The common analytical sensitivity method of the large-scale structure takes a lot of time for solving the large-scale equations, the Domain Decomposition Method (DDM) of the FEM stiffness matrix is proposed in the paper. It can make use of the high accuracy of analytical method, and improve the computational efficiency by compressing the large-scale matrix into the smaller. In the large structure, the pseud load of the numerical sensitivity has a lot of zeros, and the displacement components with relevant to design variables have non-zeros. The displacement components are reordered behind others through reordering nodes, by which the stiffness matrix is assembled. Then the stiffness matrix is decompressed to smaller matrixes using the Domain Decomposition Method (DDM). The computational cost can be reduced a lot by using these matrixes to calculate the sensitivities of displacement. These data can be used repeatedly in the optimization process, thus the efficiency is improved greatly.