assembly in parallel implementation

[Shu]

Dear All,

I am doing some palatalization work using FEAP. By read the source file, I find that FEAP using PetSc and create a block aij martrix to assemble stiffness matrix. Can anyone introduce this method comparing with normal aij matrix? What is the advantage to do so?

Thanks

Shu

[FEAP_Admin]

The choice of BAIJ versus AIJ is mainly dictated by the solver you are using.  For example if you are using prometheus then you have to use BAIJ.  If you are using SuperLU then you have to use AIJ.

In terms of speed, there is a bit of speed up with BAIJ but in general this is not going to be substantial.

[Shu]

Thanks for your reply. Also I have a question, when FEAP is trying to assemble the matrix, for example in a blocked way, it has:
          call MatCreateMPIBAIJ(PETSC_COMM_WORLD,nsbk,numpeq,numpeq,
     &                          PETSC_DETERMINE, PETSC_DETERMINE,
     &                          PETSC_NULL_INTEGER, mr(np(246)),
     &                          PETSC_NULL_INTEGER, mr(np(247)),
     &                          Kmatem, ierr)

Where does this np(246) and np(247) comes from? I was trying to search the source code but cannot have a clue. Can you briefly introduce how they are calculated?

Thank.
Shu

[FEAP_Admin]

When you partition your problem [graph,node,XX] this information is generated and written to the partitioned input files
by [outd,<bloc>k]; both commands are in parfeap/pmacr7.F.

Then when you start-up your parallel computation, the information for on-diagonal and off-diagonal storage is read in from the parallel input files and stored in mr(np(246)) and mr(np(247)), respectively.  The call sequence is program/pcontr.f calling parfeap/pdomain.F.  [Note there is a dummy pdomain.f in program for serial runs.]

[Shu]

Thank you very much for your help!