Replica placement and selection strategies in data grids

The thesis presents a set of dynamic replica placement and selection algorithms for Data Grid environments. We use the p-center, p-median, and a multi-objective model to select the candidate sites to place replica. We propose a replica selection algorithm based on the k-nearest neighbor rule. A predictive technique for estimating the file transfer time for different sites currently hold the replicas of a file is also presented in the thesis.;The p-median model places replicas at the sites that optimizes the request-weighted average response time. The p-center model selects the candidate sites to host replica by minimizing the maximum response time. We also consider a multi-objective approach that combines p-center and p-median objective to decide where to place a replica. A replica maintenance algorithm is proposed to relocate replicas to different sites if the performance metric degrades significantly in subsequent periods. To select the best replica from local information, a simple technique called the K-Nearest Neighbor (KNN) rule is exploited. The KNN rule selects the best replica for a file by considering previous file transfer logs indicating the history of the file and those nearby. We also propose a predictive technique to estimate the transfer time between sites. The predicted transfer time can be used as an estimate of transfer bandwidth of different sites that hold replica currently and help in selecting the best replica among different sites.;Replica placement algorithms are tested with three performance metrics: total file transfer time, the number of local file accesses, and the number of remote file accesses. The p-median and multi-objective model performs better than the p-center model for different types of job scheduling strategies. The k-nearest algorithm uses the local information for replica selection and shows a significant performance improvement over the traditional replica catalog based model that uses global information to select replicas. The neural network predictive technique is more accurate than the multi-regression model for predicting the transfer time between the requesting site to other sites that hold replicas of a file currently.