This is a collection of short examples for PyAMG. The source code for these (and more) examples is available at https://github.com/pyamg/pyamg-examples.
- Introduction
- Blackbox Solver
- Smoothed Aggregation AMG
- Classical AMG
- Rootnode AMG
- Finite Elements
- Preconditioning
- Other Applications
- Other
As a starting example, this demo considers a rotated anisotropic
diffusion problem from the pyamg.gallery
. First, a basic
smoothed aggregation solver is constructed. Then, many of the
options are modified to yield a more effective solver.
The comments in demo.py
follow several steps that walk through the demo:
- Step 1: import scipy and pyamg packages
- Step 2: setup up the system using pyamg.gallery
- Step 3: setup of the multigrid hierarchy
- Step 4: solve the system
- Step 5: print details
- Step 6: change the hierarchy
- Step 7: print details
- Step 8: plot convergence history
Details: Default AMG
--------------------
MultilevelSolver
Number of Levels: 5
Operator Complexity: 1.125
Grid Complexity: 1.127
Coarse Solver: 'pinv'
level unknowns nonzeros
0 40000 357604 [88.91%]
1 4489 39601 [9.85%]
2 529 4489 [1.12%]
3 64 484 [0.12%]
4 9 49 [0.01%]
The residual norm is 0.2110043217253744
The Multigrid Hierarchy
-----------------------
A_0: 40000x40000 P_0: 40000x4489
A_1: 4489x4489 P_1: 4489x529
A_2: 529x529 P_2: 529x64
A_3: 64x64 P_3: 64x9
A_4: 9x9
Details: Specialized AMG
------------------------
MultilevelSolver
Number of Levels: 6
Operator Complexity: 2.159
Grid Complexity: 1.201
Coarse Solver: 'pinv'
level unknowns nonzeros
0 40000 357604 [46.31%]
1 6700 226352 [29.31%]
2 1232 176222 [22.82%]
3 109 11827 [1.53%]
4 13 169 [0.02%]
5 4 16 [0.00%]
The residual norm is 1.1195909450352103e-10
This demo highlights using PyAMG's blackbox
module, which attempts to solve
an arbitrary system A x = b
with minimal input. The matrix A
can be
non-Hermitian, indefinite, Hermitian positive-definite, etc... The method tries
generic and
robust settings for smoothed_aggregation_solver(..)
. If
solve()
fails to effectively solve the system,
then it may be helpful to look at the demo in solver_diagonstics
for guidance on automatically finding better parameter settings.
To use solve()
, only the matrix A
and a right-hand side b
are needed:
x = pyamg.solve(A, b, verb=True)
The demo produces residual norms that can vary from machine to machine.
Detected a Hermitian matrix
maxiter = 400
iteration 1.0
iteration 2.0
iteration 3.0
iteration 4.0
iteration 5.0
iteration 6.0
iteration 7.0
Residuals ||r_k||_M, ||r_0||_M = 6.45e-01, 8.68e+06
Residual reduction ||r_k||_M/||r_0||_M = 7.43e-08
In this example, the first-level aggregates are shown for AMG based on smoothed aggregation.
An example mesh and adjacency matrix is loaded from square.mat
, followed by a call to
smoothed_aggregation_solver
. Then the first-level aggregates are
plotted. From the figure, most aggregates encompass entire groups of
elements in the underlying mesh. Still, there are a many aggregates that yield
"strings" in the aggregation, often impacting performance.
This example illustrates the effect, in 1D, of smoothed aggregation on tentative prolongation operators. Each of the aggregates (groups of three in this case) are plotted with the associated (smoothed) basis functions.
