Traditional GIS has many flaws in managing distributed, large-scale, heterogeneous spatial data. Nowadays, more and more countries build their own spatial data grid to provide abundant data services. But data grid system is a complicated one and involved in lots of entities which are different in many ways, including type, performance, runtime environment, etc. It is necessary to divide the whole system into different layers, which has great importance to simplify the system complexity, improving system stability and establish an efficient, extensible, self-adaptive and self-organizing system. Here, after carefully analyzing various system entities, the article describe the idea that entities are divided into three kinds service-type entities, management-type entities and draw-type entities that is the basis of clear level classification of the spatial data grid and establishment of a mixed three-tier architecture based on grid computing and P2P technologies.