Measuring the time properly for nested for-loops with @parallel

General Usage Performance
,
1

Hi, I am trying to measure the total time used to execute a nested for-loops. I used @elapsed but it seems to only return the time used by one of the inner loops. A MWE is as follows:

function f(m,n)
    s = SharedArray{Float64}(m)
    for i = 1:m
        @distributed for j = 1:n 
            s[i] = s[i] + rand(1)[1]
        end
    end 
    return s
end

@elapsed f(500, 1e7)

The above MWE returns 0.000486948, which is not what I need because f(500, 1e7) takes over a minute to run. I tried the TickTock package too and it also has a similar issue.

Therefore, I am wondering what’s the appropriate way to count the time required. In the above MWE, I want to count the time required to run f(500, 1e7), not the time used by one of the loops inside the function f.

Thanks for the help in advance!

2

The macros from BenchmarkTools seem to report the correct times:

julia> using BenchmarkTools

julia> @belapsed f(100,1e6)
4.345676228

julia> @benchmark f(100,1e6)
BechmarkTools.Trial: 2 samples with 1 evaluations.
 Range (min … max):  4.834 s …   4.922 s  β”Š GC (min … max): 5.16% … 5.45%
 Time  (median):     4.878 s              β”Š GC (median):    5.31%
 Time  (mean Β± Οƒ):   4.878 s Β± 62.011 ms  β”Š GC (mean Β± Οƒ):  5.31% Β± 0.21%

  β–ˆ                                                       β–ˆβ–ˆ 
  β–ˆβ–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–β–ˆ ▁
  4.83 s         Histogram: frequency by time        4.92 s <

 Memory estimate: 8.94 GiB, allocs estimate: 100004891.
1 Like
3

This works. Thanks!