Re: [petsc-users] CG with right preconditioning supports NONE norm type only

[Patrick Sanan]

There is indeed a way to interpret preconditioned CG as left preconditioning:

Consider

Ax = b

where A is spd.

You can write down the equivalent left-preconditioned system

   M^{-1}Ax = M^{-1}b

but of course this doesn't work directly with CG because M^{-1}A is no
longer spd in general.

However, if M^{-1} is spd, so is M and you can use it to define an
inner product. In this inner product, M{-1}A *is* spd. Here assume
real scalars for simplicity:

  _M = _M =  =  =  = _M

So, you can write use CG, using this inner product everywhere. You
don't in fact ever need to apply M, just M^{-1}, and you arrive at
standard preconditioned CG (See these notes [1]).

This is what's in PETSc as left-preconditioned CG, the only option.

This is a bit different from the way left preconditioning works for
(say) GMRES. With GMRES you could assemble A' = M^{-1}A and b' =
M^{-1}b and pass A' and b' to black-box (unpreconditioned) GMRES. With
CG, you fundamentally need to provide both M^{-1} and A (and
fortunately PETSc lets you do exactly this).

This then invites the natural question about applying the same logic
to the right preconditioned system. I saw this as an exercise in the
same course notes as above, from J. M. Melenk at TU Wien [1] .

Consider the right-preconditioned system

AM^{-1}y = b,x = M^{-1}y .

You can note that AM^-1 is spd in the M^{-1} inner product, and again
write down CG for the preconditioned system, now using M^{-1} inner
products everywhere.

You can then (as in this photo which I'm too lazy to transcribe [2]),
introduce transformed variables z = M^{-1}r, q = M^{-1}p, x = M^{-1}y
and work things out and arrive at ... standard preconditioned CG!

Thus the conclusion is that there really is only one way to
precondition CG. You can derive it as
1. left-preconditioned CG in the M norm, or
2. right-preconditioned CG in the M^{-1} norm, or
3. symmetrically-preconditioned CG (with M^{-1/2} which you don't need
to explicitly form), or
4. standard CG with a new inner product.

The last option is, I think, the most satisfying and fundamental. The
recent Málek and Strakoš book [3] makes a beautiful case as to why,
and indeed Hestenes himself first wrote about this in 1956, four years
after he and Stiefel popularized CG and before the term
"preconditioning" even existed [4].

So, in terms of concrete things to do with PETSc, I stand by my
feeling that calling CG preconditioning "left preconditioning" is
somehow misleading.

Rather, I'd suggest calling the relevant value of PCSide something
like PC_SPD . This
i) reminds the user of the fact that the preconditioner has to be spd,
ii) removes the asymmetry of calling it a left preconditioner when you
could equally-well call it a right preconditioner, and
iii) highlights the fact that preconditioning is different with CG
than with methods like GMRES.

PC_LEFT could be deprecated for CG before removing support.

If this seems reasonable, I can make a PR with these changes and
updates to the man page (and should also think about what other KSPs
this applies to - probably KSPMINRES and others).

[1] http://www.asc.tuwien.ac.at/~melenk/teach/iterative/
[2] http://patricksanan.com/rpcg.jpg
[3] http://bookstore.siam.org/sl01/
[4] 
https://scholar.google.ch/scholar?q=M.+Hestenes+1956.+The+Conjugate+Gradient+Method+for+Solving+Linear+Systems



On Wed, Mar 8, 2017 at 6:57 PM, Barry Smith  wrote:
>
>   Patrick,
>
> Thanks, this is interesting, we should try to add material to the KSPCG 
> page to capture some of the subtleties that I admit after 28 years I still 
> don't really understand. The paper A TAXONOMY FOR CONJUGATE GRADIENT METHODS 
> (attached) has some interesting discussion (particularly page 1548 
> "Therefore, only left preconditioning need be considered: Right 
> preconditioning may be effected by incorporating it into the left 
> preconditioner and inner product." I don't know exactly what this means in 
> practical terms in respect to code. (In PETSc KSP we don't explicitly talk 
> about, or use a separate "inner product"  in the Krylov methods, we only have 
> the concept of operator and preconditioner operator.)
>
>   Remembering vaguely, perhaps incorrectly, from a real long time ago "left 
> preconditioning with a spd M is just the unpreconditioned cg in the M inner 
> product" while "right preconditioning with M is unpreconditioned cg in a M^-1 
> inner product". If this is correct then it implies (I think) that right 
> preconditioned CG would produce a different set of iterates than left 
> preconditioning and hence is "missing" in terms of completeness from PETSc 
> KSP.
>
>   Barry
>
>
>
>
>> On Mar 8, 2017, at 7:37 PM, Patrick Sanan  wrote:
>>
>> On Wed, Mar 8, 2017 at 11:12 AM, Barry Smith  wrote:
>>>
 On Mar 8, 2017, at 10:47 AM, Kong, Fande  wrote:

 Thanks Barry,

 

Re: [petsc-users] CG with right preconditioning supports NONE norm type only

[Barry Smith]

> On Mar 9, 2017, at 1:49 AM, Klaij, Christiaan  wrote:
> 
> Barry,
> 
> I came across the same problem and decided to use KSPSetNormType
> instead of KSPSetPCSide.  Do I understand correctly that CG with
> KSP_NORM_UNPRECONDITIONED would be as efficient as with
> KSP_NORM_PRECONDITIONED?

Yes, it as efficient.

> Since PC_RIGHT is not supported, I was
> under the impression that the former would basically be the
> latter with an additional true residual evaluation for the
> convergence monitor, which would be less efficient.

It is true for GMRES with left preconditioning (but not right) that 
computing the unpreconditioned residual norm adds a good amount of additional 
work but it is not true for CG, this is because CG, by its nature tracks both y 
= A x and B y so either is equally available to compute its norm. GMRES with 
left preconditioner computes  B A something at each iteration but doesn't 
directly compute the residual (either preconditioned or not) it uses a 
recursive formula for the residual norm.

> 
> Chris
> 
>> On Mar 8, 2017, at 10:47 AM, Kong, Fande  wrote:
>> 
>> Thanks Barry,
>> 
>> We are using "KSPSetPCSide(ksp, pcside)" in the code.  I just tried 
>> "-ksp_pc_side right", and petsc did not error out.
>> 
>> I like to understand why CG does not work with right preconditioning? 
>> Mathematically, the right preconditioning does not make sense?
> 
>   No, mathematically it makes sense to do it on the right. It is just that 
> the PETSc code was never written to support it on the right. One reason is 
> that CG is interesting that you can run with the true residual or the 
> preconditioned residual with left preconditioning, hence less incentive to 
> ever bother writing it to support right preconditioning. For completeness we 
> should support right as well as symmetric.
> 
>  Barry
> 
> 
> dr. ir. Christiaan Klaij  | Senior Researcher | Research & Development
> MARIN | T +31 317 49 33 44 | mailto:c.kl...@marin.nl | http://www.marin.nl
> 
> MARIN news: 
> http://www.marin.nl/web/News/News-items/Your-future-is-Blue-Blueweek-April-1013.htm
> 

Re: [petsc-users] CG with right preconditioning supports NONE norm type only

[Klaij, Christiaan]

Barry,

I came across the same problem and decided to use KSPSetNormType
instead of KSPSetPCSide.  Do I understand correctly that CG with
KSP_NORM_UNPRECONDITIONED would be as efficient as with
KSP_NORM_PRECONDITIONED? Since PC_RIGHT is not supported, I was
under the impression that the former would basically be the
latter with an additional true residual evaluation for the
convergence monitor, which would be less efficient.

Chris

> On Mar 8, 2017, at 10:47 AM, Kong, Fande  wrote:
>
> Thanks Barry,
>
> We are using "KSPSetPCSide(ksp, pcside)" in the code.  I just tried 
> "-ksp_pc_side right", and petsc did not error out.
>
> I like to understand why CG does not work with right preconditioning? 
> Mathematically, the right preconditioning does not make sense?

   No, mathematically it makes sense to do it on the right. It is just that the 
PETSc code was never written to support it on the right. One reason is that CG 
is interesting that you can run with the true residual or the preconditioned 
residual with left preconditioning, hence less incentive to ever bother writing 
it to support right preconditioning. For completeness we should support right 
as well as symmetric.

