Large eddy simulation of spray combustion in an HSDI engine is carried out in this thesis. The implementation of the code was performed in logical steps that allowed both assessment of the performance of the existing KIVA-LES and later development. The analysis of the liquid annular jet confirmed existence of typical, annular jet exclusive structures like head vortices, stagnation point and recirculation in the inner zone. The influence of the swirl in the ambient domain was found to have profound impact on the development, penetration and radial spreading of the jet. Detailed results were reported in Jagus et al. (2008). The code was further validated by performing an extensive study of large eddy simulation of diesel fuel mixing in an engine environment. The reaction models were switched off in order to isolate the effects of both swirl and the different numerical treatment of LES. Reference RANS simulation was performed and significant differences were found. LES was found to capture much better the influence of the swirl on the liquid and vapour jets, a feature essentially absent in RANS results. Moreover, the predicted penetration of the liquid was much higher for the LES case and more in accordance with experimental measurements. Liquid penetration and subsequent evaporation are very important for prediction of heat release rates and encouraging results formed a good basis to performing a full simulation with models for ignition and combustion employed. The findings were analyzed in the paper by Jagus et al. (2009). Further modifications were introduced into the LES code, among them changes to the combustion model that was originally developed for RANS and calculation of the filter width. A new way of estimating the turbulent timescale (eddy turnover time) assured that the incompatibilities in the numerical treatment were eliminated and benefits of LES maximized. The new combustion model proved to give a very good agreement with experimental data, especially with regard to pressure and accumulated heat release rates. Both qualitative and quantitative results presented a significant improvement with respect to RANS results and old LES formulation. The new LES model was proved to give a very good performance on a spectrum of mesh resolutions. Encouraging results were obtained on a coarse mesh sets therefore proving that the new LES code is able to give good prediction even on mesh sizes more suitable for RANS. Overall, LES was found to be a worthy alternative to the well established RANS methods, surpassing it in many areas, such as liquid penetration prediction, temperature-turbulence coupling and prediction of volume-averaged data. It was also discovered that the improved LES code is capable of producing very good results on under-resolved mesh resolutions, a feature that is especially important in industrial applications and on serial code structure.