We aim to derive a parallel asynchronous version of the non-intrusive Global-Local coupling. This method can be viewed as an exact version of the submodeling technique, allowing a coarse description of a structure and enriching the modeling of some zones of interest with enhanced geometry, refined mesh, and complex material laws. The exact interactions between the local patches and the rest of the structure are computed through iterations. The method can also be employed as a simple way to derive a non-overlapping domain decomposition method in the case of patches covering the whole domain. Mathematically, the method is a stationary iteration that proved to be particularly robust in solving nonlinear structural mechanics problems (plasticity, crack propagation). One key feature of the method is that it only makes use of standard operations so that it can be implemented as a python script driving industrial reference software like Abaqus, Code Aster...However, from a parallel point of view, the method is not too efficient. The global computation is done alternating with the local ones (in the patches), in a synchronized manner which creates moments of inactivity that affect the performance. Deriving the asynchronous parallel version should bring several advantages, in particular the approach will be more resilient and less sensitive to load balancing. It should also make better use computing resources, whether a high-performance cluster or a network of heterogeneous workstations. To prove the convergence of the asynchronous model with linear or nonlinear behaviors in the patches, we used the paracontractions framework.From an implementation point of view, several techniques have been tackled in the history, generally based on MPI. Recently, the one-sided communication was proposed in different works. The idea is to let each CPU directly access the memory of the other CPUs in order to get or to put pieces of information using so-called passive synchronization. In this presentation, we will address all these issues. We will measure the performance thanks to different implementations of the method using a python code driving the getfem finite element software and using the mpi4py module. We will consider nonlinear thermal and structural mechanics problems in 2D and 3D. A comparison with the synchronous version of the code will be provided. A weak scalability study will be performed on an academic configuration. A turbine blade from the high-pressure compressor of a turbomachine will be studied where patches are used to introduce micro-perforations in the geometry of the leading and trailing edges. The computations are carried out with the simulation center of the LMPS laboratory, which includes a cluster and a network of heterogeneous workstations.We will show that in many configurations the asynchronous method is faster than the synchronized version even when the latter is accelerated. To our knowledge, these results are the first to be realized with an asynchronous parallelization technique for such problem cases.