Large-eddy simulation (LES) has been performed for an axisymmetric piston-cylinder assembly with and without swirl. For both cases, the LES mean and rms velocity profiles show better agreement with experimental data than profiles obtained using a Reynolds-averaged Navier–Stokes (RANS) approach with a standard k − ε turbulence model. The sum of the resolved and modeled contributions to turbulence kinetic energy (TKE) approaches grid independence for the meshes used in this study. The sensitivity of LES to key numerical and physical model parameters has been investigated. Results are especially sensitive to mesh and to the subfilter-scale (SFS) turbulence models. Satisfactory results can be obtained using simple viscosity-based SFS turbulence models, although there is room for improvement. No single model gives uniformly best agreement between model and measurements at all spatial locations and at all times. The strong sensitivity of computed mean and rms velocity profiles to variations in the SFS turbulence model suggests that better results might be obtained using more sophisticated models.