In these two examples, the pyamg.vis
module is called to display
aggregation in both two and three dimensions. demo1.py
considers the Poisson
problem on an unstructured triangulation of the unit square (from the PyAMG
gallery). In demo2.py
, the same Poisson problem is considered on an
unstructured tetrahedral mesh on the unit cube. Two VTK compliant output files
are generated in each case: output_mesh.vtu
and output_aggs.vtu
.
output_mesh.vtu
provides information on the underlying mesh (straight from
the unit square mesh), while output_aggs.vtu
holds information on the
aggregates generated from first level of Smoothed Aggregation. The process of
visulization in paraview is straightforward:
Start Paraview:
- open file:
output_mesh.vtu
- apply
- under display in the object inspector: select wireframe representation
- under display in the object inspector: select a better solid color
- open file:
output_aggs.vtu
- apply
- under display in the object inspector: select surface with edges representation
- under display in the object inspector: select a better solid color
- under display in the object inspector: increase line width to see line aggregates (if present)
- under display in the object inspector: increase point size to see point aggregates (if present)
AMG has a range of parameter choices; selecting the optimal combination can be challenging yet can lead to significant improvements in convergence. This example highlights a "solver diagnostics" function that makes finding good parameter choices a bit easier. A brute force search is applied, and depending on the matrix characteristics (e.g., symmetry and definiteness), 60-120 different solvers are constructed and then tested. As a result, this test is intended for smaller matrix problems.
The function solver_diagnostics
(solver_diagnostics.py
) has many
parameters, but the defaults should be sufficient. For this test,
only the matrix, A
, is needed, and A
can be nonsymmetric, indefinite, or
symmetric positive definite. The function detects symmetry and definiteness,
but it is safest to specify these.
The function outputs two separate files and we briefly examine this output
for the second example of rotated anisotropic diffusion when running the above
demo.py
.
Running
python demo.py --matrix 2
will run solver diagnostics on the rotated anisotropic diffusion problem.
The first output file is rot_ani_diff_diagnostic.txt
, which is a sorted table
of solver statistics for all the solvers tried. This file has detailed output
for the performance of each solver, and the parameter choices used for
each solver.
The second file defines a function rot_ani_diff_diagnostic.py
, that when
given a matrix, automatically generates and uses the best solver found.
Searching for optimal smoothed aggregation method for (2500,2500) matrix
...
User specified a symmetric matrix
User specified definiteness as positive
...
Test 1 out of 18
Test 2 out of 18
Test 3 out of 18
Test 4 out of 18
Test 5 out of 18
Test 6 out of 18
Test 7 out of 18
Test 8 out of 18
Test 9 out of 18
Test 10 out of 18
Test 11 out of 18
Test 12 out of 18
Test 13 out of 18
Test 14 out of 18
Test 15 out of 18
Test 16 out of 18
Test 17 out of 18
Test 18 out of 18
--> Diagnostic Results located in rot_ani_diff_diagnostic.txt
--> See automatically generated function definition in rot_ani_diff_diagnostic.py
The smoothed aggregation solver supports complex arithmetc and there is no conversion to an equivalent real system. For example, the highlighted demo here generates a basic gauge Laplacian from quantum chromodynamics and solves the system for a random right-hand side and random initial guess.
Using
python demo.py --solver 1
results in the following.
residual at iteration 0: 2.00e+02
residual at iteration 1: 1.21e+02
residual at iteration 2: 2.47e+01
residual at iteration 3: 5.40e+00
residual at iteration 4: 1.46e+00
residual at iteration 5: 4.60e-01
residual at iteration 6: 1.58e-01
residual at iteration 7: 5.77e-02
residual at iteration 8: 2.20e-02
residual at iteration 9: 8.70e-03
residual at iteration 10: 3.53e-03
residual at iteration 11: 1.46e-03
residual at iteration 12: 6.12e-04
residual at iteration 13: 2.59e-04
residual at iteration 14: 1.11e-04
residual at iteration 15: 4.75e-05
residual at iteration 16: 2.04e-05
residual at iteration 17: 8.82e-06
residual at iteration 18: 3.81e-06
residual at iteration 19: 1.65e-06
residual at iteration 20: 7.13e-07
MultilevelSolver
Number of Levels: 5
Operator Complexity: 1.344
Grid Complexity: 1.184
Coarse Solver: 'pinv'
level unknowns nonzeros
0 10000 50000 [74.43%]
1 1658 15132 [22.53%]
2 170 1906 [2.84%]
3 12 138 [0.21%]
4 1 1 [0.00%]
The smoothed aggregation solver supports nonsymmetric (i.e., non-Hermitian) and indefinite matrices, through recent advances in multigrid research. The demo highlighted here constructs a solver for a small nonsymmetric recirculating flow problem. The out-of-the-box example diverges, while more advanced options yield a convergent solver.
