Plume dispersion of hazardous materials within urban area resulting from accidental or intentional releases is of great concern to public health. Many researchers have developed local-scale atmospheric dispersion models using building-resolving computational fluid dynamics. However, an important issue is encountered when determining a reasonable domain size of the computational model in order to capture concentration distribution patterns influenced by urban surface geometries. In this study, we carried out Large-Eddy Simulations (LES) of plume dispersion within various urban areas with a wide range of obstacle density and building height variability. The difference of centerline mean and r.m.s. concentration distributions among various complex urban surface geometries becomes small for downwind distances from the point source greater than 1.0 km. From these results, it can be concluded that a length of a computational model should be at least 1.0 km from a point source.