Benchmarking

[fverdugo]

Hi @raar1

we can start benchmarking the SpMV, which is a key kernel.

this is a sample code:

using PartitionedArrays
using LinearAlgebra
using MPI


function main(distribute)
    n = 100
    nranks = MPI.Comm_size(MPI.COMM_WORLD)
    p = Int((nranks)^(1/3))
    parts_per_dir = (p,p,p)
    p = prod(parts_per_dir)
    ranks = distribute(LinearIndices((p,)))
    nodes_per_dir = (n,n,n)
    args = laplacian_fdm(nodes_per_dir,parts_per_dir,ranks)
    A = psparse(args...) |> fetch
    x = pones(axes(A,2))
    b = similar(x,axes(A,1))
    mul!(b,A,x)
    t = @elapsed for _ in 1:1000
        mul!(b,A,x)
    end
    @show t
end

with_mpi(main)

[fverdugo]

How to benchmark distributed code:

https://partitionedarrays.github.io/PartitionedArrays.jl/stable/usage/#Benchmarking-distributed-codes