Performance improvements

[CasBex]

Hi, I was profiling my code and noticed that most of the time was spent inside FMIImport.jl, so I decided to make some performance improvements. Functionality-wise there is no change. The example I was testing this on went from about 60s with 22 GB allocations to 18s with 2.8 GB (which is still not amazing in all honesty).

This is only for FMI2 as I don't use FMI3 myself. Some of these changes require updates to FMICore.jl which is incoming. Before merging, FMICore.jl shoud thus be updated and the Project.toml in this repo should be updated for compatibility.

Summary

• fmi2SetReal checking of the Jacobians in c.jl has been improved (biggest performance gain by far). By making a Set of $\partial$f_refs, it is possible to do lookups in O(1) instead of O(n), which reduces the for loop runtime complexity.
• Various allocations and copies have been removed if possible
• Type stability has been improved where possible
• Added methods to keyword-only performance-critical functions (fmu::FMU2) and (c::FMU2Component) which seem to be faster.

Performance tips from a user perspective

The FMI C-functions (e.g. cFmi2SetReal...) operate on Vector{T} and return this type, regardless of which type of array was provided (where T depends on the C-function). If other types ::AbstractVector{<:Real} are provided, they are converted upon every C-call which takes a lot of time and memory allocations. @ThummeTo Maybe it would be a good idea to dispatch functions instead on Vector{T} and have a single catch-all upper method dispatch on AbstractVector{<:Real} and provide a performance warning for every function?

For optimal performance it is best to preallocate a v::Vector{T} of the correct size and type and reuse this as an intermediate buffer between whatever function you have running and the FMU. (I was personally passing views of arrays to avoid allocation, but this turned out to be counterproductive because of above)

Further improvements

The remaining allocations and slowness are mostly caused by type instability/ambiguity. This happens for example in the FMU2Component which has fields A,B,C,D,E,F::Union{Nothing,FMUJacobian} This means that in every function, Julia has to check at runtime what the type is, even though we know that functions like fmi2SetReal are only called when the Jacobians are already initialised. This causes significant slowdown. This would however require more significant changes to the library than I am personally willing to make (and would be too big for one PR).

[CasBex]

Specifically the changes in c5d5387 and a895080 require this PR in FMICore.jl

[CasBex]

71fd194 to avoid mutating c.x when x is mutated. This will avoid some bugs for users who use often-changing cache-vectors for this.

[ThummeTo]

Much appreciated! Especially typing the structs is something that was on the todo list (for too long I think 😉), and many other performance optimizations too. Thank you very much for the PR. If this is ok, I would review and add some additional modifications after my holiday at the very beginning of september and then merge it. Best regards!

[CasBex]

FYI, I would set the FMICore compat to 18.0 before pushing v16.0 to the general repository

[ThummeTo]

Sure, this will be a breaking release (I will do some more modifications) 👍

[ThummeTo]

Hey @CasBex, just out of curiosity, what benchmark did you use, that shows these improvements? So even if you can not share the simulation model, it would be very interessting to have some information about this, like number of states and event indicators, number of triggered events, implicite/explicite solver and number of solver steps. The model I use for benchmarking the new changes is faster too, but not as much as in your case. Regards!

[CasBex]

It's basically the thermal model in here https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-stg.2019.0196 together with a discrete PID controller and weather data as periodic callbacks (no events in the fmu). The solver model is the default from DifferentialEquations.jl

One thing that will also affect this is that I usually run multiple FMUs at the same time and use my own interface for that combined with FMIImport.jl. In there I also made some improvements, the most significant being the one under "Performance tips from a user perspective", so that is also likely part of the speedup.

[ThummeTo]

Regarding performance: Yep, I just checked this out (and it's also part of the ccall docu) that basically for every arguement in ccall, a unsafe_convert(argtype, cconvert(argtype, argvalue)) is performed, which results in allocations for almost any argument that is used here... I will think about a good solution that doesn't restrict the user interface too much....

[ThummeTo]

Didn't know that the negative impact of using AbstractTypes is that huge! I am currently on restructuring the most important functions, there are speed-ups of around 20x in practice for some FMI function evaluations - just by replacing AbstractArray{T} with Array{T}. I will further add dispatches for AbstractTypes with performance warnings, that's a good and user friendly suggestion.