  Barry


dr. ir. Christiaan Klaij  | Senior Researcher | Research & Development
MARIN | T +31 317 49 33 44 | mailto:c.kl...@marin.nl | http://www.marin.nl

MARIN news: 
http://www.marin.nl/web/News/News-items/Your-future-is-Blue-Blueweek-April-1013.htm

Re: [petsc-users] CG with right preconditioning supports NONE norm type only

[Patrick Sanan]

On Wed, Mar 8, 2017 at 11:12 AM, Barry Smith  wrote:
>
>> On Mar 8, 2017, at 10:47 AM, Kong, Fande  wrote:
>>
>> Thanks Barry,
>>
>> We are using "KSPSetPCSide(ksp, pcside)" in the code.  I just tried 
>> "-ksp_pc_side right", and petsc did not error out.
>>
>> I like to understand why CG does not work with right preconditioning? 
>> Mathematically, the right preconditioning does not make sense?
>
>No, mathematically it makes sense to do it on the right. It is just that 
> the PETSc code was never written to support it on the right. One reason is 
> that CG is interesting that you can run with the true residual or the 
> preconditioned residual with left preconditioning, hence less incentive to 
> ever bother writing it to support right preconditioning. For completeness we 
> should support right as well as symmetric.

For standard CG preconditioning, which PETSc calls left
preconditioning, you use a s.p.d. preconditioner M to define an inner
product in the algorithm, and end up finding iterates x_k in K_k(MA;
Mb). That isn't quite the same as left-preconditioned GMRES, where you
apply standard GMRES to the equivalent system MAx=Mb, and also end up
finding iterates in K_k(MA,Mb). This wouldn't work for CG because MA
isn't s.p.d. in general, even if M and A are.

Standard CG preconditioning is often motivated as a clever way to do
symmetric preconditioning, E^TAEy = E^Tb, x=Ey, without ever needing E
explicitly, using only M=EE^T . y_k lives in K_k(E^TAE,E^Tb) and thus
x_k again lives in K_k(MA;Mb).

Thus, it's not clear that there is an candidate for a
right-preconditioned CG variant, as what PETSc calls "left"
preconditioning doesn't arise in the same way that it does for other
Krylov methods, namely using the standard algorithm on MAx=Mb. For
GMRES you would get a right-preconditioned variant by looking at the
transformed system AMy=b, x = My. This means that y_k lives in
K_k(AM,b), so x lives in K_k(MA,Mb), as before. For CG, AM wouldn't be
spd in general so this approach wouldn't make sense.

Another way to look at the difference in "left" preconditioning
between GMRES and CG is that

- introducing left preconditioning for GMRES alters both the Krylov
subspaces *and* the optimality condition: you go from minimizing || b
- Ax_k ||_2 over K_k(A;b) to minimizing || M (b-Ax_k) ||_2 over
K_k(MA;Mb).

- introducing "left" preconditioning for CG alters *only* the Krylov
subspaces: you always minimize || x - x_k ||_A , but change the space
from K_k(A;b) to K_k(MA;Mb).

Thus, my impression is that it's misleading to call standard CG
preconditioning "left" preconditioning in PETSc - someone might think
of GMRES and naturally ask why there is no right preconditioning.

One might define a new entry in PCSide to be used with CG and friends.
I can't think of any slam dunk suggestions yet, but something in the
genre of PC_INNERPRODUCT, PC_METRIC, PC_CG, or PC_IMPLICITSYMMETRIC,
perhaps.


>
>   Barry
>
>>
>> Fande,
>>
>> On Wed, Mar 8, 2017 at 9:33 AM, Barry Smith  wrote:
>>
>>   Please tell us how you got this output.
>>
>>   PETSc CG doesn't even implement right preconditioning. If you ask for it 
>> it should error out. CG supports no norm computation with left 
>> preconditioning.
>>
>>Barry
>>
>> > On Mar 8, 2017, at 10:26 AM, Kong, Fande  wrote:
>> >
>> > Hi All,
>> >
>> > The NONE norm type is supported only when CG is used with a right 
>> > preconditioner. Any reason for this?
>> >
>> >
>> >
>> > 0 Nonlinear |R| = 1.732051e+00
>> >   0 Linear |R| = 0.00e+00
>> >   1 Linear |R| = 0.00e+00
>> >   2 Linear |R| = 0.00e+00
>> >   3 Linear |R| = 0.00e+00
>> >   4 Linear |R| = 0.00e+00
>> >   5 Linear |R| = 0.00e+00
>> >   6 Linear |R| = 0.00e+00
>> >  1 Nonlinear |R| = 1.769225e-08
>> >   0 Linear |R| = 0.00e+00
>> >   1 Linear |R| = 0.00e+00
>> >   2 Linear |R| = 0.00e+00
>> >   3 Linear |R| = 0.00e+00
>> >   4 Linear |R| = 0.00e+00
>> >   5 Linear |R| = 0.00e+00
>> >   6 Linear |R| = 0.00e+00
>> >   7 Linear |R| = 0.00e+00
>> >   8 Linear |R| = 0.00e+00
>> >   9 Linear |R| = 0.00e+00
>> >  10 Linear |R| = 0.00e+00
>> >  2 Nonlinear |R| = 0.00e+00
>> > SNES Object: 1 MPI processes
>> >   type: newtonls
>> >   maximum iterations=50, maximum function evaluations=1
>> >   tolerances: relative=1e-08, absolute=1e-50, solution=1e-50
>> >   total number of linear solver iterations=18
>> >   total number of function evaluations=23
>> >   norm schedule ALWAYS
>> >   SNESLineSearch Object:   1 MPI processes
>> > type: bt
>> >   interpolation: cubic
>> >   alpha=1.00e-04
>> > maxstep=1.00e+08, minlambda=1.00e-12
>> > tolerances: relative=1.00e-08, absolute=1.00e-15, 
>> > lambda=1.00e-08
>> > maximum iterations=40
>> >   KSP Object:   1 MPI 

Re: [petsc-users] CG with right preconditioning supports NONE norm type only

[Jed Brown]

Looks fine to me.  Thanks.

Barry Smith  writes:

>A proposed fix 
> https://bitbucket.org/petsc/petsc/pull-requests/645/do-not-assume-that-all-ksp-methods-support
>
>Needs Jed's approval.
>
>Barry
>
>
>   
>> On Mar 8, 2017, at 10:33 AM, Barry Smith  wrote:
>> 
>> 
>>  Please tell us how you got this output. 
>> 
>>  PETSc CG doesn't even implement right preconditioning. If you ask for it it 
>> should error out. CG supports no norm computation with left preconditioning.
>> 
>>   Barry
>> 
>>> On Mar 8, 2017, at 10:26 AM, Kong, Fande  wrote:
>>> 
>>> Hi All,
>>> 
>>> The NONE norm type is supported only when CG is used with a right 
>>> preconditioner. Any reason for this?
>>> 
>>> 
>>> 
>>> 0 Nonlinear |R| = 1.732051e+00
>>>  0 Linear |R| = 0.00e+00
>>>  1 Linear |R| = 0.00e+00
>>>  2 Linear |R| = 0.00e+00
>>>  3 Linear |R| = 0.00e+00
>>>  4 Linear |R| = 0.00e+00
>>>  5 Linear |R| = 0.00e+00
>>>  6 Linear |R| = 0.00e+00
>>> 1 Nonlinear |R| = 1.769225e-08
>>>  0 Linear |R| = 0.00e+00
>>>  1 Linear |R| = 0.00e+00
>>>  2 Linear |R| = 0.00e+00
>>>  3 Linear |R| = 0.00e+00
>>>  4 Linear |R| = 0.00e+00
>>>  5 Linear |R| = 0.00e+00
>>>  6 Linear |R| = 0.00e+00
>>>  7 Linear |R| = 0.00e+00
>>>  8 Linear |R| = 0.00e+00
>>>  9 Linear |R| = 0.00e+00
>>> 10 Linear |R| = 0.00e+00
>>> 2 Nonlinear |R| = 0.00e+00
>>> SNES Object: 1 MPI processes
>>>  type: newtonls
>>>  maximum iterations=50, maximum function evaluations=1
>>>  tolerances: relative=1e-08, absolute=1e-50, solution=1e-50
>>>  total number of linear solver iterations=18
>>>  total number of function evaluations=23
>>>  norm schedule ALWAYS
>>>  SNESLineSearch Object:   1 MPI processes
>>>type: bt
>>>  interpolation: cubic
>>>  alpha=1.00e-04
>>>maxstep=1.00e+08, minlambda=1.00e-12
>>>tolerances: relative=1.00e-08, absolute=1.00e-15, 
>>> lambda=1.00e-08
>>>maximum iterations=40
>>>  KSP Object:   1 MPI processes
>>>type: cg
>>>maximum iterations=1, initial guess is zero
>>>tolerances:  relative=1e-05, absolute=1e-50, divergence=1.
>>>right preconditioning
>>>using NONE norm type for convergence test
>>>  PC Object:   1 MPI processes
>>>type: hypre
>>>  HYPRE BoomerAMG preconditioning
>>>  HYPRE BoomerAMG: Cycle type V
>>>  HYPRE BoomerAMG: Maximum number of levels 25
>>>  HYPRE BoomerAMG: Maximum number of iterations PER hypre call 1
>>>  HYPRE BoomerAMG: Convergence tolerance PER hypre call 0.
>>>  HYPRE BoomerAMG: Threshold for strong coupling 0.25
>>>  HYPRE BoomerAMG: Interpolation truncation factor 0.
>>>  HYPRE BoomerAMG: Interpolation: max elements per row 0
>>>  HYPRE BoomerAMG: Number of levels of aggressive coarsening 0
>>>  HYPRE BoomerAMG: Number of paths for aggressive coarsening 1
>>>  HYPRE BoomerAMG: Maximum row sums 0.9
>>>  HYPRE BoomerAMG: Sweeps down 1
>>>  HYPRE BoomerAMG: Sweeps up   1
>>>  HYPRE BoomerAMG: Sweeps on coarse1
>>>  HYPRE BoomerAMG: Relax down  symmetric-SOR/Jacobi
>>>  HYPRE BoomerAMG: Relax upsymmetric-SOR/Jacobi
>>>  HYPRE BoomerAMG: Relax on coarse Gaussian-elimination
>>>  HYPRE BoomerAMG: Relax weight  (all)  1.
>>>  HYPRE BoomerAMG: Outer relax weight (all) 1.
>>>  HYPRE BoomerAMG: Using CF-relaxation
>>>  HYPRE BoomerAMG: Not using more complex smoothers.
>>>  HYPRE BoomerAMG: Measure typelocal
>>>  HYPRE BoomerAMG: Coarsen typeFalgout
>>>  HYPRE BoomerAMG: Interpolation type  classical
>>>linear system matrix followed by preconditioner matrix:
>>>Mat Object: 1 MPI processes
>>>  type: mffd
>>>  rows=9, cols=9
>>>Matrix-free approximation:
>>>  err=1.49012e-08 (relative error in function evaluation)
>>>  Using wp compute h routine
>>>  Does not compute normU
>>>Mat Object:() 1 MPI processes
>>>  type: seqaij
>>>  rows=9, cols=9
>>>  total: nonzeros=49, allocated nonzeros=49
>>>  total number of mallocs used during MatSetValues calls =0
>>>not using I-node routines
>>> 
>>> Fande,
>>> 
>> 


