Reduce CLI latency

[visr]

In a REPL the trivial test model runs in about 0.03 seconds on the second run, but 20 seconds with the CLI. We put a lot of effort into making the solver fast to run. Little effort has been spent to quickly start a simulation. Users running the trivial test model from the CLI (v2024.11) need to wait 20 seconds each time. For short simulations or quick tests this hinders the iteration speed of modelers. Here is the output of hyperfine on trivial:

❯ hyperfine 'ribasim ribasim.toml'
Benchmark 1: ribasim ribasim.toml
  Time (mean ± σ):     20.835 s ±  0.280 s    [User: 10.951 s, System: 0.364 s]
  Range (min … max):   20.606 s … 21.525 s    10 runs

Especially on a warm filesystem cache the loading of the binary is quick, Info: Starting a Ribasim simulation appears quickly. It then takes about 10 seconds for Simulating 0% to appear, and another 10 seconds for the simulation to finish.

That indicates that it still needs to compile a lot of code during a CLI run. Ten seconds worth of initialization / validation code and ten seconds of simulation code. I hoped that https://github.com/Deltares/Ribasim/blob/v2024.11.0/build/precompile.jl would precompile more of what we need, but apparently it doesn't. We can consider looking into this with SnoopCompile.jl. Right now we don't yet use PrecompileTools.jl in Ribasim.jl. We could start doing this as well.

I assume in time having small binaries with juliac will also help here.

One way to completely avoid this issue is to run Ribasim.main("ribasim.toml") from Julia instead. If we'd support some way in the CLI to run multiple model (separate TOMLs) with one command this would similarly avoid the latency when doing multiple runs each time. It may be good to support separate TOMLs also for #47.

[visr]

If we don't use the precompile file in PackageCompiler, the build is 6 minutes faster and only 6 MB smaller (than 313 MB) but the latency is about twice as high.

Testing a simple PrecompileTools workload like

using PrecompileTools: @compile_workload
@compile_workload begin
    main(normpath(@__DIR__, "../../generated_testmodels/basic/ribasim.toml"))
end

seems to suffer from similar issues. It still compiles a lot of code when running other models.

The compile_workload did cause world age errors related to the generated function from #1982. Here is the written out version that we should consider switching to:

function config.snake_case(nt::NodeType.T)::Symbol
    if nt == NodeType.Basin
        return :basin
    elseif nt == NodeType.TabulatedRatingCurve
        return :tabulated_rating_curve
    elseif nt == NodeType.Pump
        return :pump
    elseif nt == NodeType.Outlet
        return :outlet
    elseif nt == NodeType.UserDemand
        return :user_demand
    elseif nt == NodeType.FlowDemand
        return :flow_demand
    elseif nt == NodeType.LevelDemand
        return :level_demand
    elseif nt == NodeType.FlowBoundary
        return :flow_boundary
    elseif nt == NodeType.LevelBoundary
        return :level_boundary
    elseif nt == NodeType.LinearResistance
        return :linear_resistance
    elseif nt == NodeType.ManningResistance
        return :manning_resistance
    elseif nt == NodeType.Terminal
        return :terminal
    elseif nt == NodeType.DiscreteControl
        return :discrete_control
    elseif nt == NodeType.ContinuousControl
        return :continuous_control
    elseif nt == NodeType.PidControl
        return :pid_control
    else
        error("Unknown node type: $nt")
    end
end

[visr]

I think #2091 would help here. Although benchmarking #2077 did not seem to make much of a difference, so it is probably not enough. Rather than focusing on precompilation we should first look why, after running test model A from a REPL, it still needs to compile so much code when running test model B after that. (Running A again after is fast) What code does it need to compile? ComponentArrays is one of them, but not sure what the situation is with OrdinaryDiffEq and JuMP. We probably need to use SnoopCompile.

https://docs.sciml.ai/DiffEqDocs/latest/features/low_dep/

[visr]

We can also try using --trace-compile=file.jl --trace-compile-timing on 1.12 (the 1.12-nightly juliaup channel).