[petsc-users] Partition of parallel AIJ sparce matrix

[Barry Smith]

 Yes

On Jul 31, 2014, at 11:04 AM, Qin Lu <lu_qin_2000 at yahoo.com> wrote:

> Barry,
>  
> If I understand you correctly, I should number all unknowns owned by a process contiguously before I construct the matrix, right?
>  
> Thanks,
> Qin  
> 
> From: Barry Smith <bsmith at mcs.anl.gov>
> To: Qin Lu <lu_qin_2000 at yahoo.com> 
> Cc: Matthew Knepley <knepley at gmail.com>; petsc-users <petsc-users at mcs.anl.gov> 
> Sent: Thursday, July 31, 2014 10:46 AM
> Subject: Re: [petsc-users] Partition of parallel AIJ sparce matrix
> 
> 
> On Jul 31, 2014, at 9:35 AM, Qin Lu <lu_qin_2000 at yahoo.com> wrote:
> 
> > Do you mean the user has to reorder both rows and columns (unknowns)? Or the renumbering of columns are done by PETSc?
> 
>   NO, never reorder the matrix. The degrees of freedom should be reordered to be contiquous before any matrix is ever created. Then the matrix and everything else is always computed with regard to this new ordering.
> 
>   Barry
> 
> 
> 
> 
> > Matthew's early response said that only rows need to renumbered by the user so that they are contiguous for each process. In my case, the order of unknowns have been determined before parallel partitioning, is this (just renumbering rows) Ok?
> >  
> > Thanks,
> > Qin    
> > 
> > From: Barry Smith <bsmith at mcs.anl.gov>
> > To: Qin Lu <lu_qin_2000 at yahoo.com> 
> > Cc: Matthew Knepley <knepley at gmail.com>; petsc-users <petsc-users at mcs.anl.gov> 
> > Sent: Wednesday, July 30, 2014 7:37 PM
> > Subject: Re: [petsc-users] Partition of parallel AIJ sparce matrix
> > 
> > 
> > On Jul 30, 2014, at 3:59 PM, Qin Lu <lu_qin_2000 at yahoo.com> wrote:
> > 
> > > Renumbering row will make an unknown not associate with its row (equation),
> > 
> >  It is a symmetric reordering. In matrix terms both the rows and columns are reordered the same way. In terms of spaces both the domain and range spaces are reordered the same way.
> > 
> >  Barry
> > 
> > 
> > > will this affect the solver performance? For example, will the infill pattern change in ILU factorization?
> > >  
> > > Thanks,
> > > Qin 
> > > 
> > > From: Matthew Knepley <knepley at gmail.com>
> > > To: Qin Lu <lu_qin_2000 at yahoo.com> 
> > > Cc: Barry Smith <bsmith at mcs.anl.gov>; petsc-users <petsc-users at mcs.anl.gov> 
> > > Sent: Wednesday, July 30, 2014 3:50 PM
> > > Subject: Re: [petsc-users] Partition of parallel AIJ sparce matrix
> > > 
> > > On Wed, Jul 30, 2014 at 3:30 PM, Qin Lu <lu_qin_2000 at yahoo.com> wrote:
> > > In the context of domain decompostion, if the unknowns are ordered (to reduce the number of infills, for instance) in the way that each subdomain may not own consecutive unknown index, does this mean the partition of the domain will be different from the partition of the matrix?
> > >  
> > > For example, if subdomain 1 (assigned to process 1) owns unknowns 1 and 3 (associated with equation 1 and 3), subdomain 2 (assigned to process 2) owns unknowns 2 and 4 (associated with equation 2 and 4) , how can I make each process own consecutive rows?
> > > 
> > > You renumber the rows once they are partitioned.
> > > 
> > >    Matt
> > > 
> > > 
> > >  
> > > Thanks,
> > > Qin
> > > 
> > > From: Barry Smith <bsmith at mcs.anl.gov>
> > > To: Qin Lu <lu_qin_2000 at yahoo.com> 
> > > Cc: petsc-users <petsc-users at mcs.anl.gov> 
> > > Sent: Wednesday, July 30, 2014 2:49 PM
> > > Subject: Re: [petsc-users] Partition of parallel AIJ sparce matrix
> > > 
> > > 
> > > On Jul 30, 2014, at 11:08 AM, Qin Lu <lu_qin_2000 at yahoo.com> wrote:
> > > 
> > > > Hello,
> > > >  
> > > > Does a process have to own consecutive rows of the matrix? For example, suppose the global AIJ is 4x4, partitioned by 2 processes. Does process 1 have to own rows 1 and 2, process 2 own rows 3 and 4?
> > > 
> > >  Yes
> > > 
> > > > Or process1 may own rows 1 and 3, and process 2 own row 2 and 4?
> > > 
> > >  However, the numbering of degrees of freedom is arbitrary. Just renumber you degrees of freedom so the first set is on process 0, the next on process 1 etc.
> > > 
> > >  Barry
> > > 
> > > 
> > > >  
> > > > Thanks a lot for your help!
> > > >  
> > > > Regards,
> > > > Qin
> > > 
> > > 
> > > 
> > > 
> > > 
> > > 
> > > -- 
> > > What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.
> > > -- Norbert Wiener
> > > 
> > > 
> > 
> > 
> 
>