A Predictive Technique for Replica Selection in Grid Environment

Replication in a data grid reduces access latency and bandwidth consumption. However, when different sites hold replicas of a particular file, there is a significant benefit realized by selecting the best replica from among them. The best replica is the one that optimizes the desired performance criterion such as absolute performance (i.e. speed), cost, security or transfer time. By selecting the best replica, the access latency can be minimized. We develop a predictive framework that uses data from various sources and predicts transfer times of the sites that host replicas. With this estimate, one site can request the replica from the site that has the lowest transfer time. We use a neural network (NN) for transfer time prediction of different sites that currently hold file replicas. We compare the results with a multi-regression model and the simulation results demonstrate that the neural network technique is capable of predicting transfer time more accurately than the regression based model.