Using
python demo.py --solver 1
we observe the following convergence history.
Observe that standard multigrid parameters for Hermitian systems
yield a nonconvergent stand-alone solver.
residual at iteration 0: 8.64e-01
residual at iteration 1: 6.92e-01
residual at iteration 2: 2.02e+01
residual at iteration 3: 5.89e+02
residual at iteration 4: 1.72e+04
residual at iteration 5: 5.02e+05
residual at iteration 6: 1.47e+07
residual at iteration 7: 4.28e+08
residual at iteration 8: 1.25e+10
residual at iteration 9: 3.65e+11
residual at iteration 10: 1.07e+13
residual at iteration 11: 3.12e+14
residual at iteration 12: 9.10e+15
residual at iteration 13: 2.66e+17
residual at iteration 14: 7.76e+18
residual at iteration 15: 2.27e+20
*************************************************************
Now using nonsymmetric parameters for multigrid , we obtain a
convergent stand-alone solver.
residual at iteration 0: 8.64e-01
residual at iteration 1: 1.14e-01
residual at iteration 2: 3.53e-02
residual at iteration 3: 1.61e-02
residual at iteration 4: 8.68e-03
residual at iteration 5: 5.09e-03
residual at iteration 6: 3.08e-03
residual at iteration 7: 1.89e-03
residual at iteration 8: 1.18e-03
residual at iteration 9: 7.44e-04
residual at iteration 10: 4.78e-04
residual at iteration 11: 3.14e-04
residual at iteration 12: 2.11e-04
residual at iteration 13: 1.46e-04
residual at iteration 14: 1.04e-04
residual at iteration 15: 7.55e-05
*************************************************************
Now, we use the nonsymmetric solver to accelerate GMRES.
residual at iteration 0: 5.54e+00
residual at iteration 1: 1.47e+00
residual at iteration 2: 5.00e-01
residual at iteration 3: 2.96e-01
residual at iteration 4: 1.62e-01
residual at iteration 5: 5.09e-02
residual at iteration 6: 1.20e-02
residual at iteration 7: 4.86e-03
residual at iteration 8: 1.39e-03
residual at iteration 9: 9.68e-04
residual at iteration 10: 3.05e-04
residual at iteration 11: 1.36e-04
residual at iteration 12: 2.94e-05
residual at iteration 13: 6.66e-06
residual at iteration 14: 1.48e-06
residual at iteration 15: 4.32e-07
The C/F splitting---i.e. the splitting of indices into strictly coarse nodes
(C-pts) and strictly fine nodes (F-pts)---using Ruge-Stuben coarsening is
illustrated in this example. An example mesh and adjacency graph is loaded
from square.mat
, ruge_stuben_solver()
is initiated, and the first level of
C/F splitting is plotted. Printing the multilevel object in this case shows that
the coarsening is typical: around 25% reduction in unknowns (or
coarsening-by-four), as shown below. The demo also plots the coarse-fine
splitting, with the orange C-pts and the blue F-pts.
MultilevelSolver
Number of Levels: 2
Operator Complexity: 1.327
Grid Complexity: 1.267
Coarse Solver: 'pinv'
level unknowns nonzeros
0 191 1243 [75.33%]
1 51 407 [24.67%]
The C/F splitting---i.e. the splitting of indices into strictly coarse nodes (C-pts) and strictly fine nodes (F-pts)---using Compatible Relaxation is illustrated in this example. A 2d finite-difference matrix of the Poisson problem is used and the coarse and fine splitting is plotted. Coarse nodes are highlighted orange, while fine nodes are highlighted blue. In this case, the coarsening is not aggressive, resulting in a coarsening-by-two.
