Abstract : Virtual Reality applications require a huge amount of computational power that clusters, sets of computers connected with networks, are able to provide. To take advantage of these architectures, it is possible to split applications into several parts, called components, and to map them on different cluster nodes. Performance of such applications depends on hardware performance, on their mapping and also on the synchronization and communication schemes between components. To determine if a VR application can run in an interactive way, we can map and run it on the architecture. If the application does not perform as expected we have to try another mapping. However, it is often a long and tedious process before finding a mapping with expected performance. To speed up this process we define a performance model which enables to evaluate performance of a given mapping for a distributed application on a cluster from architecture, application, and mapping descriptions. Then, we propose an approach based on constraint programming to automatically generate mappings. Constraints are defined from our model, from performance of the architecture and also from performance expected by the user. This approach enables to answer the following answers: Does at least one mapping exists with the expected performance on the given architecture ? If it does then what are these mappings? Does the application performs better if we increase the number of nodes of the architecture?