Performance Considerations for Linear Algebra
Contents
Performance Considerations for Linear Algebra#
BLAS/LAPACK#
After using LinearAlgebra, Julia speaks linear algebra fluently.
Performing linear algebra operations on a computer is, of course, an old problem. Lots of amazing libraries have been written - mostly in Fortran - which have been optimized over decades. Basically all high-level programming languages use these libraries, including R, Python, and Julia.
Linear algebra in Julia is largely implemented by calling BLAS/LAPACK functions. As per default, Julia uses OpenBLAS.
BLAS: a collection of low-level matrix and vector arithmetic operations (“multiply two matrices”, “multiply a matrix by vector”).
LAPACK: a collection of higher-level linear algebra operations. Things like matrix factorizations (LU, LLt, QR, SVD, Schur, etc) that are used to do things like “find the eigenvalues of a matrix”, or “find the singular values of a matrix”, or “solve a linear system”.
(Sparse matrices are more difficult and there exist different collections of routines, one of which is SuiteSparse.)
Why do I have to care?
There are multiple BLAS/LAPACK implementations (e.g. by different vendors) and switching between them can lead to big speedups! 😄
Since you might be leaving the world of Julia code, you loose easy inspectability and type genericity. 😢
Libblastrampoline#
BLAS and LAPACK demuxing library
Allows one to readily switch between BLAS/LAPACK implementations at runtime. Practically, one uses small wrapper packages like
Example: OpenBLAS -> MKL with MKL.jl#
using LinearAlgebra
BLAS.get_config()
LinearAlgebra.BLAS.LBTConfig
Libraries:
└ [ILP64] libopenblas64_.so
BLAS.set_num_threads(4)
using BenchmarkTools
@btime A * B setup = (A = rand(100, 100); B = rand(100, 100));
69.988 μs (2 allocations: 78.17 KiB)
using MKL
BLAS.get_config()
LinearAlgebra.BLAS.LBTConfig
Libraries:
├ [ILP64] libmkl_rt.so└ [ LP64] libmkl_rt.so
BLAS.set_num_threads(4)
@btime A * B setup = (A = rand(100, 100); B = rand(100, 100));
32.586 μs (2 allocations: 78.17 KiB)
Fast linear algebra with multiple dispatch#
N = 100
t = 1
μ = -0.5
H = diagm(0 => fill(μ, N), 1 => fill(-t, N - 1), -1 => fill(-t, N - 1))
100×100 Matrix{Float64}:
-0.5 -1.0 0.0 0.0 0.0 0.0 … 0.0 0.0 0.0 0.0 0.0 0.0
-1.0 -0.5 -1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 -1.0 -0.5 -1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 -1.0 -0.5 -1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 -1.0 -0.5 -1.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 -1.0 -0.5 … 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 -1.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 … 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
⋮ ⋮ ⋱ ⋮
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 … 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 -1.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 -0.5 -1.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 … -1.0 -0.5 -1.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 -1.0 -0.5 -1.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -1.0 -0.5 -1.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -1.0 -0.5 -1.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -1.0 -0.5
f(M, v) = v' * M * v
f (generic function with 1 method)
v = rand(N);
@btime f($H, $v);
2.043 μs (1 allocation: 896 bytes)
typeof(H)
Matrix{Float64} (alias for Array{Float64, 2})
As long as the code is generic (respects the informal AbstractArray interface), we can use the same piece of code for completely different array types. There are a number of special matrix types are available out-of-the-box and the ecosystem offers even more.
Let’s utilize the sparsity of H by indicating it through a type.
using SparseArrays
Hsparse = sparse(H)
100×100 SparseMatrixCSC{Float64, Int64} with 298 stored entries:
⠻⣦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠈⠻⣦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠈⠻⣦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠈⠻⣦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠈⠻⣦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠻⣦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠻⣦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠻⣦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠻⣦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠻⣦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠻⣦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠻⣦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠻⣦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠻⣦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠻⣦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠻⣦⡀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠻⣦⡀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠻⣦⡀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠻⣦⡀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠻⣦
typeof(Hsparse)
SparseMatrixCSC{Float64, Int64}
@btime f($Hsparse, $v);
472.520 ns (1 allocation: 896 bytes)
That’s a solid speedup!
Our H isn’t just sparse, but actually tridiagonal. Let’s try to exploit that.
Htri = Tridiagonal(H)
100×100 Tridiagonal{Float64, Vector{Float64}}:
-0.5 -1.0 ⋅ ⋅ ⋅ ⋅ … ⋅ ⋅ ⋅ ⋅ ⋅ ⋅
-1.0 -0.5 -1.0 ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅
⋅ -1.0 -0.5 -1.0 ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅
⋅ ⋅ -1.0 -0.5 -1.0 ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅
⋅ ⋅ ⋅ -1.0 -0.5 -1.0 ⋅ ⋅ ⋅ ⋅ ⋅ ⋅
⋅ ⋅ ⋅ ⋅ -1.0 -0.5 … ⋅ ⋅ ⋅ ⋅ ⋅ ⋅
⋅ ⋅ ⋅ ⋅ ⋅ -1.0 ⋅ ⋅ ⋅ ⋅ ⋅ ⋅
⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅
⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅
⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅
⋅ ⋅ ⋅ ⋅ ⋅ ⋅ … ⋅ ⋅ ⋅ ⋅ ⋅ ⋅
⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅
⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅
⋮ ⋮ ⋱ ⋮
⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅
⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅
⋅ ⋅ ⋅ ⋅ ⋅ ⋅ … ⋅ ⋅ ⋅ ⋅ ⋅ ⋅
⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅
⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅
⋅ ⋅ ⋅ ⋅ ⋅ ⋅ -1.0 ⋅ ⋅ ⋅ ⋅ ⋅
⋅ ⋅ ⋅ ⋅ ⋅ ⋅ -0.5 -1.0 ⋅ ⋅ ⋅ ⋅
⋅ ⋅ ⋅ ⋅ ⋅ ⋅ … -1.0 -0.5 -1.0 ⋅ ⋅ ⋅
⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ -1.0 -0.5 -1.0 ⋅ ⋅
⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ -1.0 -0.5 -1.0 ⋅
⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ -1.0 -0.5 -1.0
⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ -1.0 -0.5
@btime f($Htri, $v);
343.670 ns (2 allocations: 944 bytes)
Choosing the best type (and therewith an algorithm) can be tricky and one has to play around a bit.
Core messages of this Notebook#
Trying a different BLAS/LAPACK can often give speedups for free.
Indicate properties and structure of a matrix, like hermiticity or sparsity, through types. Fallback to generic types only if you run into method errors.