Why does rand() in threads slowdown speed in Julia 1.3

New to Julia
1

In Julia 1.3 each thread has its own default global RNG. Calling rand() in each thread is safe. So the following code that calculates pi returns correct result:

function par_pi(n::Int)
    hits = zeros(Int, nthreads())
    @threads for i in 1:n
        x, y = rand(), rand()
        hits[threadid()] += (x^2 + y^2 <= 1)
    end

    4.0 * sum(hits) / n
end
@time par_pi(10_000_000)

But the above code runs 4x slower than the following version that uses manually allocated RNGs:

const threadsRNG = [MersenneTwister() for i in 1:nthreads()]

function par_pi2(n::Int)
    hits = zeros(Int, nthreads())
    @threads for i in 1:n
        rng = threadsRNG[threadid()]
        x, y = rand(rng), rand(rng)
        hits[threadid()] += (x^2 + y^2 <= 1)
    end

    4.0 * sum(hits) / n
end
@time par_pi2(10_000_000)

How to explain it?

1 Like
2

@btime instead of @time?

3

@jling, thanks for replying.
I used @btime and the results as follows:

25.503 ms (59 allocations: 7.00 KiB) with new RNGs
102.228 ms (59 allocations: 7.00 KiB) with default RNGs

4

I can confirm your finding

5

I see the similar results with the latest nightly. It’s strange because @code_lowered is basically the same for both functions. Maybe worth filing a bug report.

julia> @btime par_pi(10^7)
  81.293 ms (44 allocations: 5.89 KiB)
3.140466

julia> @btime par_pi2(10^7)
  27.520 ms (44 allocations: 5.89 KiB)
3.14193

julia> versioninfo()
Julia Version 1.4.0-DEV.670
Commit 032dbe33e3 (2019-12-29 22:39 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: Intel(R) Core(TM) i7-7700 CPU @ 3.60GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-8.0.1 (ORCJIT, skylake)
Environment:
  JULIA_NUM_THREADS = 8

EDIT:
The difference in the lowered code is the following line. I’m not sure how to read it. Do the numbers indicate a different version of the function?
Main.:(var"#19#threadsfor_fun#4")
Main.:(var"#2#threadsfor_fun#3")

7

it’s just the location of function in your REPL

8

@robsmith11, thank you for confirming the result. Indeed it is strange; Two lowered code are same except for the generated variable names. I will file a bug report.

1 Like
9

I’ve filed a bug report:

https://github.com/JuliaLang/julia/issues/34216

2 Likes