Simulations of large-scale structure in the universe have played a vital role in observational cosmology since 1980s in particular. Their important role will definitely continue to be true in the 21st century. Rather the requirements for simulations in the precision cosmology era will become more progressively demanding; they are supposed to fill the missing link in an accurate and reliable manner between the ``initial condition at z=1000 revealed by WMAP and the galaxy/quasar distribution at z=0 - 6 surveyed by 2dF and SDSS. In this review, I will summarize what we have learned so far from the previous cosmological simulations, and discuss several remaining problems for the new millennium.