Julia : 用@async提升循环性能

@async 用在循环中，可以大幅提升性能，特别是循环量较大时。代码如下：

1、向量

using Distributed

n = 1000000
println("n :",n)
println("iter_no_async :")
@time value1 = iter_no_async(n)
println("iter_async :")
@time value2 = iter_async(n)
@assert value1 ==value2

function iter_no_async(n)
   arr = Array{Float64}(undef,n)
   for i in 1:n
       arr[i] = sin(i^2)
   end
   return arr
end

function iter_async(n)
   arr = Array{Float64}(undef,n)
   @async for i in 1:n
       arr[i] = sin(i^2)
   end
   return arr
end

# 100万 =>9倍
julia> @run testfun
n :1000000     
iter_no_async :
  0.037306 seconds (16.38 k allocations: 8.526 MiB)
iter_async :
  0.004689 seconds (9.04 k allocations: 8.132 MiB)
iter_async (generic function with 1 method)

# 10万=>2倍
julia> @run testfun
n :100000
iter_no_async :
  0.011192 seconds (16.38 k allocations: 1.659 MiB)
iter_async :
  0.004836 seconds (9.04 k allocations: 1.266 MiB)
iter_async (generic function with 1 method)

# 1万 => 略有提升
julia> @run testfun
n :10000
iter_no_async :
  0.007968 seconds (16.38 k allocations: 995.775 KiB)
iter_async :
  0.005080 seconds (9.04 k allocations: 592.965 KiB)
iter_async (generic function with 1 method)

# 1000 =>不明显
julia> @run testfun
n :1000
iter_no_async :
  0.008439 seconds (16.38 k allocations: 925.807 KiB)
iter_async :
  0.007028 seconds (9.03 k allocations: 522.699 KiB)
iter_async (generic function with 1 method)

比如，数组拆分，一个数组要拆分成n个平均的等份，这个时侯，用@async 效果也非常好！性能提升数十倍！

2、矩阵

using Distributed
n = 1000; m = 10
println("n :$n  m:$m")
println("iter_no_async :")

@time value1 = iter_no_async(n,m)

println("iter_async :")
@time value2 = iter_async(n,m)
@assert value1 ==value2

function iter_no_async(n,m)
   arr = Array{Float64}(undef,n,m)
   for i in 1:n
      for j in 1:m
         arr[i,j] = sin(i^2) +sin(j^2)
      end
   end
   return arr
end

function iter_async(n,m)
   arr = Array{Float64}(undef,n,m)
   @async for i in 1:n
      @async for j in 1:m
         arr[i,j] = sin(i^2) +sin(j^2)
      end
   end
   return arr
end

julia> @run testfun
n :1000  m:10  
iter_no_async :
  0.011082 seconds (20.51 k allocations: 1.164 MiB)
iter_async :
  0.005053 seconds (9.26 k allocations: 601.795 KiB)
iter_async (generic function with 1 method)

两层的@async ,性能差不多近2倍。

3、拆分

function split_array(rmap,n)
    responses = Vector{Array{Float64}}(undef, n)
    row,col = size(rmap)
    unit = floor(Int64,row/n)
    @async for i in 1:n
       if i<n
           responses[i] = rmap[(i-1)*unit + 1 : i*unit,:]
        else
           responses[i] = rmap[(i-1)*unit + 1 : row,:]
        end
    end
    return responses
end