Open sloede opened 4 years ago
Hard to do much here without some kind of reproducer. My guess is that it is related to https://github.com/JuliaLang/julia/issues/15276
Hi @sloede and @ranocha, I could not find where exactly you mentioned this issue within https://github.com/trixi-framework/Trixi.jl/pull/200, but do you have any updates on this TimerOutputs issue?
PS: I really enjoyed your JuliaCon talk on YouTube.
We couldn't find a really good way to resolve this issue. Thus, we have basically created a copy of @timeit_debug
but without the try
block in
That fixed the severe performance problems we got with the macros implemented in TimerOutputs.jl.
I directly apologize for the convoluted example, but I was not able to reproduce this with a more minimum WE.
When I time this piece of code (version A)
instead of this (version B)
where I did not change anything in comparison to A but wrap the return values in an
Any[]
container instead of returning them directly, I get a huge increase in memory allocations (and thus run time):With return values (version A):
With
out
(version B):If I do not time the call to
calc_blending_factors
(using either version A-type return values or version B), I get roughly the same timings as in B. This does seem very odd to me and I don't have the slightest idea where I should start looking for the problem.(If this is not the correct place to report such behavior, but rather e.g. Discourse, please let me know)