[petsc-dev] [petsc-users] Multiple solves with PCMG fail

[Barry Smith]

> On Oct 22, 2014, at 1:25 AM, Jed Brown <jed at jedbrown.org> wrote:
> 
> Barry Smith <bsmith at mcs.anl.gov> writes:
> 
>>  Jed
>> 
>>    There are some additional issues when the GMRES runs for zero iterations in computing eigenvalues for Cheby.
> 
> Hmm, what is right?

   I don’t know. I was never a big fan of Chebyshev smoothing anyways :-)

   I do know that the current situation is bad 

  Barry

>  Suggest that the user set
> -ksp_chebyshev_estimate_eigenvalues_random?  Automatically fall back to
> that?
> 
> Just picking arbitrary values doesn't seem right because the solver
> won't work right on the next solve.
> 
>> static PetscErrorCode KSPChebyshevComputeExtremeEigenvalues_Private(KSP kspest,PetscReal *emin,PetscReal *emax)
>> {
>>  PetscErrorCode ierr;
>>  PetscInt       n,neig;
>>  PetscReal      *re,*im,min,max;
>> 
>>  PetscFunctionBegin;
>>  ierr = KSPGetIterationNumber(kspest,&n);CHKERRQ(ierr);
>>  ierr = PetscMalloc2(n,&re,n,&im);CHKERRQ(ierr);
>>  ierr = KSPComputeEigenvalues(kspest,n,re,im,&neig);CHKERRQ(ierr);
>>  min  = PETSC_MAX_REAL;
>>  max  = PETSC_MIN_REAL;
>>  for (n=0; n<neig; n++) {
>>    min = PetscMin(min,re[n]);
>>    max = PetscMax(max,re[n]);
>>  }
>>  ierr  = PetscFree2(re,im);CHKERRQ(ierr);
>>  *emax = max;
>>  *emin = min;
>> 
>> * thread #1: tid = 0x3c16cf, 0x00000001041b6777 libpetsc.3.5.dylib`KSPSolve_Chebyshev(ksp=0x00007fb85197ac60) + 1495 at cheby.c:378, queue = 'com.apple.main-thread', stop reason = step over
>>    frame #0: 0x00000001041b6777 libpetsc.3.5.dylib`KSPSolve_Chebyshev(ksp=0x00007fb85197ac60) + 1495 at cheby.c:378
>>   375 	    cheb->emin = cheb->tform[0]*min + cheb->tform[1]*max;
>>   376 	    cheb->emax = cheb->tform[2]*min + cheb->tform[3]*max;
>>   377 	
>> -> 378 	    cheb->estimate_current = PETSC_TRUE;
>>   379 	  }
>>   380 	
>>   381 	  ksp->its = 0;
>> (lldb) p cheb->emin
>> (PetscReal) $11 = -1.7976931348623158E+307
>> (lldb) p cheb->emax
>> (PetscReal) $12 = -Inf
>> 
>> This puts giberish into the eigenvalue estimates. 
>> 
>> I am not sure how you want to handle this? 
>> 
>> The zero right hand side zero, initial guess case is a great corner case for testing complicated solvers :-) I’d forgotten about it for many years