple relatively. It bases on the principle of least error, so it canavoid identification of model structure of traditional times series analysis...