Large Eddy Simulation (LES) is being increasingly used for turbulent reacting flows in diesel engines. The accuracy of LES depends strongly on the accuracy of the subgrid-scale models, which are employed. In this study, two-dimensional DNS of autoigniting turbulent mixing layers are performed at temperature and pressure conditions similar to those observed in diesel engines. The DNS results are used to evaluate the performance of two common subgrid-scale combustion models that are employed in diesel combustion modeling - the perfectly stirred reactor (PSR) model and the unsteady flamelet progress variable (UFPV) model. It is shown that the effect of turbulence on the reaction zone is not felt during the early ignition stage, and in this stage, the PSR model performance is superior to that of the UFPV model. For a laminar mixing layer, which evolves under the influence of turbulence into a turbulent mixing layer, the PSR model performance deteriorates drastically with increasing time, and the UFPV model is shown to be a more suitable subgrid-scale model. The choice of the subgrid-scale combustion model is also shown to strongly depend on the filter size, i.e. the LES grid size. For filter sizes smaller than the mixing layer thickness, the PSR model performs reasonably well whereas the UFPV model outperforms the PSR model for larger filter sizes. (C) 2015 Elsevier Ltd. All rights reserved.