SharedArray setindex! slower than for Array

[mauro3]

Running this script:

@time 1
function f2(a)
    out = 0
    for i = eachindex(a)
        a[i] = out
    end
    a
end

n = 10^8
a = ones(Int, n)
f2([1])
@time f2(a)

a = convert(SharedArray, [1])
f2(a)
a = convert(SharedArray, ones(Int,n))
@time f2(a)

I get:

  0.000003 seconds (149 allocations: 9.013 KB)        <--- warm-up @time
  0.082624 seconds (4 allocations: 160 bytes)           <--- Array
  0.129632 seconds (4 allocations: 160 bytes)           <--- SharedArray

I don't know whether this is expected/fixable. If not, please close.

timings edited for clarity - Amit

[amitmurthy]

I would guess the shmem segments have additional checks at the OS level compared to regular process memory. In context, the timings are over 100 million iterations - probably not much we can do here.

[yuyichao]

I don't think the OS does anything about shared segments (it will be super expensive). I believe the issue is the a combination of #15369 and SharedArray being a mutable type (so some load of fields needs to be rooted).

[simonster]

Another issue is that the field that holds the underlying Array object is undefined at SharedArray construction, so accessing that field requires an undefined check.