Most biomolecular force fields are well-suited for studying static properties of equilibrium structures; however, they have hardly been tested on very-long-time dynamics. This is an important area of research, because computers shortly will be fast enough for quantitative computational studies of folding and stability of small protein domains. In this work, we present an extensive series of 28 50-ns particle-mesh Ewald simulations (a total of 1.4 mus of data) of the Villin headpiece in water, to compare the results of the OPLS-AA/L and GROMOS96 force fields, the influence of water models, protonation states, and the effect of using virtual "dummy" particles for H atoms to increase the simulation time step. In addition to normal properties such as the Calpha root-mean-square deviation, hydrogen bonding, and secondary structure, we also calculate the NMR distance restraints, J-couplings, and chemical shifts of Villin from the simulation and compare directly with the experimental data. We argue that this comparison of simulated and experimental ensembles of structures is a much more accurate tool than the classical comparison of average or momentaneous protein structures. For the Villin system simulated, we find that distance restraint violations in the simulations are comparable to, or even lower than, those from the ensemble obtained by distance geometry. Furthermore, we find that the OPLS-AA/L force field works considerably better with the TIP4P water model than with the simple point charge model. In further simulations, we have tested the effect of the protonation state of the two Glutamate residues, because the structure was determined at pH 3.7. We do find slightly better agreement of the simulations with experimental data when the Glutamate residues are protonated. Finally, we have used multiple different starting structures for most simulations and find that the results are very similar.