Numerical prediction of air-entraining and submerged vortices in pump sumps is important for engineering applications. The validation of pump sump simulations, however, still is not enough, because the simulations is very complicated; for examples, treatment of gas-liquid interface, detection method of the vortices and selection of turbulence model etc. We conducted numerical simulations of the benchmark experiments of the pump sump conducted by Matsui et al. (2006, 2016) and compared the simulation with the experimental data to investigate the effects of turbulence model, grid density and detection method of the vortices. We determined a threshold of the gas-liquid fraction function of VOF method (α) and the second invariant of velocity gradient tensor (Q2) to detect the air-entraining and submerged vortices by using vorticity, respectively. This method well detected the vortices and well reproduced the experiments for the RANS simulation using SST k-ω model. Large eddy simulation using Smagorinsky model, however, was sensitive to the grid system and difficult to reproduce the experimental vortex structures even for the finest grid system having 3.7 million cells.