Parallel computing is widely used for large scale three-dimensional simulation of seismic wave propagation. One particularity of most of these simulations is to consider a finite computing domain whereas the physical problem is unbounded. Additional numerical conditions are then required to absorb the energy at the artificial boundaries, which introduces a different formulation and a load-imbalance. In the context of finite difference method, we study the use of thread overloading approach to alleviate the imbalance. We introduce a mixed-hybrid parallel implementation based on a classical cartesian partitioning at the MPI level and a self-scheduling algorithm at the thread level to handle more than 700 threads on 8 processors. We demonstrate the efficiency of our methodology on an example of regional modeling performed on 80 processors.