Cluster systems are getting increasingly popular since they provide large computational power at a reasonable price. The cluster nodes are often SMPs with a small number of processors that can access a shared address space. The nodes are connected by a network like Myrinet or SCI, so the global address space is distributed. In this paper, we present a new programming model for such clusters of SMPs. The model allows the programmer to adapt his program to the two-level structure of the address space by providing a micro-level and a macro-level. The micro-level allows a thread formulation of multiprocessor tasks that are executed within a node of the cluster system. The macro-level allows the hierarchical structuring of multiprocessor-tasks according to the structure of the algorithm using message passing for data-exchange. We demonstrate the usefulness of the approach by runtime tests on several cluster systems with different node architectures and different interconnection networks.