@time is off by at least 10x

f1 calls f2 and f3. (this trivial logic will help with parallelization later)

f2 and f3 both mutate struct1 and struct2.

function f1!(int1,int2,struct1,struct2)
 println("\n running f2: ")
   @time f2!(int1,struct1,struct2) 
 println("\n running f3")
   @time f3!(int2,struct1,struct2)

for i in 1:n
println("\n running f1")
@time f1()

This is a sample from one iteration:

 running f1: 
 running f2: 
  0.000006 seconds (30 allocations: 4.688 KiB)
 running f3: 
  0.000005 seconds (1 allocation: 160 bytes)
  0.000170 seconds (111 allocations: 8.297 KiB)

What’s unintuitive here?

Please include all code necessary to reproduce the data (see e.g. Please read: make it easier to help you). Note also that for microbenchmarking you probably want to use the BenchmarkTools.jl package and not @time.

1 Like

I just don’t know how f1 can take 1.7e-4 sec when it’s constituents only take 0.1e-4 sec.

You’re measuring printing also. Depending on the terminal, that can be surprisingly slow. You may also be measuring compilation time.

1 Like

Thanks, I’ll try BenchmarkTools.
I’m not claiming @time is in any way broken, just confused.

I’ll see what I can do as far as an MWE, it’s just that f2 and f3 are both many lines, the structs are very complicated, etc.

Ok, I guess printing could be the culprit. Compilation time unlikely f1 is getting executed repeatedly in a loop.
Sorry, just realized that the loop structure was excluded, now fixed.
Again I’ll see what I can do about adding MWE

As a side note, you probably want to benchmark this instead of measuring the times only once. You can do that with using BenchmarkTools and @btime instead of @time.

1 Like