Hi,
please consider the attached script. Following this
<https://bitbucket.org/fenics-project/dolfin/pull-request/2/use-cholesky-rather-than-lu-decomposition/diff#chg-dolfin/fem/LinearVariationalSolver.cpp>
discussion, if method is mumps, petsc or pastix and
we have symmetric=True the linear system is solved with Cholesky
factorization (it this so?). While testing different method/symmetry
combinations I noticed that PETSc's own symmetric solver is easily
10 times slower then mumps (I don't have pastix to compare against). Can
anyone else reproduce
this? Thanks.
Regards, Miro
from dolfin import *
method = 'mumps'
symmetric = False
for N in (320, 384, 448, 512):
mesh = UnitSquareMesh(N, N)
V = FunctionSpace(mesh, 'CG', 1)
u = TrialFunction(V)
v = TestFunction(V)
a = inner(grad(u), grad(v))*dx
L = inner(Constant(2), v)*dx
u = Function(V)
bc = DirichletBC(V, Constant(0), DomainBoundary())
A, b = PETScMatrix(), PETScVector()
assemble_system(a, L, bc, A_tensor=A, b_tensor=b)
U = Function(V).vector()
solver = PETScLUSolver(A, method)
solver.parameters['symmetric'] = symmetric
#
timer = Timer('solver')
timer.start()
solver.solve(U, b)
print timer.stop()
# MUMPS symmetric < MUMPS not-symmetric < PETSc not-symmetic << PETSc symmetric
# 1.17892622948 1.76611208916 2.52074193954 3.348279953
# 1.27664494514 1.93275213242 2.79323792458 3.74768614769
# 1.33259105682 2.27041697502 3.53542280197 5.21441793442
# 23.6192109585 too long
_______________________________________________
fenics mailing list
[email protected]
http://fenicsproject.org/mailman/listinfo/fenics
