Geotechnical engineering is a crucial field in the design and construction of foundations, embankments, tunnels, and other structures involving soil and rock. However, the description of the elastoplastic response of soil, with preponderant non-linear and non-reversible deformations together with a non-associative flow rule, is complex. The difficulty is even higher in the case of non-monotonous loading paths, as for seismic events, where phenomenological constitutive relations require ad-hoc history parameters and advanced experimental tests for their calibration. This communication presents a 3D discrete element model (DEM) for simulating the behavior of soil with the aim of improving the reliability of its quantitative predictions of soil response to such complex loadings. This model is kept relatively simple in order to limit the computational cost and to tackle boundary value problems considered in engineering applications. The use of spherical particles with rolling resistance and an adequate representation of the soil initial state is shown to provide descriptions of the macroscopic soil response as satisfying as discrete models embedding more realistic particle shapes. The model is used to simulate the nonlinear interaction between a shallow foundation of building structure and the supporting soil during strong seismic loadings, as tested experimentally for the TRISEE project with a full scale physical model. An adaptative discretization technique is implemented to limit the number of particles in such a boundary value problem and make the computation possible with a conventional desktop computer. Numerical results are benchmarked against experimental measurements and previous simulations performed with the finite element method embedding phenomenological constitutive relations. Efficiency of the 3D DEM with respect to more conventional simulation methods and its usefulness for studying the behavior of geotechnical structures are discussed from the conclusions of this benchmark.