During the past two decades, the steady increase in the power of parallel super-computers participated heavily in developing 3D unsteady CFD modeling approaches. In these approaches, where the flow fluctuations are time and space resolved on a computational mesh, the cost of a simulation is directly linked to the size ratio between the largest and smallest resolved scales. Turbulent combustion and primary atomization modeling have both strongly benefited from this evolution as it enabled to increase the gas/liquid or burnt/unburnt gas interface resolution, the problem size and to include more physics. However, Direct Numerical Simulation (DNS) is still out-of-reach for most of practical configurations. Adaptive mesh refinement (AMR) is an appealing technique to reach DNS at a lower CPU cost. AMR has been originally designed for Cartesian grids and the major challenge for its use on distributed memory machines is its parallelization. The local mesh refinement indeed creates load imbalance that needs frequent repartitioning and balancing. The presentation will detail recent numerical developments on dynamic adaptation of tetrahedron-based unstructured grids. The use of tetrahedra has two advantages for practical configurations: complex geometries are easily meshed and the mesh is locally more isotropic than Cartesian grids. The proposed methodology relies on frequent sequential calls to a remeshing library (www.mmgtools.org), which adapt the mesh inside each MPI rank without modifying the interface shared with the other ranks. Then, repartitioning and transfer of cell groups is performed to ensure an optimal load balance and to modify the cells at the interface. All the underlying algorithms have been optimized to reach good performances with grids of several billion cells on more than 10'000 cores. This dynamic mesh adaptation strategy has been implemented in the YALES2 code (www.coria-cfd.fr) and applied to the modeling of premixed flames and primary atomization. In these applications, the local adaptation enabled to reduce drastically the CPU cost compared to the fixed grid approach and to reach unprecedented mesh resolutions at the interface.