Time to first execution (TTFX)

[maxchendt]

on Julia 1.9 rc1, there still needs 13 seconds for first run, and 0.3 seconds for second run, for one of my dataset.

I find that the package EfficientFrontier cut down from 7.89 seconds ( Julia 1.8.5) to 0.13 seconds (1.9 rc1). Lines like Base.precompile(Tuple{...}) may be helpful

[guilhermebodin]

Just as a reference, there is a big discussion about this in this issue jump-dev/MathOptInterface.jl#2226. They have made some precompilation statements for some solvers as a test. Maybe you could use some of the information even if not using the MOI interface.

[goulart-paul]

I made an initial attempt at this in pg/precompile, just using SnoopPrecompile on the solver's direct interface (i.e. not via MOI).

It doesn't seem to have made any improvement at all, even on julia v1.9, but I think I must be doing some really basic thing wrong. I will revisit once I have had a closer look at the issue suggested by @guilhermebodin.

[maxchendt]

How about

__precompile__()

.
.
.


precompile(Tuple{Type{Solver{Float64}}})

precompile(Tuple{typeof(Core.kwcall), NamedTuple{(:verbose,), Tuple{Bool}}, Type{Settings{Float64}}})

precompile(Tuple{typeof(setup!), Solver{Float64}, SparseArrays.SparseMatrixCSC{Float64, Int64}, Array{Float64, 1}, SparseArrays.SparseMatrixCSC{Float64, Int64}, Array{Float64, 1}, Array{SupportedCone, 1}, Settings{Float64}})

precompile(Tuple{typeof(solve!), Solver{Float64}})

[goulart-paul]

How about

__precompile__()
... etc

That also doesn't seem to improve things, and @time using Clarabel at launch (following a previously precompilation and restart) still takes about 5 seconds. It also seems to have increased the precompilation time quite a lot with no real payoff.

[maxchendt]

May be you should add things like

Base.precompile(Tuple{typeof(computeCL!), Vector{sCL{Float64}}, Vector{Status}, Problem{Float64}, Settings{Float64}})

where the computeCL! is the core of the computation, but not directly expose to users.

I do not know what are the working horses of your functions.

[goulart-paul]

jump-dev/JuMP.jl#122 has TTFX for the native solver interface (i.e. not MOI / JuMP) down to < 25ms. Package load times are also now a lot shorter.

NB: testing described on that PR is all on julia 1.9.0-rc1 and the master branches of MOI and JuMP.

[goulart-paul]

This should be fixed in v0.5.0, released today. I will close the issue but please reopen if there is still a problem.

For greatest improvement in load times you need julia v1.9 or higher. It will still work with earlier versions though.