The size, heterogeneity and dynamism of the execution platforms of scientific applications, like computational grids, make using those platforms complex. Furthermore, today there is no effective and relatively simple solution to the programming of these applications independently of the target architectures. Using the master-worker paradigm in software components can provide a high level abstraction of those platforms, in order to ease their programming and make these applications portable. However, this does not take into account the dynamism of these platforms, such as changes in the number of processors available or the network load. Therefore we propose to make the master-worker abstraction dynamically adaptable. More specifically, this paper characterizes the master-worker paradigm on distributed platforms, then describes a decision algorithm to switch between master-worker implementations at run-time.