While grids reach further to geographically separated clusters and data warehouses, grid applications pose new demands on end-to-end performance control. Data-intensive grid applications rely on the underneath network to bring together distributed computational and storage resources. From computational grids to data grids, the focus of resource utilization is shifting from computing power to network resources. This article investigates network resource sharing in grids, especially data grids. It studies traffic characteristics and quality of service( QoS) mechanisms of grid applications. A hybrid approach, which combines advance resource reservation and classical QoS differentiation, is proposed to meet grid application performance requirements. Identifying three types of resource reservation scenarios, that is, long-lived, short-lived, and flexible shortlived, the article formulates optimization problem of network resource sharing. Proven NP-complete, the problem is pursued with heuristics and simulations.