Thank you for your help. It turned out the problem was that I forgot to assemble the "x" vector before doing the scatter. It seems to be working now. Thanks, Sreeram On Wed, Dec 6, 2023 at 11:12 AM Junchao Zhang <[email protected]> wrote:
Hi, Sreeram, I made an example with your approach. It worked fine as you see the output at the end
#include <petscvec.h> int main(int argc, char **argv) { PetscInt i, j, rstart, rend, n, N, *indices; PetscMPIInt size, rank; IS ix; VecScatter vscat; Vec x, y;
PetscFunctionBeginUser; PetscCall(PetscInitialize(&argc, &argv, (char *)0, NULL)); PetscCallMPI(MPI_Comm_size(PETSC_COMM_WORLD, &size)); PetscCallMPI(MPI_Comm_rank(PETSC_COMM_WORLD, &rank));
PetscCall(VecCreate(PETSC_COMM_WORLD, &x)); PetscCall(VecSetFromOptions(x)); PetscCall(PetscObjectSetName((PetscObject)x, "Vec X")); n = (rank < 4) ? 9 : 0; PetscCall(VecSetSizes(x, n, PETSC_DECIDE));
PetscCall(VecGetOwnershipRange(x, &rstart, &rend)); for (i = rstart; i < rend; i++) PetscCall(VecSetValue(x, i, (PetscScalar)i, INSERT_VALUES)); PetscCall(VecAssemblyBegin(x)); PetscCall(VecAssemblyEnd(x)); PetscCall(VecGetSize(x, &N));
PetscCall(VecCreate(PETSC_COMM_WORLD, &y)); PetscCall(VecSetFromOptions(y)); PetscCall(PetscObjectSetName((PetscObject)y, "Vec Y")); PetscCall(VecSetSizes(y, PETSC_DECIDE, N));
PetscCall(VecGetOwnershipRange(y, &rstart, &rend)); PetscCall(PetscMalloc1(rend - rstart, &indices)); for (i = rstart, j = 0; i < rend; i++, j++) indices[j] = rank + size * j;
PetscCall(ISCreateGeneral(PETSC_COMM_WORLD, rend - rstart, indices, PETSC_OWN_POINTER, &ix)); PetscCall(VecScatterCreate(x, ix, y, NULL, &vscat));
PetscCall(VecScatterBegin(vscat, x, y, INSERT_VALUES, SCATTER_FORWARD)); PetscCall(VecScatterEnd(vscat, x, y, INSERT_VALUES, SCATTER_FORWARD));
PetscCall(ISView(ix, PETSC_VIEWER_STDOUT_WORLD)); PetscCall(VecView(x, PETSC_VIEWER_STDERR_WORLD)); PetscCall(VecView(y, PETSC_VIEWER_STDERR_WORLD));
PetscCall(VecScatterDestroy(&vscat)); PetscCall(ISDestroy(&ix)); PetscCall(VecDestroy(&x)); PetscCall(VecDestroy(&y)); PetscCall(PetscFinalize()); return 0; }
$ mpirun -n 12 ./ex100 IS Object: 12 MPI processes type: general [0] Number of indices in set 3 [0] 0 0 [0] 1 12 [0] 2 24 [1] Number of indices in set 3 [1] 0 1 [1] 1 13 [1] 2 25 [2] Number of indices in set 3 [2] 0 2 [2] 1 14 [2] 2 26 [3] Number of indices in set 3 [3] 0 3 [3] 1 15 [3] 2 27 [4] Number of indices in set 3 [4] 0 4 [4] 1 16 [4] 2 28 [5] Number of indices in set 3 [5] 0 5 [5] 1 17 [5] 2 29 [6] Number of indices in set 3 [6] 0 6 [6] 1 18 [6] 2 30 [7] Number of indices in set 3 [7] 0 7 [7] 1 19 [7] 2 31 [8] Number of indices in set 3 [8] 0 8 [8] 1 20 [8] 2 32 [9] Number of indices in set 3 [9] 0 9 [9] 1 21 [9] 2 33 [10] Number of indices in set 3 [10] 0 10 [10] 1 22 [10] 2 34 [11] Number of indices in set 3 [11] 0 11 [11] 1 23 [11] 2 35 Vec Object: Vec X 12 MPI processes type: mpi Process [0] 0. 1. 2. 3. 4. 5. 6. 7. 8. Process [1] 9. 10. 11. 12. 13. 14. 15. 16. 17. Process [2] 18. 19. 20. 21. 22. 23. 24. 25. 26. Process [3] 27. 28. 29. 30. 31. 32. 33. 34. 35. Process [4] Process [5] Process [6] Process [7] Process [8] Process [9] Process [10] Process [11] Vec Object: Vec Y 12 MPI processes type: mpi Process [0] 0. 12. 24. Process [1] 1. 13. 25. Process [2] 2. 14. 26. Process [3] 3. 15. 27. Process [4] 4. 16. 28. Process [5] 5. 17. 29. Process [6] 6. 18. 30. Process [7] 7. 19. 31. Process [8] 8. 20. 32. Process [9] 9. 21. 33. Process [10] 10. 22. 34. Process [11] 11. 23. 35.
