SrTiO3 is a typical cubic perovskite and serves as a candidate for thermoelectric materials. To improve the performance, it is necessary to reduce its inherently high lattice thermal conductivity by introducing lattice defects such as grain boundaries. However, the atomic structures and compositions of grain boundaries that effectively suppress thermal conduction in SrTiO3 have not been elucidated. Here, we have systematically calculated the thermal conductivity of 88 SrTiO3 symmetric tilt grain boundaries, including stoichiometric, TiO2-rich, and SrO-rich ones, using molecular dynamics simulations. The result shows that the excess volume of grain boundary is crucial in determining thermal conductivity, as is the case with ionic MgO. Further analysis also reveals that SrO-rich grain boundaries exhibit lower thermal conductivity than TiO2-rich ones due to their higher excess volume and weaker Sr-O bonds. Grain boundary non-stoichiometry is an important factor to control lattice thermal conduction in SrTiO3.