Huge increase in allocations when timing a function with return values

[sloede]

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)

  @timeit timer() "blending factors" begin                                                          
  begin                                                                                             
    alpha, element_ids_dg, element_ids_dgfv = calc_blending_factors(dg, dg.elements.u)              
  end   

instead of this (version B)

  out = Any[]                                                                                       
  @timeit timer() "blending factors" calc_blending_factors(out, dg, dg.elements.u)                  
  alpha, element_ids_dg, element_ids_dgfv = out

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):

     volume integral         415    19.2s  93.8%  46.3ms   8.99GiB  97.4%  22.2MiB
       blended DG-FV         415    18.9s  92.3%  45.6ms   8.87GiB  96.1%  21.9MiB
       blending factors      415    285ms  1.39%   686μs    119MiB  1.26%   294KiB
       pure DG               415   5.49ms  0.03%  13.2μs    396KiB  0.00%     976B

With out (version B):

     volume integral         415    3.26s  75.0%  7.85ms    124MiB  38.5%   307KiB
       blended DG-FV         415    2.99s  68.8%  7.21ms   1.31MiB  0.41%  3.23KiB
       blending factors      415    261ms  6.01%   629μs    119MiB  37.0%   294KiB
       pure DG               415   4.94ms  0.11%  11.9μs    454KiB  0.14%  1.09KiB

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)

[KristofferC]

Hard to do much here without some kind of reproducer. My guess is that it is related to JuliaLang/julia#15276

[jonas-schulze]

Hi @sloede and @ranocha, I could not find where exactly you mentioned this issue within trixi-framework/Trixi.jl#200, but do you have any updates on this TimerOutputs issue?

PS: I really enjoyed your JuliaCon talk on YouTube.

[ranocha]

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

https://github.com/trixi-framework/Trixi.jl/blob/70d365f405c2494e314277f22e63dcc32d04f095/src/auxiliary/auxiliary.jl#L257-L275

That fixed the severe performance problems we got with the macros implemented in TimerOutputs.jl.