At the moment, when we make a simulation, we traverse the (multiscale-) dependency graph of the models to run the models in the right order. This is fine, but we traverse the graph for every time-step if the models are not parallelizable over time.
This also means that the compiler cannot look into the models and optimize over them.
I don't know the best solution yet, but maybe we could use something like a macro to really "build" the model, i.e. define the calls to the functions, the outputs, etc... once in a new function, and then call this function to make a simulation. So instead of traversing a dependency graph, it would be as if it would be written by hand in a function. This way the compiler could optimize away a lot of things I think. Also we should @inline the models of course.
The benefits in summary:
- no need to traverse the dependency graph every time-step
- enable optimization pass by the compiler
- avoid multiple dispatch on every call to the
run! function when we call the models
We could start testing it by building a model by hand with e.g. 2-3 models in a function VS the same in a model list as usual.
At the moment, when we make a simulation, we traverse the (multiscale-) dependency graph of the models to run the models in the right order. This is fine, but we traverse the graph for every time-step if the models are not parallelizable over time.
This also means that the compiler cannot look into the models and optimize over them.
I don't know the best solution yet, but maybe we could use something like a macro to really "build" the model, i.e. define the calls to the functions, the outputs, etc... once in a new function, and then call this function to make a simulation. So instead of traversing a dependency graph, it would be as if it would be written by hand in a function. This way the compiler could optimize away a lot of things I think. Also we should
@inlinethe models of course.The benefits in summary:
run!function when we call the modelsWe could start testing it by building a model by hand with e.g. 2-3 models in a function VS the same in a model list as usual.