signature.asc
Description: PGP signature

Re: [petsc-users] CG with right preconditioning supports NONE norm type only

[Kong, Fande]

Thanks, Barry.

Fande,

On Wed, Mar 8, 2017 at 3:55 PM, Barry Smith  wrote:

>
>A proposed fix https://urldefense.proofpoint.com/v2/url?u=https-3A__
> bitbucket.org_petsc_petsc_pull-2Drequests_645_do-2Dnot-
> 2Dassume-2Dthat-2Dall-2Dksp-2Dmethods-2Dsupport=DQIFAg=
> 54IZrppPQZKX9mLzcGdPfFD1hxrcB__aEkJFOKJFd00=DUUt3SRGI0_
> JgtNaS3udV68GRkgV4ts7XKfj2opmiCY=RbF_pG6G05IcrxiELCCV36C6Cb_
> GqQZ7H84RH1hRQik=p1nuatzGn2KrF98argO7-qTt4U64Rzny3KoN-IJLOv4=
>
>Needs Jed's approval.
>
>Barry
>
>
>
> > On Mar 8, 2017, at 10:33 AM, Barry Smith  wrote:
> >
> >
> >  Please tell us how you got this output.
> >
> >  PETSc CG doesn't even implement right preconditioning. If you ask for
> it it should error out. CG supports no norm computation with left
> preconditioning.
> >
> >   Barry
> >
> >> On Mar 8, 2017, at 10:26 AM, Kong, Fande  wrote:
> >>
> >> Hi All,
> >>
> >> The NONE norm type is supported only when CG is used with a right
> preconditioner. Any reason for this?
> >>
> >>
> >>
> >> 0 Nonlinear |R| = 1.732051e+00
> >>  0 Linear |R| = 0.00e+00
> >>  1 Linear |R| = 0.00e+00
> >>  2 Linear |R| = 0.00e+00
> >>  3 Linear |R| = 0.00e+00
> >>  4 Linear |R| = 0.00e+00
> >>  5 Linear |R| = 0.00e+00
> >>  6 Linear |R| = 0.00e+00
> >> 1 Nonlinear |R| = 1.769225e-08
> >>  0 Linear |R| = 0.00e+00
> >>  1 Linear |R| = 0.00e+00
> >>  2 Linear |R| = 0.00e+00
> >>  3 Linear |R| = 0.00e+00
> >>  4 Linear |R| = 0.00e+00
> >>  5 Linear |R| = 0.00e+00
> >>  6 Linear |R| = 0.00e+00
> >>  7 Linear |R| = 0.00e+00
> >>  8 Linear |R| = 0.00e+00
> >>  9 Linear |R| = 0.00e+00
> >> 10 Linear |R| = 0.00e+00
> >> 2 Nonlinear |R| = 0.00e+00
> >> SNES Object: 1 MPI processes
> >>  type: newtonls
> >>  maximum iterations=50, maximum function evaluations=1
> >>  tolerances: relative=1e-08, absolute=1e-50, solution=1e-50
> >>  total number of linear solver iterations=18
> >>  total number of function evaluations=23
> >>  norm schedule ALWAYS
> >>  SNESLineSearch Object:   1 MPI processes
> >>type: bt
> >>  interpolation: cubic
> >>  alpha=1.00e-04
> >>maxstep=1.00e+08, minlambda=1.00e-12
> >>tolerances: relative=1.00e-08, absolute=1.00e-15,
> lambda=1.00e-08
> >>maximum iterations=40
> >>  KSP Object:   1 MPI processes
> >>type: cg
> >>maximum iterations=1, initial guess is zero
> >>tolerances:  relative=1e-05, absolute=1e-50, divergence=1.
> >>right preconditioning
> >>using NONE norm type for convergence test
> >>  PC Object:   1 MPI processes
> >>type: hypre
> >>  HYPRE BoomerAMG preconditioning
> >>  HYPRE BoomerAMG: Cycle type V
> >>  HYPRE BoomerAMG: Maximum number of levels 25
> >>  HYPRE BoomerAMG: Maximum number of iterations PER hypre call 1
> >>  HYPRE BoomerAMG: Convergence tolerance PER hypre call 0.
> >>  HYPRE BoomerAMG: Threshold for strong coupling 0.25
> >>  HYPRE BoomerAMG: Interpolation truncation factor 0.
> >>  HYPRE BoomerAMG: Interpolation: max elements per row 0
> >>  HYPRE BoomerAMG: Number of levels of aggressive coarsening 0
> >>  HYPRE BoomerAMG: Number of paths for aggressive coarsening 1
> >>  HYPRE BoomerAMG: Maximum row sums 0.9
> >>  HYPRE BoomerAMG: Sweeps down 1
> >>  HYPRE BoomerAMG: Sweeps up   1
> >>  HYPRE BoomerAMG: Sweeps on coarse1
> >>  HYPRE BoomerAMG: Relax down  symmetric-SOR/Jacobi
> >>  HYPRE BoomerAMG: Relax upsymmetric-SOR/Jacobi
> >>  HYPRE BoomerAMG: Relax on coarse Gaussian-elimination
> >>  HYPRE BoomerAMG: Relax weight  (all)  1.
> >>  HYPRE BoomerAMG: Outer relax weight (all) 1.
> >>  HYPRE BoomerAMG: Using CF-relaxation
> >>  HYPRE BoomerAMG: Not using more complex smoothers.
> >>  HYPRE BoomerAMG: Measure typelocal
> >>  HYPRE BoomerAMG: Coarsen typeFalgout
> >>  HYPRE BoomerAMG: Interpolation type  classical
> >>linear system matrix followed by preconditioner matrix:
> >>Mat Object: 1 MPI processes
> >>  type: mffd
> >>  rows=9, cols=9
> >>Matrix-free approximation:
> >>  err=1.49012e-08 (relative error in function evaluation)
> >>  Using wp compute h routine
> >>  Does not compute normU
> >>Mat Object:() 1 MPI processes
> >>  type: seqaij
> >>  rows=9, cols=9
> >>  total: nonzeros=49, allocated nonzeros=49
> >>  total number of mallocs used during MatSetValues calls =0
> >>not using I-node routines
> >>
> >> Fande,
> >>
> >
>
>

Re: [petsc-users] CG with right preconditioning supports NONE norm type only

[Barry Smith]

   A proposed fix 
https://bitbucket.org/petsc/petsc/pull-requests/645/do-not-assume-that-all-ksp-methods-support

   Needs Jed's approval.

