Staggered Scheme with ParFEAP

[ad10]

Dear all,

I want to implement a staggered scheme for the coupled problems. I can implement a staggered scheme using the PART command in serial FEAP. But Not able to use it during parallel one. What changes are to be made to run parallel? Below is my solve.sta code but it does not seem to work.

Any suggestion will be helpful!

Code: [Select]


BATCh
  PETSc ON
END


PARTitions
1 1 0
0 0 1

BATCH
  DT,,1
  prop,,1
END

BATCH
  TOL,,1e-3, 1e-3
  LOOP,,1000
    TIME
    PART,,1
      LOOP,,1000     !
        TANG,,1   !
      NEXT        !
      plot,reac
      reac,file
    PART,,2
      LOOP,,1000     !
        TANG,,1      !
      NEXT           !
    PLOT,WIPE
    PLOT CONT 3 ! Temperature contours
  NEXT
END


[Prof. S. Govindjee]

We have not implemented PARTitions in parFEAP.  However, I believe some groups have done so in the past.  You may want to inquire with Prof. Haim Waisman at Columbia University, I think his group may have an implementation.

[ad10]

Thanks, professor for the suggestion. I will try to connect with them.

[Prof. S. Govindjee]

I should point out that I believe they have done this using the split field feature in petsc.

[JStorm]

I guess that this feature would be very interesting for some users, including me.

We have  realised the staggered scheme at the element level.
Depending on the current partion flag from a user macro, only the corresponding sub-matrix of the element matrix is created, whereas the rest gets the identity matrix and zero residuum.
On top of that, we store and restore the corresponding convergence values for each partition, e.g. initial residuum norm and energy.
This is approach is working well.
However, it is inefficient because the full size system of equation is solved, independently on which partition is active.

[Prof. S. Govindjee]

I wonder how inefficient this really is.  The identity part should solve in one iteration and there is no communication for off-diagonal terms.  I wonder if you can make a test where you construct a matrix and then another one that only has the parts that you need and then compare the solve times.

[JStorm]

I wonder how inefficient this really is.  The identity part should solve in one iteration and there is no communication for off-diagonal terms.  I wonder if you can make a test where you construct a matrix and then another one that only has the parts that you need and then compare the solve times.

This is definitely worth a try.
I am not very optimistic because the identity is placed in a sparse matrix with more activated components.
Thus, we have many zero components.

[Prof. S. Govindjee]

True, but you can set the 'profile' so that there is only one non-zero on the extra rows; thus the storage should be quite minimal and they should now interact at all with the other equations.