Abstract: The timely availability of data is very crucial for the efficient execution of jobs in grids, especially in the context of data grids. In the replica selection procedure it is required to fetch data cached in replica across the grid in the least possible time. This selection problem has been investigated in literature thoroughly and most approaches minimize total execution time by minimizing delays in transferring accessed data. Usually this is done by minimizing lookup time by either using a classification technique like K-Nearest Neighbor rules [8] or using predictive techniques such as regression [12] or using neural network techniques [16]. All these methods attempt to predict the total transferring time instead of using traditional models which look up catalogs. In this paper, a totally new approach to replica selection is proposed that aims to optimize transfer by probing for current network congestion status and opts for the most efficient set of replica’s sites that will work concurrently to transfer requested files or their parts. To achieve this goal, an association technique [3] is used to extract the replica’s sites that have shown best possible efficiency and will work together to accelerate the response to the user’s request. Our simulation results show an improvement of 29% as compared other reported work like the methods mentioned above and 40% better than traditional models that use replica lookup time.