In this example we look at several strength of connection measures, including symmetric
,
evolution
, affinity
, and algebraic_distance
. From the output, we see the large impact
on convergence and the variability due to parameter selection.
running symmetric: theta=0.0
running symmetric: theta=0.25
running evolution: epsilon=4.0
running affinity: epsilon=3.0, R=10, alpha=0.5, k=20
running affinity: epsilon=4.0, R=10, alpha=0.5, k=20
running algebraic_distance: epsilon=2.0, p=inf, R=10, alpha=0.5, k=20
running algebraic_distance: epsilon=3.0, p=inf, R=10, alpha=0.5, k=20
We demonstrate the use of AMG based on Approximate Ideal Restriction
(AIR) to solve upwind discretiazations of advection. Here we consider
a simple 2d first-order upwind finite difference discretization of
the steady advection problem
AIR is ideal for upwind discretizations of advection-dominated problems. For pure advection, often AIR does not need Krylov acceleration; we see this by converging to 1e-10 residual tolerance in as little as 7 iterations. AIR operator complexity tends to be large; here we compare using AIR with distance-1 and distance-2 restriction, and with and without second pass coarsening. Distance-2 restriction and second-pass coarsening will both increase operator complexity but also improve convergence.
python demo
500 x 500 mesh:
Distance-1 AIR using RS coarsening *without* second pass.
Levels in hierarchy: 10
Operator complexity: 2.604073699237941
Number of iterations: 39
Average convergence factor: 0.5452519246046812
Distance-1 AIR using RS coarsening *with* second pass.
Levels in hierarchy: 13
Operator complexity: 3.7809599131373113
Number of iterations: 8
Average convergence factor: 0.044070522107121105
Distance-2 AIR using RS coarsening *without* second pass.
Levels in hierarchy: 10
Operator complexity: 2.9532603668876214
Number of iterations: 11
Average convergence factor: 0.11031359469804951
Distance-2 AIR using RS coarsening *with* second pass.
Levels in hierarchy: 14
Operator complexity: 3.77129643903191
Number of iterations: 7
Average convergence factor: 0.029339749752205272
The rootnode_solver
is a mixture of both classical and aggregation-based
approaches to AMG, with the intent to combine their strengths, while minimizing
their respective drawbacks. As a result, this solver is more robust for some
problem types, especially anisotropic diffusion.
In terms of use, the interface to pyamg.aggregation.rootnode_solver(...)
is
identical to pyamg.aggregation.smoothed_aggregation_solver(...)
, meaning that
the above aggregation examples can be easily changed by simply replacing calls
to smoothed_aggregation_solver()
with rootnode_solver()
.
This example compares the rootnode coarsening to classical AMG's coarsening (see the Coarse Fine Splitting Example) and to smoothed aggregation's coarsening (see the Aggregation Example). The rootnode approach mixes classical AMG and smoothed aggregation, and hence has an associated C/F splitting that splits the indices into strictly coarse (C) nodes and strictly fine (F) nodes, and also has an associated aggregation that disjointly splits the nodes into strongly connected neighborhoods. Essentially, each aggregate has one "root" C-node associated with it, that is injected between the fine and coarse grids.
An example mesh and adjacency graph is loaded from square.mat
, and the
rootnode_solver()
is initiated. Then, the first-level C/F splitting and the
first-level aggregation are plotted. Coarse nodes are highlighted orange, while
fine nodes are highlighted blue.
In general (as well as in this example), the C/F splitting for rootnode contains far fewer coarse nodes than for classical AMG. In general, this fewer number of coarse nodes is compensated by having a somewhat denser interpolation operator than for classical AMG.
