In order to perform structural analysis on parallel computers, with shared or distributed memory, an iterative approach related to domain decomposition methods has been designed. The studied structure is described as an assembly defined from two mechanical entities: sub-structures and interfaces. The problem is here formulated in the case of linear elasticity, for which we are only interested in the final configuration. After a description of the principles of this method, a multi-level approach is introduced in order to improve performances. Effectively, when decomposing into sub-structures, the presence of two scales can be noticed (one related to the dimension of the structure, and one related to the dimension of the sub-structures). Adding a large scale problem on the whole structure allows to quickly broadcast information among sub-structures because it takes into account the large variation length effects. The goal is to illustrate the feasibility and the performances of such an approach, especially for a large number of sub-structures. A close attention is devoted to the influence of the discretisation choice for the unknowns of the problem, which are both displacement and efforts on interfaces.