Mostly, turbulence models, including subgrid-scale (SGS) models and Reynolds stress (and heat flux) models for numerical simulations of turbulent flows perform less well than expected due to the lack of physically meaningful constraints and universality, especially for the separated flows. In this paper, special focus is placed on a constrained SGS (CSGS) model for large-eddy simulation (LES) of separated flow phenomena. In this method, traditional SGS models can be employed in the far-wall region, whereas the mean SGS models are constrained by prescribed external Reynolds quantities (stresses and heat fluxes) in the near-wall region, Several Reynolds stress (and heat flux) models, including Sparlat-Allmaras (SA) model, Menter's k-w Shear Stress Transport (SST) model, k-w SST coupled with Gamma predictive transition model, etc. are evaluated as the model constraints. The CSGS models are compared with frequently used SGS models, such as Smargorinsky model and Wall Adaptive Local-Eddy (WALE) Viscosity model, in simulations of both internal and external flows. It is found that the CSGS models can help eliminate the non-physical Log-layer mismatch phenomenon reported in hybrid RANS-LES methods (e.g., detached-eddy simulation, referred to as DES), and can predict the mean velocity profile, friction force and other statistical quantities more accurately than traditional SGS and hybrid RANS-LES models. Moreover, the practical performance of the CSGS model is closely associated with the Reynolds constraint models rather than the base SGS models.