This study compares the conventional isotropic dynamic eddy viscosity model(DEVM) and anisotropy-resolving nonlinear explicit algebraic subgrid-scale(SGS) stress model (EASSM) of Marstorp et al. (J. Fluid Mech., vol. 639,2009, pp. 403–432) in large-eddy simulations (LESs) of flow separation in achannel with streamwise periodic hill-shaped constrictions and spanwise homogeneity(periodic hill flow). The results are validated with well-resolved LESdata of Breuer et al (Computers & Fluids, vol. 38, 2009, pp. 433-457). Threedifferent resolutions ranging from moderate to very coarse are used. LESs arecarried out with the Code Saturne, an unstructured collocated finite volumesolver for incompressible flows with a second-order central difference schemein space and a second-order discretisation in time. It has inherent numericaldissipation due to the low-order of the numerical method. LESs with no SGSmodel (NSM) are also carried out to analyse the influence of the SGS modelsin the presence of discretisation errors. LESs with the NSM show that the inherentnumerical dissipation is sufficient to give a reasonable prediction of themean velocity profiles at the finest resolution. The LES predictions of the meanvelocity and Reynolds stresses with the EASSM are found to be much moreaccurate than the ones with the DEVM at all resolutions. Although the SGSdissipation produced by the EASSM is found to be considerably lower than bythe DEVM, the EASSM predictions show appreciable improvements over theNSM, indicating the importance of the nonlinear part of the model. At thecoarsest resolution, where the SGS anisotropy is large, LES with the EASSMshows a reasonable prediction of the mean separation and reattachment points,whereas LES with the isotropic DEVM predicts a considerably delayed separationand early flow reattachment with a small separation bubble and the LESwith NSM does not display flow separation. At finer resolutions, the DEVMand NSM predict a shorter separation bubble than the EASSM, which has agood agreement with the well-resolved reference LES data. Hence, a correctprediction of the separation and reattachment by LES requires resolving the SGS anisotropy either by a fine grid or by an anisotropy-resolving SGS modelsuch as the EASSM.