Using NARX Neural Network Based Load Prediction to Improve Scheduling Decision in Grid Environments
Jin Huang, Hai Jin, Xia Xie, Qin Zhang · 2007
In grid environment, applications are in active competition with unknown background workloads introduced by other users. To achieve good performance, performance models are used to predict the possible status of the resources, and to make decisions of the selection of a performance-efficient application execution strategy. In this paper, we present a scheduling decision method that utilizes the NARX neural network based load prediction to define data mappings appropriate for dynamic resources. This method uses the information of the predicted CPU load interval and variance of future resource capabilities to obtain the CPU load decision, which can be used to guide the scheduling decision. As to the predictor used here, the NARX neural network based predictor learns the model of the system from the external input information and the system itself. It inherits the mapping capability of feed forward networks and, at the same time, captures the dynamic features of load information. In this work, our predictor shows good performance for time series prediction.