A remarkable class of soft real time applications exhibits a very dynamical behaviour due to the variations in the treated data. Moreover, such programs have to be able to run on hundreds of different platforms. As a consequence, classical real time scheduling algorithms are not flexible enough since they are based on the exact knowledge of the tasks' timing parameters. Some of the approaches proposed so far in the literature guarantee temporal isolation, but they make a static assignment of resources to each task, which, once again, is based on an a priori knowledge. The authors propose a closed loop method for online adapting of the fraction of assigned resource to the task requirements. The approach is based on adaptive control techniques and has resulted in being effective in a significant set of real life experiments