This demo considers different strength measures in the SA-AMG setup phase for finite element (Q1) discretizations of anisotropic diffusion. In particular, the Classic Strength Measure is compared to the Evolution Measure. For this example, we see that total work is reduced by using the Evolution Measure and that a scalable convergence rate is observed with rootnode:
Running Grid = (100 x 100)
Running Grid = (200 x 200)
Running Grid = (300 x 300)
Running Grid = (400 x 400)
AMG Scalability Study for Ax = 0, x_init = rand
Emphasis on Robustness of Evolution Strength
Measure and Root-Node Solver
Rotated Anisotropic Diffusion in 2D
Anisotropic Coefficient = 1.000e-03
Rotation Angle = 0.393
n | nnz | rho | OpCx | Work
--------------------------------------------------------
Classic strength
--------------------------------------------------------
10000 | 88804 | 0.86 | 1.6 | 24
40000 | 357604 | 0.87 | 1.6 | 25
90000 | 806404 | 0.87 | 1.6 | 27
160000 | 1435204 | 0.87 | 1.6 | 27
--------------------------------------------------------
Evolution strength
--------------------------------------------------------
10000 | 88804 | 0.56 | 1.8 | 7
40000 | 357604 | 0.67 | 1.8 | 10
90000 | 806404 | 0.69 | 1.8 | 11
160000 | 1435204 | 0.72 | 1.8 | 13
--------------------------------------------------------
Evolution strength with Rootnode
--------------------------------------------------------
10000 | 88804 | 0.46 | 1.8 | 5.4
40000 | 357604 | 0.49 | 1.9 | 6
90000 | 806404 | 0.5 | 1.9 | 6.3
160000 | 1435204 | 0.5 | 1.9 | 6.5
We consider the 2D linear elasticity problem from the pyamg gallery in this
example (corresponding to a simple finite element discretization on a regular
grid). Three near null space modes are fed to the
smoothed_aggregation_solver()
(relating to rotation and two types of
translation). Smoothed aggregation and root node are ideal for this problem
and the results are apparent. Very low operator complexities are observed and
the convergence is quick, whether you choose the root node or smoothed
aggregation solver.
Using
python demo --solver 2
results in the following.
MultilevelSolver
Number of Levels: 5
Operator Complexity: 1.125
Grid Complexity: 1.127
Coarse Solver: 'pinv'
level unknowns nonzeros
0 80000 1430416 [88.91%]
1 8978 158404 [9.85%]
2 1058 17956 [1.12%]
3 128 1936 [0.12%]
4 18 196 [0.01%]
Number of iterations: 19d
residual at iteration 0: 1.63e+02
residual at iteration 1: 1.13e+02
residual at iteration 2: 8.19e+00
residual at iteration 3: 1.12e+00
residual at iteration 4: 2.57e-01
residual at iteration 5: 6.78e-02
residual at iteration 6: 1.86e-02
residual at iteration 7: 5.19e-03
residual at iteration 8: 1.46e-03
residual at iteration 9: 4.11e-04
residual at iteration 10: 1.17e-04
residual at iteration 11: 3.32e-05
residual at iteration 12: 9.52e-06
residual at iteration 13: 2.73e-06
residual at iteration 14: 7.89e-07
residual at iteration 15: 2.29e-07
residual at iteration 16: 6.65e-08
residual at iteration 17: 1.94e-08
residual at iteration 18: 5.68e-09
This example shows how to effectively use multilevel solvers to precondition a Krylov method. The first example considers the Poisson problem from the pyamg gallery and uses a constant near-nullspace vector for SA-AMG. The second example is 2D linear elasticity also from the pyamg gallery and uses the typical three rigid body modes (rotation and translation in x and y) to coach SA-AMG. Since both problems are symmetric and positive definite, CG acceleration is used. The residual histories show a clear improvement in using the SA-AMG preconditioners in both cases.
Using
python demo.py --solver 1
produces the following.
Matrix: Poisson
MultilevelSolver
Number of Levels: 6
Operator Complexity: 1.337
Grid Complexity: 1.188
Coarse Solver: 'pinv'
level unknowns nonzeros
0 250000 1248000 [74.82%]
1 41750 373416 [22.39%]
2 4704 41554 [2.49%]
3 532 4526 [0.27%]
4 65 509 [0.03%]
5 9 65 [0.00%]
Matrix: Elasticity
MultilevelSolver
Number of Levels: 5
Operator Complexity: 1.281
Grid Complexity: 1.191
Coarse Solver: 'pinv'
level unknowns nonzeros
0 80000 1430416 [78.08%]
1 13467 356409 [19.45%]
2 1587 40401 [2.21%]
3 192 4356 [0.24%]
4 27 441 [0.02%]
In this this example, smoothed aggregation AMG is used to precondition the
LOBPCG eigensolver to find the lowest nine eigenmodes of a Poisson problem.
