The development of a parallel unstructured multigrid solver for the solution of the compressible FavreAveraged Navier-Stokes equations is described. The flow domain is discretized using unstructured hybrid grids made up of hexahedral, pentahedral, tetrahedral and triangular prismatic cells handled by a single unifying edge-based data-structure. Turbulence is modeled using either zero or two-equation turbulent models. The different sub-grid levels are obtained by agglomeration and the domain decomposition in each subgrid is self contained in the corresponding fine-grid sub-domain avoiding the need for inter-processor/inter-grid level communication. It is shown that this approach does not significantly affects the multigrid convergence rate making it attractive from an implementation point of view. Special emphasis is made in controlling the agglomeration process in the presence of structured or semi-structured patterns and its application to turbomachinery configurations. I Introduction C omputational fluid dynamics simulations play a central role in the design of complex engineering components in general, and in particular, of modern gas turbines. These applications push regularly simulation technologies beyond current limits in terms of speed, capacity and geometric complexity. Unstructured methods are not only superior to deal with arbitrary geometries, from a mesh generation point of view, but provide a more homogeneous framework for the solver than multiblock methods. However it is acknowledged that the higher computational resources required by unstructured methods, specially in terms of memory, has slowdown a wider use, although improvements on the robustness and convergence rate of multigrid methods on unstructured grids and the appearance of low-cost parallel hardware have partially overcome some of these limitations. The homogeneous data-structure of unstructured methods has two advantages that are usually overlooked, namely, superior scalability on parallel machines and lower maintaining cost of the software, which is an added value for small organizations. Nowadays it is generally acknowledged that the most economic means of solving large computational problems are parallel, distributed memory architectures. This is specially true in current days where personal computer (PC) boxes are an inexpensive “out-of-the-shelf” source to construct clusters of CPUs. Explicit methods that use spatial discretizations with compact stencils are specially suitable for parallel implementations since they minimize the communication over