In my opinion, if you use variable step solvers then no. If you use fixed step solvers, or consider fixed step discrete systems usually one chooses time step in a way that is the divisor of delay value. However I have seen results for discrete systems with delays non-commesurate with time step, however I'm not sure about their practicality.

Thanks Jerzy for your response. It is well taken.

'Time-step' is an abstract notion that SD theoreticians carried over from control engineering to ensure that models did not mis-behave. If your time-units are too long - years, say, in a business analysis - the fact that parameters remain fixed over each time-unit means that you are stuck with that constant value for a long period when it should [in real-world terms] have changed. This not only means you get clunky results as the model jumps for each time-unit, but also means the results get far, far away from reality. Breaking the time-units into fractions [1/2, 1/4 etc] makes recalculation happen more often, so the clunkiness disappears and results get ever-closer to matching reality. 

Unfortunately, time-step or dT as it is known from integral calculus, is a meaningless notion in the real-world, so a preferable approach is to build the model in the first place with short-enough time-units that the clunkiness is no worse than in the real-world data. For example, many business values are reported no more frequently than monthly, so the error of having a monthly model with a time-step of 1.0 is no worse than the uncertainty in the data. Then, you get the advantage that the numbers you produce are instantly recognisable by the people involved in the case, and your delays can be realistic numbers of actual time-units -- it takes 3 months to hire people, 1 month to receive payment for sales, 24 months to install new capacity, and so on.  

Thanks Professor Warren for delailed and enlightening response. I met you twice this year at SD events Tokyo and Delft.

I think the issue concerns logical time and chronological time and the degree of representativeness of the model when the two time achieves a closer convergence. Time that exist in the world of the model (of the dynamic system) which may be called logical time. And time that exist in the real world (of which the simulation is a representation) is chronological time. The conversion between logical and chronological time is not smooth and has to be discretized. Even though the level of granularity of the discretization (requires some calibration on the part of the modeler) my improve the degree of representation of the simulation and may be improved by empirical evidence regarding chronological time.