   Barry


  
> On Mar 8, 2017, at 10:33 AM, Barry Smith  wrote:
> 
> 
>  Please tell us how you got this output. 
> 
>  PETSc CG doesn't even implement right preconditioning. If you ask for it it 
> should error out. CG supports no norm computation with left preconditioning.
> 
>   Barry
> 
>> On Mar 8, 2017, at 10:26 AM, Kong, Fande  wrote:
>> 
>> Hi All,
>> 
>> The NONE norm type is supported only when CG is used with a right 
>> preconditioner. Any reason for this?
>> 
>> 
>> 
>> 0 Nonlinear |R| = 1.732051e+00
>>  0 Linear |R| = 0.00e+00
>>  1 Linear |R| = 0.00e+00
>>  2 Linear |R| = 0.00e+00
>>  3 Linear |R| = 0.00e+00
>>  4 Linear |R| = 0.00e+00
>>  5 Linear |R| = 0.00e+00
>>  6 Linear |R| = 0.00e+00
>> 1 Nonlinear |R| = 1.769225e-08
>>  0 Linear |R| = 0.00e+00
>>  1 Linear |R| = 0.00e+00
>>  2 Linear |R| = 0.00e+00
>>  3 Linear |R| = 0.00e+00
>>  4 Linear |R| = 0.00e+00
>>  5 Linear |R| = 0.00e+00
>>  6 Linear |R| = 0.00e+00
>>  7 Linear |R| = 0.00e+00
>>  8 Linear |R| = 0.00e+00
>>  9 Linear |R| = 0.00e+00
>> 10 Linear |R| = 0.00e+00
>> 2 Nonlinear |R| = 0.00e+00
>> SNES Object: 1 MPI processes
>>  type: newtonls
>>  maximum iterations=50, maximum function evaluations=1
>>  tolerances: relative=1e-08, absolute=1e-50, solution=1e-50
>>  total number of linear solver iterations=18
>>  total number of function evaluations=23
>>  norm schedule ALWAYS
>>  SNESLineSearch Object:   1 MPI processes
>>type: bt
>>  interpolation: cubic
>>  alpha=1.00e-04
>>maxstep=1.00e+08, minlambda=1.00e-12
>>tolerances: relative=1.00e-08, absolute=1.00e-15, 
>> lambda=1.00e-08
>>maximum iterations=40
>>  KSP Object:   1 MPI processes
>>type: cg
>>maximum iterations=1, initial guess is zero
>>tolerances:  relative=1e-05, absolute=1e-50, divergence=1.
>>right preconditioning
>>using NONE norm type for convergence test
>>  PC Object:   1 MPI processes
>>type: hypre
>>  HYPRE BoomerAMG preconditioning
>>  HYPRE BoomerAMG: Cycle type V
>>  HYPRE BoomerAMG: Maximum number of levels 25
>>  HYPRE BoomerAMG: Maximum number of iterations PER hypre call 1
>>  HYPRE BoomerAMG: Convergence tolerance PER hypre call 0.
>>  HYPRE BoomerAMG: Threshold for strong coupling 0.25
>>  HYPRE BoomerAMG: Interpolation truncation factor 0.
>>  HYPRE BoomerAMG: Interpolation: max elements per row 0
>>  HYPRE BoomerAMG: Number of levels of aggressive coarsening 0
>>  HYPRE BoomerAMG: Number of paths for aggressive coarsening 1
>>  HYPRE BoomerAMG: Maximum row sums 0.9
>>  HYPRE BoomerAMG: Sweeps down 1
>>  HYPRE BoomerAMG: Sweeps up   1
>>  HYPRE BoomerAMG: Sweeps on coarse1
>>  HYPRE BoomerAMG: Relax down  symmetric-SOR/Jacobi
>>  HYPRE BoomerAMG: Relax upsymmetric-SOR/Jacobi
>>  HYPRE BoomerAMG: Relax on coarse Gaussian-elimination
>>  HYPRE BoomerAMG: Relax weight  (all)  1.
>>  HYPRE BoomerAMG: Outer relax weight (all) 1.
>>  HYPRE BoomerAMG: Using CF-relaxation
>>  HYPRE BoomerAMG: Not using more complex smoothers.
>>  HYPRE BoomerAMG: Measure typelocal
>>  HYPRE BoomerAMG: Coarsen typeFalgout
>>  HYPRE BoomerAMG: Interpolation type  classical
>>linear system matrix followed by preconditioner matrix:
>>Mat Object: 1 MPI processes
>>  type: mffd
>>  rows=9, cols=9
>>Matrix-free approximation:
>>  err=1.49012e-08 (relative error in function evaluation)
>>  Using wp compute h routine
>>  Does not compute normU
>>Mat Object:() 1 MPI processes
>>  type: seqaij
>>  rows=9, cols=9
>>  total: nonzeros=49, allocated nonzeros=49
>>  total number of mallocs used during MatSetValues calls =0
>>not using I-node routines
>> 
>> Fande,
>> 
> 

Re: [petsc-users] CG with right preconditioning supports NONE norm type only

[Barry Smith]

   Jed,

 This seems wrong. It is called in KSPCreate() and seems to imply all KSP 
methods support no-norm with right preconditioning (or left for that matter). 
But CG doesn't support right period and FGMES does not support left at all.
Shouldn't those two lines be removed (and maybe they need to be added in the 
create for certain KSP methods).

PetscErrorCode KSPNormSupportTableReset_Private(KSP ksp)
{
  PetscErrorCode ierr;

  PetscFunctionBegin;
  ierr = 
PetscMemzero(ksp->normsupporttable,sizeof(ksp->normsupporttable));CHKERRQ(ierr);
  ierr = KSPSetSupportedNorm(ksp,KSP_NORM_NONE,PC_LEFT,1);CHKERRQ(ierr);
  ierr = KSPSetSupportedNorm(ksp,KSP_NORM_NONE,PC_RIGHT,1);CHKERRQ(ierr);
  ksp->pc_side  = ksp->pc_side_set;
  ksp->normtype = ksp->normtype_set;
  PetscFunctionReturn(0);
}



