Benchmarking questions

General Usage
1

Dear all, I was testing some acceleration of for loops (and reducing the allocations), and I made a less complex example for testing:

Julia>julia> function test_1(a,b)
           @views for i in 1:a, j in 1:b
               c = i + j
           end
       end
test_1 (generic function with 1 method)
julia> @benchmark test_1(5,5)
BenchmarkTools.Trial: 10000 samples with 1000 evaluations.
 Range (min … max):  1.500 ns … 8.291 ns  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     1.583 ns             ┊ GC (median):    0.00%
 Time  (mean ± σ):   1.566 ns ± 0.084 ns  ┊ GC (mean ± σ):  0.00% ± 0.00%

                     ▇                 █▁
  ▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁██▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁██▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▂ ▂
  1.5 ns         Histogram: frequency by time       1.62 ns <

 Memory estimate: 0 bytes, allocs estimate: 0.

julia> function test_2(a,b)
           for i in 1:a, j in 1:b
               c = i + j
           end
       end
test_2 (generic function with 1 method)

julia> @benchmark test_2(5,5)
BenchmarkTools.Trial: 10000 samples with 1000 evaluations.
 Range (min … max):  1.500 ns … 6.792 ns  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     1.583 ns             ┊ GC (median):    0.00%
 Time  (mean ± σ):   1.566 ns ± 0.076 ns  ┊ GC (mean ± σ):  0.00% ± 0.00%

                     ▇                 █▁
  ▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁██▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁██▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▂ ▂
  1.5 ns         Histogram: frequency by time       1.62 ns <

 Memory estimate: 0 bytes, allocs estimate: 0.

Shouldn’t I expect the function test_1 to be faster due to the usage of @views?

2

@views only makes things faster when you are actually allocating any slices. There aren’t any here.

3

Just to expand on what’s above, note that scalars don’t allocate.

So if you access a variable holding a number or a single element of a vector (for example, x=[1,2,3] and you access x[1]), then you don’t need views.

You’d need a view if you have a slice with two or more elements, such as x[1:2] .

4

Note also that your benchmarks are incorrect since Julia optimizes your loops away because your functions don’t return anything, and so independently by the magnitude of a and b it takes a small amount of ns:


julia> @benchmark test_1(1000000000,10000000000)
BenchmarkTools.Trial: 10000 samples with 1000 evaluations.
 Range (min … max):  4.700 ns … 73.700 ns  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     5.700 ns              ┊ GC (median):    0.00%
 Time  (mean ± σ):   5.514 ns ±  1.185 ns  ┊ GC (mean ± σ):  0.00% ± 0.00%

  ▅ █       ▁ ▂  ▂ ▃  ▃ ▄  ▆ █  ▇ ▅  ▄ ▃  ▁                  ▂
  █▁█▁▁▅▁█▁▁█▁█▁▁█▁█▁▁█▁█▁▁█▁█▁▁█▁█▁▁█▁█▁▁█▁█▁▁▇▁▇▁▁▆▁▇▁▁█▁█ █
  4.7 ns       Histogram: log(frequency) by time        7 ns <

 Memory estimate: 0 bytes, allocs estimate: 0.
1 Like