Optimizing the well configuration of an oil field in reservoir engineering consists in maximizing the Net Present Value function, which couples drilling/operational costs and production profits, associated with produced oil as well as with produced and injected water volumes. Such profits are computed from fluid flow reservoir simulations. The decision variables are the number of active wells, their type (producer/injector) and their location. Standard MINLP formulation of such a problem involving complex PDE leads to a huge discretized equation system, which cannot be solved in practice. Thus, we separate the computation of the NPV objective function for a given well configuration and the optimization phase. We delegate the former to the fluid flow simulator and the latter to a black-box optimizer. We present computational optimization results on a 3D realistic reservoir case by using the direct search implementation of NOMAD solver. The impact of tailored exploration phase and adapted surrogate models are illustrated.