Thank you Rob. 

The computation converged with no problem when I used simplec scheme. I have imposed mass flow inlet BC and pressure outlet for outlets. With the coupled scheme, computation converges when I block artificially reverse flow. I don't have mesh actually to show you, sorry. As a solution to the problem, I have refined mesh near outlets and extended outlets but it doesn't work. The mesh is tetrahedral with inflation layers.  The initialization is standard, computation is steady, flow is incompressible, and a turbulence model is K-Omega SST . All spatial schemes are second order.