[petsc-dev] pbjacobi error, in parallel, on kokkos/cuda

Mark Adams mfadams at lbl.gov
Tue Feb 18 10:07:44 CST 2025 

Also, this uses the branch: adams/mat-rap-blocksize
that has fixes to get the block sizes moved up in P'AP.

On Tue, Feb 18, 2025 at 9:29 AM Mark Adams <mfadams at lbl.gov> wrote:

> I've got a bug in pbjacobi that only shows up on* the Galerkin coarse
> grid *(I have not been able to reproduce it a fine grid at least), and in
> *parallel*, and on *GPUs*/kokkos.
>
> I have modified ex55 to take P from GAMG and give it (one) to PCMG with
> Galerkin coarse grids,and solve (code and command lines appended).
> I see this with ex56 (bs=3 & 6), ex55 (bs=2), but ex54 (bs=1) is fine
> (does pbjacobi switch to jacobi?)
>
> With 4 processors I get these valgrind errors only on these bad solves (no
> false positives). Note that *only one process has errors*, and note some
> solver output before and after:
>
> I'm going to keeps digging but ideas are welcome,
> Thanks,
> Mark
>
> [0] <pc:gamg> PCSetUp_GAMG(): (null): 1) N=4, n data cols=2, nnz/row
> (ave)=4, 1 active pes
> [0] <pc:gamg> PCSetUp_GAMG(): (null): 2 levels, operator complexity = 1.04
> [0] <pc:gamg> PCSetUp_GAMG(): (null): PCSetUp_GAMG: call
> KSPChebyshevSetEigenvalues on level 0 (N=32) with emax = 2.26125 emin =
> 0.0198344
> [0] <pc:gamg> PCSetUp_MG(): Using outer operators to define finest grid
> operator
>   because PCMGGetSmoother(pc,nlevels-1,&ksp);KSPSetOperators(ksp,...); was
> not called.
> [0] <pc:mg> PCSetUp_MG(): Using outer operators to define finest grid
> operator
>   because PCMGGetSmoother(pc,nlevels-1,&ksp);KSPSetOperators(ksp,...); was
> not called.
> ==978424== Invalid read of size 16
> ==978424==    at 0x42B9AFA2: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42AF0F12: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42C3EE7B: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42EEC474: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42B4B495: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42EC748F: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42B44781: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42CF766D: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42445504: ??? (in
> /opt/nvidia/hpc_sdk/Linux_x86_64/23.9/cuda/12.2/targets/x86_64-linux/lib/libcudart.so.12.2.53)
> ==978424==    by 0x42417A04: ??? (in
> /opt/nvidia/hpc_sdk/Linux_x86_64/23.9/cuda/12.2/targets/x86_64-linux/lib/libcudart.so.12.2.53)
> ==978424==    by 0x42468730: cudaMemcpy (in
> /opt/nvidia/hpc_sdk/Linux_x86_64/23.9/cuda/12.2/targets/x86_64-linux/lib/libcudart.so.12.2.53)
> ==978424==    by 0x4896979: cuda_memcpy_wrapper<>
> (Kokkos_Cuda_Instance.hpp:365)
> ==978424==    by 0x4896979: Kokkos::Impl::DeepCopyCuda(void*, void const*,
> unsigned long) (Kokkos_CudaSpace.cpp:62)
> ==978424==  Address 0xcb722dfc is 1,644 bytes inside a block of size 1,652
> alloc'd
> ==978424==    at 0x4E0A926: memalign (in
> /usr/lib/valgrind/vgpreload_memcheck-amd64-linux.so)
> ==978424==    by 0x4E0AA69: posix_memalign (in
> /usr/lib/valgrind/vgpreload_memcheck-amd64-linux.so)
> ==978424==    by 0x57E3BD9: PetscMallocAlign (mal.c:52)
> ==978424==    by 0x57E892F: PetscTrMallocDefault (mtr.c:175)
> ==978424==    by 0x57E5A77: PetscMallocA (mal.c:421)
> ==978424==    by 0x63B02F5: MatInvertBlockDiagonal_SeqAIJ (aij.c:3333)
> ==978424==    by 0x69B0599: MatInvertBlockDiagonal (matrix.c:10908)
> ==978424==    by 0x629D585: MatInvertBlockDiagonal_MPIAIJ (mpiaij.c:2588)
> ==978424==    by 0x69B0599: MatInvertBlockDiagonal (matrix.c:10908)
> ==978424==    by 0x7C7F726: PCSetUp_PBJacobi_Host (pbjacobi.c:256)
> ==978424==    by 0x7D737D8: PCSetUp_PBJacobi_Kokkos
> (pbjacobi_kok.kokkos.cxx:90)
> ==978424==    by 0x7C803F8: PCSetUp_PBJacobi (pbjacobi.c:296)
> ==978424==
> ==978424== Invalid read of size 16
> ==978424==    at 0x42B9AFB4: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42AF0F12: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42C3EE7B: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42EEC474: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42B4B495: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42EC748F: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42B44781: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42CF766D: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42445504: ??? (in
