Performing numerical analysis of structures made of composite materials leads to difficult problems. On one hand, in order to take into account the mecanisms induced by the material itself, a modelling at the material component scale is often necessary. On the other hand, the entire structure scale is generally much larger than the material components one. The coexistence of these two scales makes the global numerical problem very large. A way to solve that kind of problem at low cost was initiated by Sanchez-Palencia for periodic structure materials: work is done on a global homogeneized scale coupled with a local re-analysis. The "micro-macro" strategy that we proposed can also be related to the domain decomposition methods. These kinds of techniques are well suited to parallel architecture computers. Today, the most powerful methods are the multilevel ones. A global coarse problem is superposed to the substructured problem, then an iterative solver uses this global information. Our parallel and two-level technique is based on the mechanical problem at the continuum level. It can be described with the three following points: (a) a decomposition of the structure into simple constituents, substructures and interfaces, each of these components possessing its own variables and equations; (b) a mixed approach of the interface problem so that the interface problem deals with both displacements and forces; (c) the separation of the micro and macro scales on the interfaces. The "macro" homogeneized problem is global on the entire structure, and has a low cost due to its finite dimension. The "micro" problems are related to each substructure independantly but they remain independant of edge effects. Both problems are coupled and solved with an iterative technique. The proposed algorithm uses two search directions: one related to each level. They are the parameters of the method. Their choice is discussed with examples and performance evaluations. The performances are illustrated on highly heterogeneous structure calculations for which homogeneization is a crucial point. Moreover, boundary effects are automatically taken into account with this strategy. Additional examples will give a comparison, in term of convergence performances, between this method and the FETI method for homogeneous or heterogeneous cases.