A package for dense and sparse distributed linear algebra and optimization. The underlying functionality is provided by the C++ library Elemental written originally by Jack Poulson and now maintained by LLNL.
The package is installed with Pkg.add("Elemental"). For Julia versions 1.3 and later, Elemental uses the binaries provided by BinaryBuilder, which are linked against the MPI (mpich) provided through BinaryBuilder.
Each of these examples should be run in a separate Julia session.
This example runs on a single processor, and initializes MPI under the hood. However, explicit use of MPI.jl is not required in this case, compared to the other examples below.
julia> using LinearAlgebra, Elemental
julia> A = Elemental.Matrix(Float64)
0x0 Elemental.Matrix{Float64}
julia> Elemental.gaussian!(A, 100, 80);
julia> U, s, V = svd(A);
julia> convert(Matrix{Float64}, s)[1:10]
10-element Array{Float64,1}:
19.8989
18.2702
17.3665
17.0475
16.4513
16.3197
16.0989
15.8353
15.5947
15.5079In this example, @mpi_do has to be used to send the parallel instructions to all processors.
julia> using MPI, MPIClusterManagers, Distributed
julia> man = MPIManager(np = 4);
julia> addprocs(man);
julia> @everywhere using LinearAlgebra, Elemental
julia> @mpi_do man A = Elemental.DistMatrix(Float64);
julia> @mpi_do man Elemental.gaussian!(A, 1000, 800);
julia> @mpi_do man U, s, V = svd(A);
julia> @mpi_do man println(s[1])
From worker 5: 59.639990420817696
From worker 4: 59.639990420817696
From worker 2: 59.639990420817696
From worker 3: 59.639990420817696This example is slightly different from the ones above in that it only calculates the singular values. However,
it uses the DistributedArrays.jl package, and has a single thread of control. Note, we do not need to use @mpi_do
explicitly in this case.
julia> using MPI, MPIClusterManagers, Distributed
julia> man = MPIManager(np = 4);
julia> addprocs(man);
julia> using DistributedArrays, Elemental
julia> A = drandn(1000, 800);
julia> Elemental.svdvals(A)[1:5]
5-element SubArray{Float64,1,DistributedArrays.DArray{Float64,2,Array{Float64,2}},Tuple{UnitRange{Int64}},0}:
59.4649
59.1984
59.0309
58.7178
58.389The iterative SVD algorithm is implemented in pure Julia, but the factorized matrix as well as the Lanczos vectors are stored as distributed matrices in Elemental. Notice, that TSVD.jl doesn't depend on Elemental and is only using Elemental.jl through generic function calls.
julia> using MPI, MPIClusterManagers, Distributed
julia> man = MPIManager(np = 4);
julia> addprocs(man);
julia> @mpi_do man using Elemental, TSVD, Random
julia> @mpi_do man A = Elemental.DistMatrix(Float64);
julia> @mpi_do man Elemental.gaussian!(A, 5000, 2000);
julia> @mpi_do man Random.seed!(123) # to avoid different initial vectors on the workers
julia> @mpi_do man r = tsvd(A, 5);
julia> @mpi_do man println(r[2][1:5])
From worker 3: [1069.6059089732858,115.44260091060129,115.08319164529792,114.87007788947226,114.48092348847719]
From worker 5: [1069.6059089732858,115.44260091060129,115.08319164529792,114.87007788947226,114.48092348847719]
From worker 2: [1069.6059089732858,115.44260091060129,115.08319164529792,114.87007788947226,114.48092348847719]
From worker 4: [1069.6059089732858,115.44260091060129,115.08319164529792,114.87007788947226,114.48092348847719]@mpi_do man A = Elemental.DistMatrix(Float32)
@mpi_do man B = Elemental.DistMatrix(Float32)
@mpi_do man copyto!(A, Float32[2 1; 1 2])
@mpi_do man copyto!(B, Float32[4, 5])Run distributed ridge regression ½|A*X-B|₂² + λ|X|₂²
@mpi_do man X = Elemental.ridge(A, B, 0f0)Run distributed lasso regression ½|A*X-B|₂² + λ|X|₁ (only supported in recent versions of Elemental)
@mpi_do man X = Elemental.bpdn(A, B, 0.1f0)Right now, the best way to see if a specific function is available, is to look through the source code. We are looking for help to prepare Documenter.jl based documentation for this package, and also to add more functionality from the Elemental library.