> On Mar 8, 2017, at 11:19 AM, Matthew Knepley  wrote:
> 
> On Wed, Mar 8, 2017 at 10:47 AM, Kong, Fande  wrote:
> Thanks Barry,
> 
> We are using "KSPSetPCSide(ksp, pcside)" in the code.  I just tried 
> "-ksp_pc_side right", and petsc did not error out. 
> 
> I like to understand why CG does not work with right preconditioning? 
> Mathematically, the right preconditioning does not make sense?
> 
> cd src/snes/examples/tutorials
> knepley/feature-plasma-example 
> $:/PETSc3/petsc/petsc-dev/src/snes/examples/tutorials$ ./ex5 -ksp_view 
> -ksp_type cg -ksp_pc_side right -ksp_error_if_not_converged
> KSP Object: 1 MPI processes
>   type: cg
>   maximum iterations=1, initial guess is zero
>   tolerances:  relative=1e-05, absolute=1e-50, divergence=1.
>   right preconditioning
>   using NONE norm type for convergence test
> PC Object: 1 MPI processes
>   type: ilu
> out-of-place factorization
> 0 levels of fill
> tolerance for zero pivot 2.22045e-14
> matrix ordering: natural
> factor fill ratio given 1., needed 1.
>   Factored matrix follows:
> Mat Object: 1 MPI processes
>   type: seqaij
>   rows=16, cols=16
>   package used to perform factorization: petsc
>   total: nonzeros=64, allocated nonzeros=64
>   total number of mallocs used during MatSetValues calls =0
> not using I-node routines
>   linear system matrix = precond matrix:
>   Mat Object: 1 MPI processes
> type: seqaij
> rows=16, cols=16
> total: nonzeros=64, allocated nonzeros=64
> total number of mallocs used during MatSetValues calls =0
>   not using I-node routines
> [0]PETSC ERROR: - Error Message 
> --
> [0]PETSC ERROR:   
> [0]PETSC ERROR: KSPSolve has not converged
> [0]PETSC ERROR: See http://www.mcs.anl.gov/petsc/documentation/faq.html for 
> trouble shooting.
> [0]PETSC ERROR: Petsc Development GIT revision: v3.7.5-3127-ge9f6087  GIT 
> Date: 2017-02-11 13:06:34 -0600
> [0]PETSC ERROR: ./ex5 on a arch-c-exodus-master named 
> MATTHEW-KNEPLEYs-MacBook-Air-2.local by knepley Wed Mar  8 11:17:43 2017
> [0]PETSC ERROR: Configure options --PETSC_ARCH=arch-c-exodus-master 
> --download-chaco --download-cmake --download-ctetgen --download-eigen 
> --download-exodusii --download-gmp --download-hdf5 --download-metis 
> --download-mpfr --download-mpich --download-netcdf --download-p4est 
> --download-parmetis --download-pragmatic --download-triangle --useThreads=1 
> --with-cc="/Users/knepley/MacSoftware/bin/ccache gcc -Qunused-arguments" 
> --with-cxx="/Users/knepley/MacSoftware/bin/ccache g++ -Qunused-arguments" 
> --with-fc="/Users/knepley/MacSoftware/bin/ccache gfortran" 
> --with-shared-libraries
> [0]PETSC ERROR: #1 KSPSolve() line 847 in 
> /PETSc3/petsc/petsc-dev/src/ksp/ksp/interface/itfunc.c
> [0]PETSC ERROR: #2 SNESSolve_NEWTONLS() line 224 in 
> /PETSc3/petsc/petsc-dev/src/snes/impls/ls/ls.c
> [0]PETSC ERROR: #3 SNESSolve() line 3967 in 
> /PETSc3/petsc/petsc-dev/src/snes/interface/snes.c
> [0]PETSC ERROR: #4 main() line 187 in 
> /PETSc3/petsc/petsc-dev/src/snes/examples/tutorials/ex5.c
> [0]PETSC ERROR: PETSc Option Table entries:
> [0]PETSC ERROR: -ksp_error_if_not_converged
> [0]PETSC ERROR: -ksp_pc_side right
> [0]PETSC ERROR: -ksp_type cg
> [0]PETSC ERROR: -ksp_view
> [0]PETSC ERROR: -malloc_test
> [0]PETSC ERROR: End of Error Message ---send entire error 
> message to petsc-ma...@mcs.anl.gov--
> 
> So we are not getting an error
> 
>   Matt
>  
> 
> Fande,
> 
> On Wed, Mar 8, 2017 at 9:33 AM, Barry Smith  wrote:
> 
>   Please tell us how you got this output.
> 
>   PETSc CG doesn't even implement right preconditioning. If you ask for it it 
> should error out. CG supports no norm computation with left preconditioning.
> 
>Barry
> 
> > On Mar 8, 2017, at 10:26 AM, Kong, Fande  wrote:
> >
> > Hi All,
> >
> > The NONE norm type is supported only when CG is used with a right 
> > preconditioner. Any reason for this?
> >
> >
> >
> > 0 Nonlinear |R| 

Re: [petsc-users] CG with right preconditioning supports NONE norm type only

[Barry Smith]

> On Mar 8, 2017, at 10:47 AM, Kong, Fande  wrote:
> 
> Thanks Barry,
> 
> We are using "KSPSetPCSide(ksp, pcside)" in the code.  I just tried 
> "-ksp_pc_side right", and petsc did not error out. 
> 
> I like to understand why CG does not work with right preconditioning? 
> Mathematically, the right preconditioning does not make sense?

   No, mathematically it makes sense to do it on the right. It is just that the 
PETSc code was never written to support it on the right. One reason is that CG 
is interesting that you can run with the true residual or the preconditioned 
residual with left preconditioning, hence less incentive to ever bother writing 
it to support right preconditioning. For completeness we should support right 
as well as symmetric.

  Barry

> 
> Fande,
> 
> On Wed, Mar 8, 2017 at 9:33 AM, Barry Smith  wrote:
> 
>   Please tell us how you got this output.
> 
>   PETSc CG doesn't even implement right preconditioning. If you ask for it it 
> should error out. CG supports no norm computation with left preconditioning.
> 
>Barry
> 
> > On Mar 8, 2017, at 10:26 AM, Kong, Fande  wrote:
> >
> > Hi All,
> >
> > The NONE norm type is supported only when CG is used with a right 
> > preconditioner. Any reason for this?
> >
> >
> >
> > 0 Nonlinear |R| = 1.732051e+00
> >   0 Linear |R| = 0.00e+00
> >   1 Linear |R| = 0.00e+00
> >   2 Linear |R| = 0.00e+00
> >   3 Linear |R| = 0.00e+00
> >   4 Linear |R| = 0.00e+00
> >   5 Linear |R| = 0.00e+00
> >   6 Linear |R| = 0.00e+00
> >  1 Nonlinear |R| = 1.769225e-08
> >   0 Linear |R| = 0.00e+00
> >   1 Linear |R| = 0.00e+00
> >   2 Linear |R| = 0.00e+00
> >   3 Linear |R| = 0.00e+00
> >   4 Linear |R| = 0.00e+00
> >   5 Linear |R| = 0.00e+00
> >   6 Linear |R| = 0.00e+00
> >   7 Linear |R| = 0.00e+00
> >   8 Linear |R| = 0.00e+00
> >   9 Linear |R| = 0.00e+00
> >  10 Linear |R| = 0.00e+00
> >  2 Nonlinear |R| = 0.00e+00
> > SNES Object: 1 MPI processes
> >   type: newtonls
> >   maximum iterations=50, maximum function evaluations=1
> >   tolerances: relative=1e-08, absolute=1e-50, solution=1e-50
> >   total number of linear solver iterations=18
> >   total number of function evaluations=23
> >   norm schedule ALWAYS
> >   SNESLineSearch Object:   1 MPI processes
> > type: bt
> >   interpolation: cubic
> >   alpha=1.00e-04
> > maxstep=1.00e+08, minlambda=1.00e-12
> > tolerances: relative=1.00e-08, absolute=1.00e-15, 
> > lambda=1.00e-08
> > maximum iterations=40
> >   KSP Object:   1 MPI processes
> > type: cg
> > maximum iterations=1, initial guess is zero
> > tolerances:  relative=1e-05, absolute=1e-50, divergence=1.
> > right preconditioning
> > using NONE norm type for convergence test
> >   PC Object:   1 MPI processes
> > type: hypre
> >   HYPRE BoomerAMG preconditioning
> >   HYPRE BoomerAMG: Cycle type V
> >   HYPRE BoomerAMG: Maximum number of levels 25
> >   HYPRE BoomerAMG: Maximum number of iterations PER hypre call 1
> >   HYPRE BoomerAMG: Convergence tolerance PER hypre call 0.
> >   HYPRE BoomerAMG: Threshold for strong coupling 0.25
> >   HYPRE BoomerAMG: Interpolation truncation factor 0.
> >   HYPRE BoomerAMG: Interpolation: max elements per row 0
> >   HYPRE BoomerAMG: Number of levels of aggressive coarsening 0
> >   HYPRE BoomerAMG: Number of paths for aggressive coarsening 1
> >   HYPRE BoomerAMG: Maximum row sums 0.9
> >   HYPRE BoomerAMG: Sweeps down 1
> >   HYPRE BoomerAMG: Sweeps up   1
> >   HYPRE BoomerAMG: Sweeps on coarse1
> >   HYPRE BoomerAMG: Relax down  symmetric-SOR/Jacobi
> >   HYPRE BoomerAMG: Relax upsymmetric-SOR/Jacobi
> >   HYPRE BoomerAMG: Relax on coarse Gaussian-elimination
> >   HYPRE BoomerAMG: Relax weight  (all)  1.
> >   HYPRE BoomerAMG: Outer relax weight (all) 1.
> >   HYPRE BoomerAMG: Using CF-relaxation
> >   HYPRE BoomerAMG: Not using more complex smoothers.
> >   HYPRE BoomerAMG: Measure typelocal
> >   HYPRE BoomerAMG: Coarsen typeFalgout
> >   HYPRE BoomerAMG: Interpolation type  classical
> > linear system matrix followed by preconditioner matrix:
> > Mat Object: 1 MPI processes
> >   type: mffd
> >   rows=9, cols=9
> > Matrix-free approximation:
> >   err=1.49012e-08 (relative error in function evaluation)
> >   Using wp compute h routine
> >   Does not compute normU
> > Mat Object:() 1 MPI processes
> >   type: seqaij
> >   rows=9, cols=9
> >   total: nonzeros=49, allocated nonzeros=49
> >   total number of mallocs used during MatSetValues calls =0
> > not 

Re: [petsc-users] CG with right preconditioning supports NONE norm type only

[Barry Smith]

  We should be getting an error.


