Accurate computations of turbulent flows using the Large-Eddy Simulation (LES) technique with an appropriate SubFilter Scale (SFS) model require low artificial dissipation such that the physical energy cascade process is not perturbed by numerical artifacts. To realize this in practical simulations, energy-conserving numerical schemes and high-quality computational grids are needed. If unstructured meshes are used, the latter requirement often makes grid generation for complex geometries very difficult. Structured Cartesian grids offer the advantage that uncertainties in mesh quality are reduced to choosing appropriate resolution. However, two intrinsic challenges of the structured approach are local mesh refinement and representation of complex geometries. In this work, the effectiveness of numerical methods which can be expected to reduce both drawbacks is assessed in engine flows, using a multi-physics inhouse code. The overset grid approach is utilized to arbitrarily combine grid patches of different spacing to a flow domain of complex shape during mesh generation. Walls are handled by an Immersed Boundary (IB) method, which is combined with a wall function to treat underresolved boundary layers. A statistically stationary Spark Ignition (SI) engine port flow is simulated at Reynolds numbers typical for engine operation. Good agreement of computed and measured integral flow quantities like overall pressure loss and tumble number is found. A comparison of simulated velocity fields to Particle Image Velocimetry (PIV) measurement data concludes the validation of the enhanced numerical framework for both mean velocity and turbulent fluctuations. The performance of two SFS models, the dynamic Smagorinsky model with Lagrangian averaging along pathlines and the coherent structure model, is tested on different grids. Sensitivity of pressure loss and tumble ratio to the wall treatment and mesh refinement is presented. It is shown that increased wall friction introduced by applying a wall model is overcompensated by some secondary effects, which lead to an overall reduction of pressure loss in the investigated engine geometry. Finally, dynamics of the statistically stationary valve jets are analyzed using Proper Orthogonal Decomposition (POD). Two distinct flow patterns are identified and the relevance for Cycle-to-Cycle Variations (CCV) is discussed.