Grids rely on a complex interconnection of IP domains that may exhibit changing performance characteristics and may offer different quality of service (QoS) facilities. We examine the case of a biomedical application distributed over a grid and show it may suffer from uncontrolled communication performance. Then we present the QoSINUS service that dynamically allocates the network resources to Grid flows in order to match their specific QoS requirements under different load conditions. The aim of this approach is to optimize the end to end performances the heterogeneous mix of grid flows get from the network to enhance the individual application's performance as well as the overall grid infrastructure performance and utilization level. The QoSINUS service is based on the programmable network approach that offers flexibility, evolutivity and enables dynamic adaptation to network load variations. Finally, results of QoSINUS experiments conducted in the context of the eToile french grid testbed based on the high speed and DiffServ capable research network infrastructure, VTHD, are presented.