> On Mar 8, 2017, at 11:19 AM, Matthew Knepley  wrote:
> 
> On Wed, Mar 8, 2017 at 10:47 AM, Kong, Fande  wrote:
> Thanks Barry,
> 
> We are using "KSPSetPCSide(ksp, pcside)" in the code.  I just tried 
> "-ksp_pc_side right", and petsc did not error out. 
> 
> I like to understand why CG does not work with right preconditioning? 
> Mathematically, the right preconditioning does not make sense?
> 
> cd src/snes/examples/tutorials
> knepley/feature-plasma-example 
> $:/PETSc3/petsc/petsc-dev/src/snes/examples/tutorials$ ./ex5 -ksp_view 
> -ksp_type cg -ksp_pc_side right -ksp_error_if_not_converged
> KSP Object: 1 MPI processes
>   type: cg
>   maximum iterations=1, initial guess is zero
>   tolerances:  relative=1e-05, absolute=1e-50, divergence=1.
>   right preconditioning
>   using NONE norm type for convergence test
> PC Object: 1 MPI processes
>   type: ilu
> out-of-place factorization
> 0 levels of fill
> tolerance for zero pivot 2.22045e-14
> matrix ordering: natural
> factor fill ratio given 1., needed 1.
>   Factored matrix follows:
> Mat Object: 1 MPI processes
>   type: seqaij
>   rows=16, cols=16
>   package used to perform factorization: petsc
>   total: nonzeros=64, allocated nonzeros=64
>   total number of mallocs used during MatSetValues calls =0
> not using I-node routines
>   linear system matrix = precond matrix:
>   Mat Object: 1 MPI processes
> type: seqaij
> rows=16, cols=16
> total: nonzeros=64, allocated nonzeros=64
> total number of mallocs used during MatSetValues calls =0
>   not using I-node routines
> [0]PETSC ERROR: - Error Message 
> --
> [0]PETSC ERROR:   
> [0]PETSC ERROR: KSPSolve has not converged
> [0]PETSC ERROR: See http://www.mcs.anl.gov/petsc/documentation/faq.html for 
> trouble shooting.
> [0]PETSC ERROR: Petsc Development GIT revision: v3.7.5-3127-ge9f6087  GIT 
> Date: 2017-02-11 13:06:34 -0600
> [0]PETSC ERROR: ./ex5 on a arch-c-exodus-master named 
> MATTHEW-KNEPLEYs-MacBook-Air-2.local by knepley Wed Mar  8 11:17:43 2017
> [0]PETSC ERROR: Configure options --PETSC_ARCH=arch-c-exodus-master 
> --download-chaco --download-cmake --download-ctetgen --download-eigen 
> --download-exodusii --download-gmp --download-hdf5 --download-metis 
> --download-mpfr --download-mpich --download-netcdf --download-p4est 
> --download-parmetis --download-pragmatic --download-triangle --useThreads=1 
> --with-cc="/Users/knepley/MacSoftware/bin/ccache gcc -Qunused-arguments" 
> --with-cxx="/Users/knepley/MacSoftware/bin/ccache g++ -Qunused-arguments" 
> --with-fc="/Users/knepley/MacSoftware/bin/ccache gfortran" 
> --with-shared-libraries
> [0]PETSC ERROR: #1 KSPSolve() line 847 in 
> /PETSc3/petsc/petsc-dev/src/ksp/ksp/interface/itfunc.c
> [0]PETSC ERROR: #2 SNESSolve_NEWTONLS() line 224 in 
> /PETSc3/petsc/petsc-dev/src/snes/impls/ls/ls.c
> [0]PETSC ERROR: #3 SNESSolve() line 3967 in 
> /PETSc3/petsc/petsc-dev/src/snes/interface/snes.c
> [0]PETSC ERROR: #4 main() line 187 in 
> /PETSc3/petsc/petsc-dev/src/snes/examples/tutorials/ex5.c
> [0]PETSC ERROR: PETSc Option Table entries:
> [0]PETSC ERROR: -ksp_error_if_not_converged
> [0]PETSC ERROR: -ksp_pc_side right
> [0]PETSC ERROR: -ksp_type cg
> [0]PETSC ERROR: -ksp_view
> [0]PETSC ERROR: -malloc_test
> [0]PETSC ERROR: End of Error Message ---send entire error 
> message to petsc-ma...@mcs.anl.gov--
> 
> So we are not getting an error
> 
>   Matt
>  
> 
> Fande,
> 
> On Wed, Mar 8, 2017 at 9:33 AM, Barry Smith  wrote:
> 
>   Please tell us how you got this output.
> 
>   PETSc CG doesn't even implement right preconditioning. If you ask for it it 
> should error out. CG supports no norm computation with left preconditioning.
> 
>Barry
> 
> > On Mar 8, 2017, at 10:26 AM, Kong, Fande  wrote:
> >
> > Hi All,
> >
> > The NONE norm type is supported only when CG is used with a right 
> > preconditioner. Any reason for this?
> >
> >
> >
> > 0 Nonlinear |R| = 1.732051e+00
> >   0 Linear |R| = 0.00e+00
> >   1 Linear |R| = 0.00e+00
> >   2 Linear |R| = 0.00e+00
> >   3 Linear |R| = 0.00e+00
> >   4 Linear |R| = 0.00e+00
> >   5 Linear |R| = 0.00e+00
> >   6 Linear |R| = 0.00e+00
> >  1 Nonlinear |R| = 1.769225e-08
> >   0 Linear |R| = 0.00e+00
> >   1 Linear |R| = 0.00e+00
> >   2 Linear |R| = 0.00e+00
> >   3 Linear |R| = 0.00e+00
> >   4 Linear |R| = 0.00e+00
> >   5 Linear |R| = 0.00e+00
> >   6 Linear |R| = 0.00e+00
> >   7 Linear |R| = 0.00e+00
> >   8 Linear |R| = 0.00e+00
> >   9 Linear |R| = 0.00e+00
> >  10 Linear |R| = 0.00e+00
> >  2 Nonlinear |R| = 

Re: [petsc-users] CG with right preconditioning supports NONE norm type only

[Matthew Knepley]

On Wed, Mar 8, 2017 at 10:47 AM, Kong, Fande  wrote:

> Thanks Barry,
>
> We are using "KSPSetPCSide(ksp, pcside)" in the code.  I just tried
> "-ksp_pc_side right", and petsc did not error out.
>
> I like to understand why CG does not work with right preconditioning?
> Mathematically, the right preconditioning does not make sense?
>

cd src/snes/examples/tutorials
knepley/feature-plasma-example
$:/PETSc3/petsc/petsc-dev/src/snes/examples/tutorials$ ./ex5 -ksp_view
-ksp_type cg -ksp_pc_side right -ksp_error_if_not_converged
KSP Object: 1 MPI processes
  type: cg
  maximum iterations=1, initial guess is zero
  tolerances:  relative=1e-05, absolute=1e-50, divergence=1.
  right preconditioning
  using NONE norm type for convergence test
PC Object: 1 MPI processes
  type: ilu
out-of-place factorization
0 levels of fill
tolerance for zero pivot 2.22045e-14
matrix ordering: natural
factor fill ratio given 1., needed 1.
  Factored matrix follows:
Mat Object: 1 MPI processes
  type: seqaij
  rows=16, cols=16
  package used to perform factorization: petsc
  total: nonzeros=64, allocated nonzeros=64
  total number of mallocs used during MatSetValues calls =0
not using I-node routines
  linear system matrix = precond matrix:
  Mat Object: 1 MPI processes
type: seqaij
rows=16, cols=16
total: nonzeros=64, allocated nonzeros=64
total number of mallocs used during MatSetValues calls =0
  not using I-node routines
[0]PETSC ERROR: - Error Message
--
[0]PETSC ERROR:
[0]PETSC ERROR: KSPSolve has not converged
[0]PETSC ERROR: See http://www.mcs.anl.gov/petsc/documentation/faq.html for
trouble shooting.
[0]PETSC ERROR: Petsc Development GIT revision: v3.7.5-3127-ge9f6087  GIT
Date: 2017-02-11 13:06:34 -0600
[0]PETSC ERROR: ./ex5 on a arch-c-exodus-master named
MATTHEW-KNEPLEYs-MacBook-Air-2.local by knepley Wed Mar  8 11:17:43 2017
[0]PETSC ERROR: Configure options --PETSC_ARCH=arch-c-exodus-master
--download-chaco --download-cmake --download-ctetgen --download-eigen
--download-exodusii --download-gmp --download-hdf5 --download-metis
--download-mpfr --download-mpich --download-netcdf --download-p4est
--download-parmetis --download-pragmatic --download-triangle --useThreads=1
--with-cc="/Users/knepley/MacSoftware/bin/ccache gcc -Qunused-arguments"
--with-cxx="/Users/knepley/MacSoftware/bin/ccache g++ -Qunused-arguments"
--with-fc="/Users/knepley/MacSoftware/bin/ccache gfortran"
--with-shared-libraries
[0]PETSC ERROR: #1 KSPSolve() line 847 in
/PETSc3/petsc/petsc-dev/src/ksp/ksp/interface/itfunc.c
[0]PETSC ERROR: #2 SNESSolve_NEWTONLS() line 224 in
/PETSc3/petsc/petsc-dev/src/snes/impls/ls/ls.c
[0]PETSC ERROR: #3 SNESSolve() line 3967 in
/PETSc3/petsc/petsc-dev/src/snes/interface/snes.c
[0]PETSC ERROR: #4 main() line 187 in
/PETSc3/petsc/petsc-dev/src/snes/examples/tutorials/ex5.c
[0]PETSC ERROR: PETSc Option Table entries:
[0]PETSC ERROR: -ksp_error_if_not_converged
[0]PETSC ERROR: -ksp_pc_side right
[0]PETSC ERROR: -ksp_type cg
[0]PETSC ERROR: -ksp_view
[0]PETSC ERROR: -malloc_test
[0]PETSC ERROR: End of Error Message ---send entire
error message to petsc-ma...@mcs.anl.gov--

