Help to reduce number of allocations

General Usage
17

There’s also Julia’s built-in performance and memory profiler, if you add this at the end of your script:

using Profile
teste() # run w/o profiling first to ensure compiled & warmed up
@profile teste()
Profile.print(maxdepth=6)

You’ll get something like the below (trimmed a bit):

361 ...usr/share/julia/stdlib/v1.0/Profile/src/Profile.jl:25; top-level scope
  14  allocation_test.jl:9; teste()
    13 ./arraymath.jl:47; +(::Array{Float64,2}, ::Array{Float64,2})
    1  .../usr/share/julia/stdlib/v1.0/Random/src/Random.jl:241; rand
  321 allocation_test.jl:10; teste()
    318 ...are/julia/stdlib/v1.0/LinearAlgebra/src/dense.jl:1296; pinv(::Array{Float64,2})
    3   .../stdlib/v1.0/LinearAlgebra/src/uniformscaling.jl:90; +(::UniformScaling{Bool}, ::Array{Float64,2})
  22  allocation_test:11; teste()
    2 ./arraymath.jl:47; +(::Array{Float64,2}, ::Array{Float64,2})
    6 ./operators.jl:502; *(::Array{Float64,2}, ::Array{Float64,2}, ::Adjoint{Float64,Array{Float64,2}})
    5 ./operators.jl:502; *
    5 ...a/stdlib/v1.0/LinearAlgebra/src/uniformscaling.jl:124; -(::UniformScaling{Bool}, ::Array{Float64,2})

I find this tool very useful. Increase the maxdepth for more detail. There’s also a graphical interface from within Atom if you prefer that.

Then there’s memory allocation tracking. If you start Julia with the flag --track-allocation=user and put this at the end of your script:

using Profile
teste() # run w/o profiling first to ensure compiled & warmed up
Profile.clear_malloc_data()
teste()

Then quit Julia, and locate the newly created .mem file. It will have annotated each line with how much it allocated:

        - function teste()
    50160     P = Vector{Matrix}(undef, 6000)
        -
     2752     P[1] = zeros(18,18)
        -
        0     @inbounds for k = 2:6000
 33018496         Pp   = P[k-1] + rand(18,18)
        0         K    = pinv(I + Pp)
 16597072         P[k] = (I - K)*Pp*(I - K)' + K*I*K'
        -     end
        -
        0     P
        - end

Unfortunately, as you can see above, it’s not very accurate since first of all it shows 0 bytes allocated for the pinv line, and second the total number of bytes allocated is way off (it should be ~357 MB). Not sure what’s going on there, I have the impression that this worked well in Julia 0.6 but hasn’t quite worked since. When/if this works, it is an awesome tool. Apart from the obvious use case of seeing what allocates, you can also use it to detect type instabilities.

1 Like