Dynamic load balancing is an important step conditioning the performance of parallel programs, like adaptive mesh refinement codes. If the global workload varies drastically over time (such that memory is exceeded), it can be relevant to adjust the number of processors while maintaining the load balanced. We propose two different solutions, that extend classic graph repartitioning approaches to accept a variable number of processors: one based on biased partitioning method and one based on a diffusive method. We call this problem: the MxN graph repartitioning problem. Finally, an experimental study on real-life graphs validates our algorithms against state-of-the-art methods.