4. Asynchronous sub-domain and multi-decomposition methods

As we have seen, solving large-scale applications on a parallel computer requires the problem to be broken down into interconnected sub-problems. For parallel algorithms to be effective, the decomposition must be sufficiently granular. Indeed, sub-problem sizes that are too small will have a negative effect on method efficiency, increasing the cost of communication and synchronization between computational processes, and thus degrading the performance of parallel methods. This is why developers group computational tasks together to avoid this type of algorithmic behavior. For example, in the numerical solution of pseudo-linear problems, this type of grouping of computational tasks leads to the development of sub-domain methods or, more generally, multi-decomposition methods.