--Junchao Zhang
On Tue, Dec 5, 2023 at 10:09 PM Sreeram R Venkat <[email protected]> wrote:
Yes, I have an example code at github.com/s769/petsc-test. Only thing is, when I described the example before, I simplified the actual use case in the code to make things simpler. Here are the extra details relevant to this code:
- We assume a 2D processor grid, given by the command-line args -proc_rows and -proc_cols - The total length of the vector is n_m*n_t given by command-line args -nm and -nt. n_m corresponds to a space index and n_t a time index. - In the "Start" phase, the vector is divided into n_m blocks each of size n_t (indexed space->time). The blocks are partitioned over the first row of processors. For example if -nm = 4 and -proc_cols = 4, each processor in the first row will get one block of size n_t. Each processor in the first row gets n_m_local blocks of size n_t, where the sum of all n_m_locals for the first row of processors is n_m. - In the "Finish" phase, the vector is divided into n_t blocks each of size n_m (indexed time->space; this is the reason for the permutation of indices). The blocks are partitioned over all processors. Each processor will get n_t_local blocks of size n_m, where the sum of all n_t_locals for all processors is n_t.
I think the basic idea is similar to the previous example, but these details make things a bit more complicated. Please let me know if anything is unclear, and I can try to explain more.
Thanks for your help, Sreeram
On Tue, Dec 5, 2023 at 9:30 PM Junchao Zhang <[email protected]> wrote:
I think your approach is correct. Do you have an example code?
--Junchao Zhang
On Tue, Dec 5, 2023 at 5:15 PM Sreeram R Venkat <[email protected]> wrote:
Hi, I have a follow up question on this.
Now, I'm trying to do a scatter and permutation of the vector. Under the same setup as the original example, here are the new Start and Finish states I want to achieve: Start Finish Proc | Entries Proc | Entries 0 | 0,...,8 0 | 0, 12, 24 1 | 9,...,17 1 | 1, 13, 25 2 | 18,...,26 2 | 2, 14, 26 3 | 27,...,35 3 | 3, 15, 27 4 | None 4 | 4, 16, 28 5 | None 5 | 5, 17, 29 6 | None 6 | 6, 18, 30 7 | None 7 | 7, 19, 31 8 | None 8 | 8, 20, 32 9 | None 9 | 9, 21, 33 10 | None 10 | 10, 22, 34 11 | None 11 | 11, 23, 35
So far, I've tried to use ISCreateGeneral(), with each process giving an idx array corresponding to the indices it wants (i.e. idx on P0 looks like [0,12,24] P1 -> [1,13, 25], and so on). Then I use this to create the VecScatter with VecScatterCreate(x, is, y, NULL, &scatter).
However, when I try to do the scatter, I get some illegal memory access errors.
Is there something wrong with how I define the index sets?
Thanks, Sreeram
On Thu, Oct 5, 2023 at 12:57 PM Sreeram R Venkat <[email protected]> wrote:
Thank you. This works for me.
Sreeram
On Wed, Oct 4, 2023 at 6:41 PM Junchao Zhang <[email protected]> wrote:
Hi, Sreeram, You can try this code. Since x, y are both MPI vectors, we just need to say we want to scatter x[0:N] to y[0:N]. The 12 index sets with your example on the 12 processes would be [0..8], [9..17], [18..26], [27..35], [], ..., []. Actually, you can do it arbitrarily, say, with 12 index sets [0..17], [18..35], .., []. PETSc will figure out how to do the communication.
PetscInt rstart, rend, N; IS ix; VecScatter vscat; Vec y; MPI_Comm comm; VecType type;
PetscObjectGetComm((PetscObject)x, &comm); VecGetType(x, &type); VecGetSize(x, &N); VecGetOwnershipRange(x, &rstart, &rend);
VecCreate(comm, &y); VecSetSizes(y, PETSC_DECIDE, N); VecSetType(y, type);
ISCreateStride(PetscObjectComm((PetscObject)x), rend - rstart, rstart, 1, &ix); VecScatterCreate(x, ix, y, ix, &vscat);
--Junchao Zhang
On Wed, Oct 4, 2023 at 6:03 PM Sreeram R Venkat <[email protected]> wrote:
Suppose I am running on 12 processors, and I have a vector "v" of size 36 partitioned over the first 4. v still uses the PETSC_COMM_WORLD, so it has a layout of (9, 9, 9, 9, 0, 0, ..., 0). Now, I would like to repartition it over all 12 processors, so that the layout becomes (3, 3, 3, ..., 3). I've been trying to use VecScatter to do this, but I'm not sure what IndexSets to use for the sender and receiver.
The result I am trying to achieve is this:
Assume the vector is v = <0, 1, 2, ..., 35>
Start Finish Proc | Entries Proc | Entries 0 | 0,...,8 0 | 0, 1, 2 1 | 9,...,17 1 | 3, 4, 5 2 | 18,...,26 2 | 6, 7, 8 3 | 27,...,35 3 | 9, 10, 11 4 | None 4 | 12, 13, 14 5 | None 5 | 15, 16, 17 6 | None 6 | 18, 19, 20 7 | None 7 | 21, 22, 23 8 | None 8 | 24, 25, 26 9 | None 9 | 27, 28, 29 10 | None 10 | 30, 31, 32 11 | None 11 | 33, 34, 35
Appreciate any help you can provide on this.
Thanks, Sreeram