Execution performance is difficult to generalize because it is a function of the specific attributes of the simulation.

Hypothetically, if I implemented the same functionality for a GUI using the same methods described above, I would expect that using the SET_PARAM API would have the least impact on simulation execution time.

The reason I believe this is because Simulink does not need to call the MATLAB Interpreter throughout the execution of the simulation. Instead the exchange between Simulink and the MATLAB Interpreter is much more limited. But this also limits you ability to do things like plot signals in the GUI during the course of simulation.

Both the S-function approach and attaching event listeners to the Simulink model require that the MATLAB Interpreter be invoked to execute MATLAB Code throughout the course of the Simulink execution.

This exchange between the Simulink engine and the MATLAB Interpreter can slow things down.

But as with any question about performance, the question really becomes “what is fast enough?” While implementing a given GUI functionality using the three methods described above, each implementation might produce differing performance results. But would you notice the difference? Maybe, maybe not. Again we end up back to the difficulty of generalizing questions regarding performance.