Second (vs. third) "using" issue

[PallHaraldsson]

Nobody has mentioned second plot issues (as in mbaz/Gaston.jl#133), that I recall, or for using as I discovered with Plots.jl (also confirmed in recent but not most recent Julia 1.5):

Longer text on here: https://discourse.julialang.org/t/using-plots-takes-20-seconds-or-more/7061/17?u=palli

What can be done to get minimal "allocations" on second use?

$ julia-1.4.0/bin/julia 

julia> @time using Plots
 10.637071 seconds (8.87 M allocations: 494.417 MiB, 2.51% gc time)

julia> @time using Plots
  1.475938 seconds (1.72 M allocations: 81.089 MiB, 3.81% gc time)

julia> @time using Plots
  0.000227 seconds (246 allocations: 13.312 KiB)

Only mitigated by -O0 but by about 50%:

$ time julia-1.4.0/bin/julia -O0

julia> @time using Plots
  9.048337 seconds (8.87 M allocations: 493.852 MiB, 2.98% gc time)

julia> @time using Plots
  0.772250 seconds (1.72 M allocations: 81.089 MiB, 4.60% gc time)

julia> @time using Plots
  0.000295 seconds (246 allocations: 13.312 KiB)

julia> @time plot(1:3)
  4.354561 seconds (10.44 M allocations: 517.449 MiB, 5.26% gc time)

julia> @time plot(1:3)
  0.001798 seconds (9.85 k allocations: 284.570 KiB)

julia> @time plot(1:3)
  0.002011 seconds (9.85 k allocations: 284.570 KiB)

It seems I ruled out issue with Julia itself or @time.

[isentropic]

I'm just curious but why do you want to using multiple times?

[isentropic]

This happens to other packages too, I cheked DifferentialEquations for example. Do you think that this issue is Plots.jl specific?

[PallHaraldsson]

It's not Plots.jl specific. I mean, yes, Plots does do something specific, so that this happens for it, but such doesn't happen for all packages (why I ruled out Julia itself). E.g. not for RandomNumbers.jl but does for VectorizedRNG.jl.

I wouldn't do using more than once on purpose (except here for testing), but in general I could see it happening in (large) programs. E.g. you include a package that also does it.

While the number of allocations, and amount of bytes is larger, the speed (of second using) is a bit better on my recent Julia 1.5:

julia> @time using Plots
 10.756484 seconds (18.60 M allocations: 937.382 MiB, 3.84% gc time)

julia> @time using Plots
  0.543384 seconds (1.01 M allocations: 46.835 MiB, 2.70% gc time)

julia> @time using Plots
  0.000324 seconds (366 allocations: 19.484 KiB)

$ julia -O0

julia> @time using Plots  # with -O1 almost as fast, so possibly a better default
  8.539219 seconds (18.60 M allocations: 937.408 MiB, 6.61% gc time)

julia> @time using Plots
  0.288531 seconds (1.01 M allocations: 46.835 MiB, 3.79% gc time)

julia> @time using Plots
  0.000327 seconds (366 allocations: 19.484 KiB)


# since non-default optimization level helps, what does the (very slow) precompilation actually do? I suppose it doesn't fully compile, but what exactly is left out? If the optimization, then above should be equal.

julia> @time using Plots
[ Info: Precompiling Plots [91a5bcdd-55d7-5caf-9e0b-520d859cae80]
 88.142653 seconds (18.67 M allocations: 947.484 MiB, 0.58% gc time)

[isentropic]

Let's keep this open, but I do not expect this to be investigated as it perhaps is related to Julia then. I do not think that packages small or large do something wrong that causes this issue.

[BeastyBlacksmith]

Some insight can be gained from #2676