With preconditioning (M=M
in the loppcg
call), the computation of the
eigensubspace is extremely fast.
In this example, we compute a partition of a basic cracked mesh (crack_mesh.mat
)
using the Fiedler vector (the second lowest eigenmode of the graph laplacian).
We construct a SA-AMG preconditioner to assist LOBPCG in finding the Fiedler
vector. Positive/negative values of the Fiedler vector are plotted in
different colors, illustrating the natural splitting this mesh.
The example focusses on the indefinite Helmholtz wave problem. The first
part highlights the value in using waves to represent the near-null space, B
.
In addition, we observe the waves to resemble the (lowest) right singular vectors
of the problem.
In the case of 2D, discontinuous Galerkin is used, yielding multiple degrees of freedom at each spatial location. As a result, the fine level (level-0) aggregates of the discontinuous elements, largely group neighboring vertices. The wave-like near null-space is then enforced on the first coarse grid (level-1), resulting in four modes.
Running 2D Helmholtz Example
-- 10.00 Points-per-wavelength
-- 2.73e-01 = h, 2.50 = omega
-- Discretized with a local discontinuous Galerkin method
on annulus-shaped domain
Using only a constant mode for interpolation yields an inefficient solver.
This is due to aliasing oscillatory, but algebraically smooth, modes on the coarse levels.
residual at iteration 0: 3.14e+01
residual at iteration 1: 4.00e+00
residual at iteration 2: 2.53e+00
residual at iteration 3: 9.74e-01
residual at iteration 4: 2.76e-01
residual at iteration 5: 1.22e-01
residual at iteration 6: 5.91e-02
residual at iteration 7: 2.64e-02
residual at iteration 8: 1.48e-02
residual at iteration 9: 6.75e-03
residual at iteration 10: 3.30e-03
residual at iteration 11: 1.73e-03
residual at iteration 12: 6.48e-04
residual at iteration 13: 3.75e-04
residual at iteration 14: 1.44e-04
residual at iteration 15: 6.55e-05
residual at iteration 16: 3.85e-05
residual at iteration 17: 1.34e-05
residual at iteration 18: 7.11e-06
residual at iteration 19: 3.81e-06
residual at iteration 20: 9.36e-07
Note the improved performance from using planewaves in B.
residual at iteration 0: 3.05e+01
residual at iteration 1: 2.62e-01
residual at iteration 2: 2.42e-03
residual at iteration 3: 2.88e-05
residual at iteration 4: 3.29e-07
residual at iteration 5: 3.87e-09
MultilevelSolver
Number of Levels: 4
Operator Complexity: 1.435
Grid Complexity: 1.411
Coarse Solver: 'pinv2'
level unknowns nonzeros
0 2880 52016 [69.67%]
1 880 10480 [14.04%]
2 256 9856 [13.20%]
3 48 2304 [3.09%]
In this example we look at a 2D DG discretization of a Poisson problem.
The mesh consists of 46 elements, with p=5, leading to 21 degrees of freedom
per element. The first figure shows that aggregation is local, leading to
a continuous first level. The first coarse level (level1) candidate vector B
is
also shown.
Diffusion problem discretized with p=5 and the local
discontinuous Galerkin method.
Observe that standard SA parameters for this p=5 discontinuous
Galerkin system yield an inefficient solver.
residual at iteration 0: 2.98e+02
residual at iteration 1: 1.06e+01
residual at iteration 2: 5.08e+00
residual at iteration 3: 2.71e+00
residual at iteration 4: 1.64e+00
residual at iteration 5: 1.01e+00
residual at iteration 6: 5.20e-01
residual at iteration 7: 3.57e-01
residual at iteration 8: 2.19e-01
residual at iteration 9: 1.25e-01
residual at iteration 10: 8.53e-02
residual at iteration 11: 5.44e-02
residual at iteration 12: 3.50e-02
residual at iteration 13: 2.59e-02
residual at iteration 14: 1.79e-02
residual at iteration 15: 1.15e-02
residual at iteration 16: 6.57e-03
residual at iteration 17: 4.40e-03
residual at iteration 18: 2.49e-03
residual at iteration 19: 1.37e-03
residual at iteration 20: 8.23e-04
Now use appropriate parameters, especially 'energy' prolongation
smoothing and a distance based strength measure on level 0. This
yields a much more efficient solver.
