maphys issueshttps://gitlab.inria.fr/solverstack/maphys/maphys/issues2019-01-25T16:01:11Zhttps://gitlab.inria.fr/solverstack/maphys/maphys/issues/15FIll Ratio in Maphys2019-01-25T16:01:11ZAFrahFIll Ratio in MaphysHow can I find the fill ratio in Maphys solver which is the number of nonzeros in the final matrix divided by the number of nonzeros in the original matrix. Can I say like this:
(IINFO(6), local matrix nnz(max) + IINFO(13), local schur nnz(max))/ (nonzeros in the original matrix).
What if the matrix is symmetric... should I multiply the numerator by two or what?
Thank you,How can I find the fill ratio in Maphys solver which is the number of nonzeros in the final matrix divided by the number of nonzeros in the original matrix. Can I say like this:
(IINFO(6), local matrix nnz(max) + IINFO(13), local schur nnz(max))/ (nonzeros in the original matrix).
What if the matrix is symmetric... should I multiply the numerator by two or what?
Thank you,https://gitlab.inria.fr/solverstack/maphys/maphys/issues/6Multithreading with new pastix2017-12-08T18:11:39ZKUHN MatthieuMultithreading with new pastixConcerns branch feature@newpastix.
- [x] With the new pastix, to enable multithreading, the number of threads in use have to be set before pastixInit. However, the current SDS factorization and sparse precond factorization call XMPH_SDS_set_multithread (setting multithreading) after XMPH_SDS_set_matrix (calling pastixInit). Some investigation are required to find the appropriate order which will preserve both pastix and mumps functionalities.
- [x] About multithreading again, manual binding has to be investigated too. The subroutine pastixInitWithAffinity is going to be reachable from fortran very soon.
- [ ] Sparse preconditioning with PILUT: requires pastix perm, unavailable for this moment.
Concerns branch feature@newpastix.
- [x] With the new pastix, to enable multithreading, the number of threads in use have to be set before pastixInit. However, the current SDS factorization and sparse precond factorization call XMPH_SDS_set_multithread (setting multithreading) after XMPH_SDS_set_matrix (calling pastixInit). Some investigation are required to find the appropriate order which will preserve both pastix and mumps functionalities.
- [x] About multithreading again, manual binding has to be investigated too. The subroutine pastixInitWithAffinity is going to be reachable from fortran very soon.
- [ ] Sparse preconditioning with PILUT: requires pastix perm, unavailable for this moment.
KUHN MatthieuKUHN Matthieuhttps://gitlab.inria.fr/solverstack/maphys/maphys/issues/4Unknowns disappear during analysis2017-09-28T12:41:06ZMARAIT GillesUnknowns disappear during analysisWhen running MaPHyS testcase:
mpirun -np 4 ./dmph_examplekv real_bcsstk17.in
Unknowns disappear. It can be seen when dumping the matrices after analysis. For example, for the 3rd domain :
3839 2276 -1.129781473007000022E+05
3839 2279 8.198356871495000087E+04
3846 2200 -1.063664884091000073E+06
3846 2279 -4.604417188240000047E+06
unknows between 3839 and 3846 are gone.
The output solution given does not seem very satisfying :
== ESTIMATION OF COMMITED ERROR ==
Backward error: |A.x-b|/|b| norm2 = 8.1775593529E-08
Backward error centralized: |A.x-b|/|b| norm2 = 8.1775593533E-08
Computed (x) vs analytical (x*) solution: |x-x*|/|x*| norm2 = 1.5351223696E-05
Computed (x) vs analytical (x*) solution: |x-x*|/|x*| norm_inf = 7.0400129879E-05
== FIRST VALUES OF ESTIMATED VS ANALYTICAL SOLUTION ==
| Computed (x) | Theoretical (x*)
_____________________________________________________
|sol( 1)| | 1.0000000000E+00 | 1.0000000000E+00
|sol( 2)| | 1.2274289747E+00 | 2.0000000000E+00
|sol( 3)| | 3.0395067110E+00 | 3.0000000000E+00
...
NB: this bug does not appear when running with Paddle.When running MaPHyS testcase:
mpirun -np 4 ./dmph_examplekv real_bcsstk17.in
Unknowns disappear. It can be seen when dumping the matrices after analysis. For example, for the 3rd domain :
3839 2276 -1.129781473007000022E+05
3839 2279 8.198356871495000087E+04
3846 2200 -1.063664884091000073E+06
3846 2279 -4.604417188240000047E+06
unknows between 3839 and 3846 are gone.
The output solution given does not seem very satisfying :
== ESTIMATION OF COMMITED ERROR ==
Backward error: |A.x-b|/|b| norm2 = 8.1775593529E-08
Backward error centralized: |A.x-b|/|b| norm2 = 8.1775593533E-08
Computed (x) vs analytical (x*) solution: |x-x*|/|x*| norm2 = 1.5351223696E-05
Computed (x) vs analytical (x*) solution: |x-x*|/|x*| norm_inf = 7.0400129879E-05
== FIRST VALUES OF ESTIMATED VS ANALYTICAL SOLUTION ==
| Computed (x) | Theoretical (x*)
_____________________________________________________
|sol( 1)| | 1.0000000000E+00 | 1.0000000000E+00
|sol( 2)| | 1.2274289747E+00 | 2.0000000000E+00
|sol( 3)| | 3.0395067110E+00 | 3.0000000000E+00
...
NB: this bug does not appear when running with Paddle.