Dynamic reconfiguration is the ability to modify a parallel or distributed system while it is running. We adopt the framework developed by Jeff Kramer and colleagues at the system architecture level: changes must occur in a consistent state, which is brought about by ``freezing'' some system components. The goal is to reduce system disruption, i.e., to minimize the part of the system to be ``frozen'' and the time taken by reconfiguration operations. Towards the first goal we take a connection based approach instead of a component based one. To reduce time, we refine the reconfiguration algorithm by executing changes in parallel as much as possible. Our model also handles hierarchic systems.