So we are not getting an error

  Matt


>
> Fande,
>
> On Wed, Mar 8, 2017 at 9:33 AM, Barry Smith  wrote:
>
>>
>>   Please tell us how you got this output.
>>
>>   PETSc CG doesn't even implement right preconditioning. If you ask for
>> it it should error out. CG supports no norm computation with left
>> preconditioning.
>>
>>Barry
>>
>> > On Mar 8, 2017, at 10:26 AM, Kong, Fande  wrote:
>> >
>> > Hi All,
>> >
>> > The NONE norm type is supported only when CG is used with a right
>> preconditioner. Any reason for this?
>> >
>> >
>> >
>> > 0 Nonlinear |R| = 1.732051e+00
>> >   0 Linear |R| = 0.00e+00
>> >   1 Linear |R| = 0.00e+00
>> >   2 Linear |R| = 0.00e+00
>> >   3 Linear |R| = 0.00e+00
>> >   4 Linear |R| = 0.00e+00
>> >   5 Linear |R| = 0.00e+00
>> >   6 Linear |R| = 0.00e+00
>> >  1 Nonlinear |R| = 1.769225e-08
>> >   0 Linear |R| = 0.00e+00
>> >   1 Linear |R| = 0.00e+00
>> >   2 Linear |R| = 0.00e+00
>> >   3 Linear |R| = 0.00e+00
>> >   4 Linear |R| = 0.00e+00
>> >   5 Linear |R| = 0.00e+00
>> >   6 Linear |R| = 0.00e+00
>> >   7 Linear |R| = 0.00e+00
>> >   8 Linear |R| = 0.00e+00
>> >   9 Linear |R| = 0.00e+00
>> >  10 Linear |R| = 0.00e+00
>> >  2 Nonlinear |R| = 0.00e+00
>> > SNES Object: 1 MPI processes
>> >   type: newtonls
>> >   maximum iterations=50, maximum function evaluations=1
>> >   tolerances: relative=1e-08, absolute=1e-50, solution=1e-50
>> >   total number of linear solver iterations=18
>> >   

Re: [petsc-users] CG with right preconditioning supports NONE norm type only

[Kong, Fande]

Thanks Barry,

We are using "KSPSetPCSide(ksp, pcside)" in the code.  I just tried
"-ksp_pc_side right", and petsc did not error out.

I like to understand why CG does not work with right preconditioning?
Mathematically, the right preconditioning does not make sense?

Fande,

On Wed, Mar 8, 2017 at 9:33 AM, Barry Smith  wrote:

>
>   Please tell us how you got this output.
>
>   PETSc CG doesn't even implement right preconditioning. If you ask for it
> it should error out. CG supports no norm computation with left
> preconditioning.
>
>Barry
>
> > On Mar 8, 2017, at 10:26 AM, Kong, Fande  wrote:
> >
> > Hi All,
> >
> > The NONE norm type is supported only when CG is used with a right
> preconditioner. Any reason for this?
> >
> >
> >
> > 0 Nonlinear |R| = 1.732051e+00
> >   0 Linear |R| = 0.00e+00
> >   1 Linear |R| = 0.00e+00
> >   2 Linear |R| = 0.00e+00
> >   3 Linear |R| = 0.00e+00
> >   4 Linear |R| = 0.00e+00
> >   5 Linear |R| = 0.00e+00
> >   6 Linear |R| = 0.00e+00
> >  1 Nonlinear |R| = 1.769225e-08
> >   0 Linear |R| = 0.00e+00
> >   1 Linear |R| = 0.00e+00
> >   2 Linear |R| = 0.00e+00
> >   3 Linear |R| = 0.00e+00
> >   4 Linear |R| = 0.00e+00
> >   5 Linear |R| = 0.00e+00
> >   6 Linear |R| = 0.00e+00
> >   7 Linear |R| = 0.00e+00
> >   8 Linear |R| = 0.00e+00
> >   9 Linear |R| = 0.00e+00
> >  10 Linear |R| = 0.00e+00
> >  2 Nonlinear |R| = 0.00e+00
> > SNES Object: 1 MPI processes
> >   type: newtonls
> >   maximum iterations=50, maximum function evaluations=1
> >   tolerances: relative=1e-08, absolute=1e-50, solution=1e-50
> >   total number of linear solver iterations=18
> >   total number of function evaluations=23
> >   norm schedule ALWAYS
> >   SNESLineSearch Object:   1 MPI processes
> > type: bt
> >   interpolation: cubic
> >   alpha=1.00e-04
> > maxstep=1.00e+08, minlambda=1.00e-12
> > tolerances: relative=1.00e-08, absolute=1.00e-15,
> lambda=1.00e-08
> > maximum iterations=40
> >   KSP Object:   1 MPI processes
> > type: cg
> > maximum iterations=1, initial guess is zero
> > tolerances:  relative=1e-05, absolute=1e-50, divergence=1.
> > right preconditioning
> > using NONE norm type for convergence test
> >   PC Object:   1 MPI processes
> > type: hypre
> >   HYPRE BoomerAMG preconditioning
> >   HYPRE BoomerAMG: Cycle type V
> >   HYPRE BoomerAMG: Maximum number of levels 25
> >   HYPRE BoomerAMG: Maximum number of iterations PER hypre call 1
> >   HYPRE BoomerAMG: Convergence tolerance PER hypre call 0.
> >   HYPRE BoomerAMG: Threshold for strong coupling 0.25
> >   HYPRE BoomerAMG: Interpolation truncation factor 0.
> >   HYPRE BoomerAMG: Interpolation: max elements per row 0
> >   HYPRE BoomerAMG: Number of levels of aggressive coarsening 0
> >   HYPRE BoomerAMG: Number of paths for aggressive coarsening 1
> >   HYPRE BoomerAMG: Maximum row sums 0.9
> >   HYPRE BoomerAMG: Sweeps down 1
> >   HYPRE BoomerAMG: Sweeps up   1
> >   HYPRE BoomerAMG: Sweeps on coarse1
> >   HYPRE BoomerAMG: Relax down  symmetric-SOR/Jacobi
> >   HYPRE BoomerAMG: Relax upsymmetric-SOR/Jacobi
> >   HYPRE BoomerAMG: Relax on coarse Gaussian-elimination
> >   HYPRE BoomerAMG: Relax weight  (all)  1.
> >   HYPRE BoomerAMG: Outer relax weight (all) 1.
> >   HYPRE BoomerAMG: Using CF-relaxation
> >   HYPRE BoomerAMG: Not using more complex smoothers.
> >   HYPRE BoomerAMG: Measure typelocal
> >   HYPRE BoomerAMG: Coarsen typeFalgout
> >   HYPRE BoomerAMG: Interpolation type  classical
> > linear system matrix followed by preconditioner matrix:
> > Mat Object: 1 MPI processes
> >   type: mffd
> >   rows=9, cols=9
> > Matrix-free approximation:
> >   err=1.49012e-08 (relative error in function evaluation)
> >   Using wp compute h routine
> >   Does not compute normU
> > Mat Object:() 1 MPI processes
> >   type: seqaij
> >   rows=9, cols=9
> >   total: nonzeros=49, allocated nonzeros=49
> >   total number of mallocs used during MatSetValues calls =0
> > not using I-node routines
> >
> > Fande,
> >
>
>

Re: [petsc-users] CG with right preconditioning supports NONE norm type only

[Barry Smith]

  Please tell us how you got this output. 

