Calculation with UD differencing scheme yields results that are closer to the experimentally measured velocity profile. However, with MARS scheme the curve behavior is very similar to the one of experimental data. Which scheme would you choose?

Basically, one should not use UD for computations. However, sometimes it is the only choice due to bad convergence. But generally speaking, stick to MARS or other second order scheme !

If your flow doesn't have sharp gradients or the only mesh you can afford is a coarse one, then it's UD. If you wish to dampen out oscillating effects (ie. simulate a transient phenomena as steady state to just get 'bulk' values -with 2nd orders will not converge), then that's another reason to use UD. But don't expect anything other than 'in-the-ballpark' figures for your results.