Regression: e.g. rand(100) a little slower

[PallHaraldsson]

rand() is a bit faster on Julia 1.2, so that's a good thing, while rand(1) up to rand(100) is slower, so it may be regarding allocation[s], not be about rand?

First Julia 1.2 (yes, my laptop is a bit loaded (with Firefox running) and is very old, can anyone replicate on recent hardware?):

julia> @btime rand(1);
  97.928 ns (1 allocation: 96 bytes)

julia> @btime rand(1);
  98.043 ns (1 allocation: 96 bytes)

julia> @btime rand(1);
  98.409 ns (1 allocation: 96 bytes)

julia> @btime rand(10);
  130.490 ns (1 allocation: 160 bytes)

julia> @btime rand(100);
  579.181 ns (1 allocation: 896 bytes)

julia> @btime rand(100);
  564.536 ns (1 allocation: 896 bytes)

julia> @btime rand();
  7.411 ns (0 allocations: 0 bytes)

julia> @btime rand();
  7.480 ns (0 allocations: 0 bytes)

julia> versioninfo()
Julia Version 1.2.0-DEV.55
Commit 7acb991b98 (2018-12-21 23:33 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: Intel(R) Core(TM)2 Duo CPU     T8100  @ 2.10GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-6.0.1 (ORCJIT, penryn)

Older Julia:

julia> @btime rand();
  11.255 ns (0 allocations: 0 bytes)

julia> @btime rand();
  11.325 ns (0 allocations: 0 bytes)

julia> @btime rand(1);
  78.665 ns (1 allocation: 96 bytes)

julia> @btime rand(1);
  78.214 ns (1 allocation: 96 bytes)

julia> @btime rand(10);
  114.000 ns (1 allocation: 160 bytes)

julia> @btime rand(100);
  489.624 ns (1 allocation: 896 bytes)

julia> @btime rand(100);
  481.225 ns (1 allocation: 896 bytes)

julia> versioninfo()
Julia Version 0.5.2
Commit f4c6c9d4bb (2017-05-06 16:34 UTC)
Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: Intel(R) Core(TM)2 Duo CPU     T8100  @ 2.10GHz
  WORD_SIZE: 64
  BLAS: libopenblas (NO_LAPACKE DYNAMIC_ARCH NO_AFFINITY Penryn)
  LAPACK: liblapack.so.3
  LIBM: libopenlibm
  LLVM: libLLVM-3.7.1 (ORCJIT, penryn)

Then this form is faster for Julia 1.2 (and a bit faster than same for Julia 0.5, so I think I confimed not really about rand):

julia> @btime rand!(rng, a);
  378.034 ns (0 allocations: 0 bytes)

julia> @btime rand!(rng, a);
  374.184 ns (0 allocations: 0 bytes)

julia> @btime rand!(rng, a, 100);  # This worked in Julia 0.5, a is 100-element sized.
ERROR: ArgumentError: Sampler for this object is not defined

[inkydragon]

rand() works fine in my computer @1.0.3&0.7

julia> versioninfo()
Julia Version 1.0.3
Commit 099e826241 (2018-12-18 01:34 UTC)
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: Intel(R) Core(TM) i7-5600U CPU @ 2.60GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-6.0.0 (ORCJIT, broadwell)

julia> @btime rand(1);
  48.178 ns (1 allocation: 96 bytes)

julia> @btime rand(1);
  48.380 ns (1 allocation: 96 bytes)

julia> @btime rand(10);
  61.366 ns (1 allocation: 160 bytes)

julia> @btime rand(100);
  349.239 ns (1 allocation: 896 bytes)

julia> @btime rand(100);
  333.566 ns (1 allocation: 896 bytes)

julia> @btime rand();
  5.300 ns (0 allocations: 0 bytes)

julia> @btime rand();
  5.300 ns (0 allocations: 0 bytes)

julia>

julia> versioninfo()
Julia Version 0.7.0
Commit a4cb80f3ed (2018-08-08 06:46 UTC)
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: Intel(R) Core(TM) i7-5600U CPU @ 2.60GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-6.0.0 (ORCJIT, broadwell)

julia> @btime rand(1);
  91.113 ns (1 allocation: 96 bytes)

julia> @btime rand(1);
  91.488 ns (1 allocation: 96 bytes)

julia> @btime rand(10);
  114.753 ns (1 allocation: 160 bytes)

julia> @btime rand(100);
  355.185 ns (1 allocation: 896 bytes)

julia> @btime rand(100);
  202.945 ns (1 allocation: 896 bytes)

julia> @btime rand();
  10.009 ns (0 allocations: 0 bytes)

julia> @btime rand();
  9.009 ns (0 allocations: 0 bytes)

julia>

[xanfus]

No regression.

Julia Version 0.7.0
Commit a4cb80f (2018-08-08 06:46 UTC)
Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: Intel(R) Xeon(R) CPU E5-2630 0 @ 2.30GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-6.0.0 (ORCJIT, sandybridge)

@btime rand(100);
  352.276 ns (1 allocation: 896 bytes)

Julia Version 1.2.0-DEV.75
Commit 8221c87* (2018-12-31 18:52 UTC)
Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: Intel(R) Xeon(R) CPU E5-2630 0 @ 2.30GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-6.0.1 (ORCJIT, sandybridge)

@btime rand(100);
  247.824 ns (1 allocation: 896 bytes)

rand(1) takes 58-59 ns in both setups. I can't add BenchmarkTools to 0.5.2. I can add it to 1.0+. Why some actions continue to use proxies despite system-wide tweaks performed?

[ViralBShah]

I benchmarked 0.6 and 1.0.3. I note that rand() and rand(100) take about the same time, whereas rand(1) is about 15% faster in 0.6. That suggests that this does not have to do with the RNG, and with small arrays. I have opened #30623 about this slowness, which is demonstrated with a simpler example.