In the distributed simulation system, lots of redundant data flowing on the net seriously restricts the system scalability. The relevance filtering base on even-grid reduces the redundant data while it is inaccurate in region matching. Because the sizes of all cells in even-grid are identical, it is difficult to fit all entities in the simulation. A relevance filtering method based on a multi-level grid is put forward which overcomes the above shortcomings and is characterized by agility and speediness.