Why is the result of the operation between `SharedMatrix`es not a `SharedMatrix`?

Specific Domains Julia at Scale
1

I created two sharedmatrix as the below

sm1 = SharedMatrix{Float64}( rand(1000, 1000) )
sm2 = SharedMatrix{Float64}( rand(1000, 1000) )
sm3 = sm1 * sm2 ---- (1)

However, sm3 is of type Matrix{Float64} not of type SharedMatrix{Float64}

So I have tried like the below

sm3 = SharedMatrix{Float64}(sm1*sm2) ----- (2)

But, this still has a performance problem.

(2) is much slower than (1).

I use version 0.6.2

Does anyone help me? I want to find a way that returns the result of type SharedMatrix and as fast as (1) do.

2 Likes
2

The conversion from SharedArray to Array happens because of the call to BLAS, but I think it should be converted back before returning. However performance-wise, the following 2 functions are very similar:

julia> using BenchmarkTools

julia> function f1(a::SharedMatrix{T}, b::SharedMatrix{S}) where {T, S}
           c = a*b
           return SharedMatrix{promote_type(T,S)}(c)
       end
f1 (generic function with 1 method)

julia> function f2(a::SharedMatrix{T}, b::SharedMatrix{S}) where {T, S}
           c = a*b
           return c
       end
f2 (generic function with 1 method)

julia> sm1 = SharedMatrix{Float64}(rand(1000,1000));

julia> sm2 = SharedMatrix{Float64}(rand(1000,1000));

julia> @btime sm3 = f1($sm1, $sm2);
  86.727 ms (835 allocations: 7.66 MiB)

julia> @btime sm4 = f2($sm1, $sm2);
  85.523 ms (2 allocations: 7.63 MiB)

julia> sm1 = SharedMatrix{Float64}(rand(10000,10000));

julia> sm2 = SharedMatrix{Float64}(rand(10000,10000));

julia> @time sm3 = f1(sm1, sm2);
 51.399829 seconds (813 allocations: 762.967 MiB, 0.14% gc time)

julia> @time sm4 = f2(sm1, sm2);
 50.057509 seconds (6 allocations: 762.940 MiB, 0.12% gc time)

Edit: when you convert, the array is actually being copied so yes I agree it might be a performance problem. I hope someone more knowledgeable will give you a better answer.