Execution time prediction for Simulink models

Hi,

Is there a method to predict if a simulink model A consumes more time than a Simulink model B?

In my case, I have many models generated by RTW. I am running these models in many threads. Before starting the simulation, I am trying to predict which model will consume more CPU and then affect it to a special thread.

If we can prove that state variables and zero-crossing events of model A are superior than those of model B, can we say that model A will consume more time than model B? I am not sure that is true...

Another solution is to use the WCET technique (Worst Case Execution Time), but I am searching for a simpler technique.

Thank you,

Best Reguards, Hassen

Answer 1

[Andreas Goser]

That is a very interesting question and I hope many others contribute to it. My first idea is to simulate a model for a brief period and then extrapolate the time.

Example: Model simulation time is 1000s. Simulate the model for the first 10s and multiply the execution time by 100 for a prediction.

Pro: Simple :-)

Cons: Would only be reasonable for fixed-step. Need to estimate the overhead for startingg a simulation. Does not take into account complex behaviour after the "presimulated" time like Stateflow state machines, enabled subsystems etc.

Answer 2

[Hassen]

Thank you Andreas for answering.

I agree with your idea and I tested it. Fortunately, I am using only models with fixed-step solvers.

But, I am searching for something "more intelligent" without doing pre-simulation. At the initialisation of a model, information about zero-crossing events and number of state variables is available. I am not sure but I think that with fixed step solvers, zero-crossing events are disabled. So, may be comparaing the number of state variables could give me an idea about the complexity of the model. Indeed, I suppose that the user of the models (binary models generated with RTW) have no idea about the complexity of each model and also have not the possibility to make pre-simulation.

So here an example showing what I want to do: Let's take 3 models A, B and C. I suppose that the user have only 2 cores. The model A have the biggest execution time so it's assigned to one core, the models B and C are assigned in the other core.

Best Regards, Hassen

Answer 3

[Hassen]

Any other ideas?