Slow loading time

[jebej]

OrdinaryDiffEq now takes ~10s to load, mostly taken up by DiffEqBase + ForwardDiff, FiniteDiff.

julia> @time using DiffEqBase
  4.237307 seconds (11.94 M allocations: 688.936 MiB, 3.81% gc time)

julia> @time using ForwardDiff, FiniteDiff
  3.776792 seconds (9.31 M allocations: 472.495 MiB, 1.59% gc time)

Separately, the packages load ~2s faster, meaning there probably is a lot of invalidation happening in DiffEqBase.

julia> @time using DiffEqBase
  4.340304 seconds (11.94 M allocations: 688.958 MiB, 3.80% gc time)

julia> @time using ForwardDiff, FiniteDiff
  1.762233 seconds (4.85 M allocations: 293.860 MiB, 2.32% gc time)

[jebej]

@ChrisRackauckas I can't see where Zygote is being loaded in DiffEqBase, do you mean ZygoteRules?

[ChrisRackauckas]

Separately, the packages load ~2s faster, meaning there probably is a lot of invalidation happening in DiffEqBase.

Or it's the @requires portion.

@ChrisRackauckas I can't see where Zygote is being loaded in DiffEqBase, do you mean ZygoteRules?

Yes, I think it's just ZygoteRules, so it seems like it avoids the known load time issues with Zygote.

[jebej]

I can't seem to identify any one package in DiffEqBase, it just looks like all of them are contributing. Maybe loading the package vs versions would make it easier to troubleshoot. It's also possible that it didn't jump from 5s to 10s and instead that gradually happened over time.

[ChrisRackauckas]

FWIW, I made sure we got a fix into Julia v1.6 which should improve inference and compile times a lot: JuliaLang/julia#35976 so it's worth checking how things are on the nightly.

[ChrisRackauckas]

If it helps a lot, we can beg for it to go into v1.5.

[ChrisRackauckas]

it was like 24 for me on v1.4, and then:

julia> @time using OrdinaryDiffEq
 12.086512 seconds (15.86 M allocations: 928.801 MiB, 2.80% gc time)

on v1.5, so v1.5 is about 2x faster for me!

[ChrisRackauckas]

This has been steadily improving. We can follow it in the other compilation and latency issues.