residual at iteration 0: 2.98e+02
residual at iteration 1: 1.32e+00
residual at iteration 2: 8.72e-02
residual at iteration 3: 1.02e-02
residual at iteration 4: 6.38e-04
residual at iteration 5: 6.65e-05
residual at iteration 6: 6.01e-06
residual at iteration 7: 6.16e-07
residual at iteration 8: 4.92e-08
residual at iteration 9: 5.41e-09
MultilevelSolver
Number of Levels: 5
Operator Complexity: 1.623
Grid Complexity: 1.807
Coarse Solver: 'pinv2'
level unknowns nonzeros
0 966 35338 [61.63%]
1 652 19602 [34.19%]
2 94 2006 [3.50%]
3 27 341 [0.59%]
4 7 49 [0.09%]
This example highlights the lowest order edge AMG implementation of the
Reitzinger-Schoberl algorithm. From the convergence figure we observe
significant improvements over out-of-the-box AMG due to the use of
the specialized relaxation method (hiptmair_smoother
).
This demo shows the performance of finite element assembly (by way of scikit-fem)
and multigrid setup/solve (within PCG). The solver is run to a tolerance of 1e-8
.
The figure shows:
DoFs
: the total number of degrees of freedom in the systemAssembly
: the total time to assemble the FE matrix (scikit-fem)Solve prep
: the total time to condense the system to non-Dirichlet nodes (scikit-fem)Solve setup
: the total time for the AMG setup phase (pyamg)Solve
: the total time for the AMG solve phase (pyamg) withing PCG (scikit-fem)
| DoFs | Assembly | Solve prep | Solve setup | Solve |
|-------------|-------------|-------------|-------------|-------------|
| 8 | 0.00169 | 0.00045 | 0.00011 | 0.00049 |
| 27 | 0.00142 | 0.00027 | 0.00148 | 0.00054 |
| 64 | 0.00193 | 0.00025 | 0.00210 | 0.00061 |
| 216 | 0.00393 | 0.00025 | 0.00225 | 0.00079 |
| 512 | 0.00760 | 0.00028 | 0.00331 | 0.00099 |
| 1000 | 0.01450 | 0.00034 | 0.00462 | 0.00152 |
| 2744 | 0.03646 | 0.00055 | 0.00776 | 0.00357 |
| 5832 | 0.07144 | 0.00088 | 0.01526 | 0.00664 |
| 12167 | 0.14817 | 0.00150 | 0.01945 | 0.01439 |
| 27000 | 0.33348 | 0.00307 | 0.05211 | 0.03602 |
| 54872 | 0.68135 | 0.00625 | 0.07649 | 0.07217 |
| 110592 | 1.41124 | 0.01306 | 0.18654 | 0.16615 |
| 238328 | 3.15382 | 0.03606 | 0.61186 | 0.40605 |
| 474552 | 6.58809 | 0.06941 | 1.00435 | 0.86555 |
This is a short example on profiling the setup phase of AMG.
Here, we use pyinstrument
to analyze the construction of
a smoothed aggregation solver:
_ ._ __/__ _ _ _ _ _/_ Recorded: 12:00:00 Samples: 598
/_//_/// /_\ / //_// / //_'/ // Duration: 1.320 CPU time: 7.091
/ _/ v4.4.0
Program: demo.py --savefig
�[31m1.319�[0m �[48;5;24m�[38;5;15m<module>�[0m �[2mdemo.py:1�[0m
└─ �[31m1.319�[0m smoothed_aggregation_solver�[0m �[2mpyamg/aggregation/aggregation.py:26�[0m
[236 frames hidden] �[2mpyamg, <__array_function__ internals>...�[0m
�[33m0.748�[0m _approximate_eigenvalues�[0m �[2mpyamg/util/linalg.py:156�[0m