  PETSc CG doesn't even implement right preconditioning. If you ask for it it 
should error out. CG supports no norm computation with left preconditioning.

   Barry

> On Mar 8, 2017, at 10:26 AM, Kong, Fande  wrote:
> 
> Hi All,
> 
> The NONE norm type is supported only when CG is used with a right 
> preconditioner. Any reason for this?
> 
> 
> 
> 0 Nonlinear |R| = 1.732051e+00
>   0 Linear |R| = 0.00e+00
>   1 Linear |R| = 0.00e+00
>   2 Linear |R| = 0.00e+00
>   3 Linear |R| = 0.00e+00
>   4 Linear |R| = 0.00e+00
>   5 Linear |R| = 0.00e+00
>   6 Linear |R| = 0.00e+00
>  1 Nonlinear |R| = 1.769225e-08
>   0 Linear |R| = 0.00e+00
>   1 Linear |R| = 0.00e+00
>   2 Linear |R| = 0.00e+00
>   3 Linear |R| = 0.00e+00
>   4 Linear |R| = 0.00e+00
>   5 Linear |R| = 0.00e+00
>   6 Linear |R| = 0.00e+00
>   7 Linear |R| = 0.00e+00
>   8 Linear |R| = 0.00e+00
>   9 Linear |R| = 0.00e+00
>  10 Linear |R| = 0.00e+00
>  2 Nonlinear |R| = 0.00e+00
> SNES Object: 1 MPI processes
>   type: newtonls
>   maximum iterations=50, maximum function evaluations=1
>   tolerances: relative=1e-08, absolute=1e-50, solution=1e-50
>   total number of linear solver iterations=18
>   total number of function evaluations=23
>   norm schedule ALWAYS
>   SNESLineSearch Object:   1 MPI processes
> type: bt
>   interpolation: cubic
>   alpha=1.00e-04
> maxstep=1.00e+08, minlambda=1.00e-12
> tolerances: relative=1.00e-08, absolute=1.00e-15, 
> lambda=1.00e-08
> maximum iterations=40
>   KSP Object:   1 MPI processes
> type: cg
> maximum iterations=1, initial guess is zero
> tolerances:  relative=1e-05, absolute=1e-50, divergence=1.
> right preconditioning
> using NONE norm type for convergence test
>   PC Object:   1 MPI processes
> type: hypre
>   HYPRE BoomerAMG preconditioning
>   HYPRE BoomerAMG: Cycle type V
>   HYPRE BoomerAMG: Maximum number of levels 25
>   HYPRE BoomerAMG: Maximum number of iterations PER hypre call 1
>   HYPRE BoomerAMG: Convergence tolerance PER hypre call 0.
>   HYPRE BoomerAMG: Threshold for strong coupling 0.25
>   HYPRE BoomerAMG: Interpolation truncation factor 0.
>   HYPRE BoomerAMG: Interpolation: max elements per row 0
>   HYPRE BoomerAMG: Number of levels of aggressive coarsening 0
>   HYPRE BoomerAMG: Number of paths for aggressive coarsening 1
>   HYPRE BoomerAMG: Maximum row sums 0.9
>   HYPRE BoomerAMG: Sweeps down 1
>   HYPRE BoomerAMG: Sweeps up   1
>   HYPRE BoomerAMG: Sweeps on coarse1
>   HYPRE BoomerAMG: Relax down  symmetric-SOR/Jacobi
>   HYPRE BoomerAMG: Relax upsymmetric-SOR/Jacobi
>   HYPRE BoomerAMG: Relax on coarse Gaussian-elimination
>   HYPRE BoomerAMG: Relax weight  (all)  1.
>   HYPRE BoomerAMG: Outer relax weight (all) 1.
>   HYPRE BoomerAMG: Using CF-relaxation
>   HYPRE BoomerAMG: Not using more complex smoothers.
>   HYPRE BoomerAMG: Measure typelocal
>   HYPRE BoomerAMG: Coarsen typeFalgout
>   HYPRE BoomerAMG: Interpolation type  classical
> linear system matrix followed by preconditioner matrix:
> Mat Object: 1 MPI processes
>   type: mffd
>   rows=9, cols=9
> Matrix-free approximation:
>   err=1.49012e-08 (relative error in function evaluation)
>   Using wp compute h routine
>   Does not compute normU
> Mat Object:() 1 MPI processes
>   type: seqaij
>   rows=9, cols=9
>   total: nonzeros=49, allocated nonzeros=49
>   total number of mallocs used during MatSetValues calls =0
> not using I-node routines
> 
> Fande,
> 

[petsc-users] CG with right preconditioning supports NONE norm type only

[Kong, Fande]

Hi All,

The NONE norm type is supported only when CG is used with a right
preconditioner. Any reason for this?



















































































*0 Nonlinear |R| = 1.732051e+00  0 Linear |R| = 0.00e+00  1
Linear |R| = 0.00e+00  2 Linear |R| = 0.00e+00  3 Linear
|R| = 0.00e+00  4 Linear |R| = 0.00e+00  5 Linear |R| =
0.00e+00  6 Linear |R| = 0.00e+00 1 Nonlinear |R| =
1.769225e-08  0 Linear |R| = 0.00e+00  1 Linear |R| =
0.00e+00  2 Linear |R| = 0.00e+00  3 Linear |R| =
0.00e+00  4 Linear |R| = 0.00e+00  5 Linear |R| =
0.00e+00  6 Linear |R| = 0.00e+00  7 Linear |R| =
0.00e+00  8 Linear |R| = 0.00e+00  9 Linear |R| =
0.00e+00 10 Linear |R| = 0.00e+00 2 Nonlinear |R| =
0.00e+00SNES Object: 1 MPI processes  type: newtonls  maximum
iterations=50, maximum function evaluations=1  tolerances:
relative=1e-08, absolute=1e-50, solution=1e-50  total number of linear
solver iterations=18  total number of function evaluations=23  norm
schedule ALWAYS  SNESLineSearch Object:   1 MPI processestype: bt
interpolation: cubic  alpha=1.00e-04maxstep=1.00e+08,
minlambda=1.00e-12tolerances: relative=1.00e-08,
absolute=1.00e-15, lambda=1.00e-08maximum iterations=40  KSP
Object:   1 MPI processestype: cgmaximum iterations=1, initial
guess is zerotolerances:  relative=1e-05, absolute=1e-50,
divergence=1.right preconditioningusing NONE norm type for
convergence test  PC Object:   1 MPI processestype: hypre  HYPRE
BoomerAMG preconditioning  HYPRE BoomerAMG: Cycle type V  HYPRE
BoomerAMG: Maximum number of levels 25  HYPRE BoomerAMG: Maximum number
of iterations PER hypre call 1  HYPRE BoomerAMG: Convergence tolerance
PER hypre call 0.  HYPRE BoomerAMG: Threshold for strong coupling
0.25  HYPRE BoomerAMG: Interpolation truncation factor 0.  HYPRE
BoomerAMG: Interpolation: max elements per row 0  HYPRE BoomerAMG:
Number of levels of aggressive coarsening 0  HYPRE BoomerAMG: Number of
paths for aggressive coarsening 1  HYPRE BoomerAMG: Maximum row sums
0.9  HYPRE BoomerAMG: Sweeps down 1  HYPRE BoomerAMG:
Sweeps up   1  HYPRE BoomerAMG: Sweeps on coarse1
HYPRE BoomerAMG: Relax down  symmetric-SOR/Jacobi  HYPRE
BoomerAMG: Relax upsymmetric-SOR/Jacobi  HYPRE BoomerAMG:
Relax on coarse Gaussian-elimination  HYPRE BoomerAMG: Relax
weight  (all)  1.  HYPRE BoomerAMG: Outer relax weight (all)
1.  HYPRE BoomerAMG: Using CF-relaxation  HYPRE BoomerAMG: Not
using more complex smoothers.  HYPRE BoomerAMG: Measure type
local  HYPRE BoomerAMG: Coarsen typeFalgout  HYPRE
BoomerAMG: Interpolation type  classicallinear system matrix followed
by preconditioner matrix:Mat Object: 1 MPI processes  type:
mffd  rows=9, cols=9Matrix-free approximation:
err=1.49012e-08 (relative error in function evaluation)  Using wp
compute h routine  Does not compute normUMat Object:
() 1 MPI processes  type: seqaij  rows=9, cols=9  total:
nonzeros=49, allocated nonzeros=49  total number of mallocs used during
MatSetValues calls =0not using I-node routines*

Fande,