Due to network bandwidth growth, grid computing has been envisionned. As grids are aggregation of computer clusters based on high performance local networks and interconnected by very high speed core networks, access links to core networks remain bottleneck locations. In order to optimize grid computing overall performance, we propose to organize data transfers from one cluster to an other. This organization must allow ressource schedulers to schedule data transferts in addition of tasks. Based on a grid network resource reservation model this paper presents the design f a grid gateway, located at the LAN/WAN interface. To implement this approach, we have develop a prototype based on Network Processors IntelIXP2400 which is able to control flows at 1gigabits/s. We discuss the main design choices and experimental results.