In this paper, we present a fully decentralized algorithm for fair resource sharing between multiple bag-of-tasks applications in a grid environment. This algorithm is inspired from related work on multi-path routing in communication network. An interesting feature of this algorithm is that it allows the choice of wide variety of fairness criteria and achieves both optimal path selection and flow control. In addition, this algorithm only requires local information at each slave computing tasks and at each buffer of the network links while minimal computation is done by the schedulers. A naive adaptation is unstable and inefficient though. Fortunately, a simple and effective scaling mechanism is sufficient to circumvent this issue. This scaling mechanism is motivated by a careful study of the subtle differences with the classical multi-path routing problem. We prove its efficiency through a detailed analysis of a simple simulation.