feat: add PETScSNES
#482
feat: add PETScSNES
#482
Conversation
|
This 150 lines of code https://github.com/JuliaParallel/PETSc.jl/blob/main/examples/SNES_ex2.jl translates to 30 lines in nonlinearsolve with automatic sparsity detection # This implements src/snes/examples/tutorials/ex2.c from PETSc using the PETSc.jl package, using SNES
#
# This solves the equations sequentially
#
# Newton method to solve u'' + u^{2} = f, sequentially.
using NonlinearSolve, PETSc, LinearAlgebra, SparseConnectivityTracer
n = 21
u0 = fill(0.5, n)
function form_residual!(resid, x, _)
n = length(x)
xp = LinRange(0.0, 1.0, n)
F = 6xp .+ (xp .+ 1e-12) .^ 6
dx = 1 / (n - 1)
resid[1] = x[1]
for i in 2:(n - 1)
resid[i] = (x[i - 1] - 2x[i] + x[i + 1]) / dx^2 + x[i] * x[i] - F[i]
end
resid[n] = x[n] - 1
return
end
nlfunc = NonlinearFunction{true}(form_residual!; sparsity = TracerSparsityDetector())
nlprob = NonlinearProblem(nlfunc, u0)
solve(nlprob, NewtonRaphson())
solve(nlprob, PETScSNES()) |
|
PETSc is faster than us in the sparse case at least on my local computer. Once all the refactoring is done we will have to revisit the benchmarks |
|
Looking at the profile all of the time goes in the UMFPACK |
docs/src/tutorials/snes_ex2.md
Outdated
| ## Runtimes | ||
|
|
||
| ### Dense Jacobian | ||
|
|
||
| ```@example snes_ex2 | ||
| @benchmark solve($(nlprob_dense), $(NewtonRaphson()); abstol = 1e-8) | ||
| nothing # hide | ||
| ``` | ||
|
|
||
| ```@example snes_ex2 | ||
| @benchmark solve($(nlprob_dense), $(PETScSNES()); abstol = 1e-8) | ||
| nothing # hide | ||
| ``` | ||
|
|
||
| ### Sparse Jacobian | ||
|
|
||
| ```@example snes_ex2 | ||
| @benchmark solve($(nlprob_sparse), $(NewtonRaphson()); abstol = 1e-8) | ||
| nothing # hide | ||
| ``` | ||
|
|
||
| ```@example snes_ex2 | ||
| @benchmark solve($(nlprob_sparse), $(PETScSNES()); abstol = 1e-8) | ||
| nothing # hide |
There was a problem hiding this comment.
What's the result? Just curious
There was a problem hiding this comment.
not by much on CI. On my laptop it is 300us vs 260us (PETSc being faster).
There was a problem hiding this comment.
How optimized is our sparse solver selection for LinearSolve?
There was a problem hiding this comment.
Relatively optimal? What sparse linear solver is PETSc using here? Try all of UMFPACK, KLU, and MKLPardiso. Other thing to try is using MKL with Pardiso. Then note that there's a new sparse solver ParU that @rayegun is getting wrapped SciML/LinearSolve.jl#394.
One thing to double check is what the extra time is spent in. Is it only doing one symbolic factorization?
There was a problem hiding this comment.
Let's get this in, since it is non-blocking. I anyways need to redo benchmarks for the paper, so I can revisit it after the splitting is done.
PETSc0.2 resolves old compatibility issues we were having with wrapping this library.Warning
I am still seeing occasional segfaults with PETSc. Very common in 1.11. And for some reason PETSc hijacks
Ctrl + Cand trying to use it leads to a segfault even when no code is runningfixes #160
TODOs
PETSc.jl