OrdinaryDiffEq precompilation time is super slow

[anandijain]

like super uber duper mega fantastically slow. anytime anyone updates the teeny tiniest little package, we get the most exciting opportunity of watching paint dry for 10 minutes.

i realized we don't have any issues on this since the 22 seconds -> 3 megaissue, which has regressed so significantly its worth keeping something open on it.

i realize i can have Preferences, but I don't think this is sufficient, since very few "normal" users will do find and do this on their own. IIRC someone was suggesting looking into how to improve parallel precomp of individual solvers by potentially making them their own packages. yingbo was suggesting maybe removing some perform_steps! and have them just hit the fallback

is there a thread i've missed on what we can do given that 1.9 will come out soon?

[ChrisRackauckas]

We're finding concrete things causing the problem, like #1885.

i realize i can have Preferences, but I don't think this is sufficient, since very few "normal" users will do find and do this on their own. IIRC someone was suggesting looking into how to improve parallel precomp of individual solvers by potentially making them their own packages. yingbo was suggesting maybe removing some perform_steps! and have them just hit the fallback

The issue is something to do with inference in the auto-switch algorithm If you open up Cthulhu and help ensure every dynamic dispatch is handled by branching then I think the majority may go away.

And there's JuliaLang/julia#48821 which gives exponential growth in the time to compile things with the integrator type due to type intersection calculations O(2^n). Cody is taking a look at that.