Data Grid, which consists of several geographically distributed datacenters linked by high speed network, is an ideal platform for the data-intensive and computingintensive scientific computing. Besides improving the computing performance and the data processing capabilities, the replication service among the nodes improves failure resistance and increase system availability. Replica selection is one important problem in replication optimization, because Grid application may need to retrieve data from many distributed nodes and do computation on their own local machine in parallel. In this paper, we focus on how to determine an appropriate set of replicas that at least cover the data, and farthest utilize the system parallel computing capacity. As we believe there is a trade-off between increasing parallelism and reducing redundancy as more replicas involved in computation, we put forward a fast replica selection algorithm inspired by the Utility Theory in economics to balance the two conflict demands.