Less Data Can Be Better for Domain Generalization