Speed of Convergence faster with smaller time step
 

I have read here many times that convergence is sped up by using larger time steps, assuming the physics can handle it. However, I have noticed that convergence occurs faster and better with smaller time scales in both steady state and transient. This effect has been observed with excellent mesh stats, and simple as well as complex physics. Perhaps speed is measured by actual time to solve, but it seems to be faster in the sense of fewer iterations.
Am I doing something wrong or is this normal?

This is unusual, but occurs occasionally. In steady state, if a simulation has too large time steps it will have troubles converging and a smaller time step will help. This is particularly the case when a simulation is starting up. The best approach is usually to start with small time steps and increase them when it is starting to converge well. You can often end up with very large time steps compared to where you started.

In transient runs you are limited in the size time step you can take as convergence is harder and temporal accuracy suffers. Time step size for transient runs should be set by a sensitivity analysis, but a quick way to find something which is close is to use adaptive time steps to home in on 3-5 coeff loops per iteration.