> /opt/nvidia/hpc_sdk/Linux_x86_64/23.9/cuda/12.2/targets/x86_64-linux/lib/libcudart.so.12.2.53)
> ==978424==    by 0x42417A04: ??? (in
> /opt/nvidia/hpc_sdk/Linux_x86_64/23.9/cuda/12.2/targets/x86_64-linux/lib/libcudart.so.12.2.53)
> ==978424==    by 0x42468730: cudaMemcpy (in
> /opt/nvidia/hpc_sdk/Linux_x86_64/23.9/cuda/12.2/targets/x86_64-linux/lib/libcudart.so.12.2.53)
> ==978424==    by 0x4896979: cuda_memcpy_wrapper<>
> (Kokkos_Cuda_Instance.hpp:365)
> ==978424==    by 0x4896979: Kokkos::Impl::DeepCopyCuda(void*, void const*,
> unsigned long) (Kokkos_CudaSpace.cpp:62)
> ==978424==  Address 0xcb722e0c is 8 bytes after a block of size 1,652
> alloc'd
> ==978424==    at 0x4E0A926: memalign (in
> /usr/lib/valgrind/vgpreload_memcheck-amd64-linux.so)
> ==978424==    by 0x4E0AA69: posix_memalign (in
> /usr/lib/valgrind/vgpreload_memcheck-amd64-linux.so)
> ==978424==    by 0x57E3BD9: PetscMallocAlign (mal.c:52)
> ==978424==    by 0x57E892F: PetscTrMallocDefault (mtr.c:175)
> ==978424==    by 0x57E5A77: PetscMallocA (mal.c:421)
> ==978424==    by 0x63B02F5: MatInvertBlockDiagonal_SeqAIJ (aij.c:3333)
> ==978424==    by 0x69B0599: MatInvertBlockDiagonal (matrix.c:10908)
> ==978424==    by 0x629D585: MatInvertBlockDiagonal_MPIAIJ (mpiaij.c:2588)
> ==978424==    by 0x69B0599: MatInvertBlockDiagonal (matrix.c:10908)
> ==978424==    by 0x7C7F726: PCSetUp_PBJacobi_Host (pbjacobi.c:256)
> ==978424==    by 0x7D737D8: PCSetUp_PBJacobi_Kokkos
> (pbjacobi_kok.kokkos.cxx:90)
> ==978424==    by 0x7C803F8: PCSetUp_PBJacobi (pbjacobi.c:296)
> ==978424==
> ==978424== Invalid read of size 4
> ==978424==    at 0x42B9B103: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42AF0F12: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42C3EE7B: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42EEC474: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42B4B495: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42EC748F: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42B44781: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42CF766D: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42445504: ??? (in
> /opt/nvidia/hpc_sdk/Linux_x86_64/23.9/cuda/12.2/targets/x86_64-linux/lib/libcudart.so.12.2.53)
> ==978424==    by 0x42417A04: ??? (in
> /opt/nvidia/hpc_sdk/Linux_x86_64/23.9/cuda/12.2/targets/x86_64-linux/lib/libcudart.so.12.2.53)
> ==978424==    by 0x42468730: cudaMemcpy (in
> /opt/nvidia/hpc_sdk/Linux_x86_64/23.9/cuda/12.2/targets/x86_64-linux/lib/libcudart.so.12.2.53)
> ==978424==    by 0x4896979: cuda_memcpy_wrapper<>
> (Kokkos_Cuda_Instance.hpp:365)
> ==978424==    by 0x4896979: Kokkos::Impl::DeepCopyCuda(void*, void const*,
> unsigned long) (Kokkos_CudaSpace.cpp:62)
> ==978424==  Address 0xcb722e1c is 12 bytes after a block of size 1,664 in
> arena "client"
> ==978424==
> ==978424== Invalid read of size 4
> ==978424==    at 0x42B9B107: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42AF0F12: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42C3EE7B: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42EEC474: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42B4B495: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42EC748F: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42B44781: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42CF766D: ??? (in /usr/lib64/libcuda.so.550.127.08)
> ==978424==    by 0x42445504: ??? (in
> /opt/nvidia/hpc_sdk/Linux_x86_64/23.9/cuda/12.2/targets/x86_64-linux/lib/libcudart.so.12.2.53)
> ==978424==    by 0x42417A04: ??? (in
> /opt/nvidia/hpc_sdk/Linux_x86_64/23.9/cuda/12.2/targets/x86_64-linux/lib/libcudart.so.12.2.53)
> ==978424==    by 0x42468730: cudaMemcpy (in
> /opt/nvidia/hpc_sdk/Linux_x86_64/23.9/cuda/12.2/targets/x86_64-linux/lib/libcudart.so.12.2.53)
> ==978424==    by 0x4896979: cuda_memcpy_wrapper<>
> (Kokkos_Cuda_Instance.hpp:365)
> ==978424==    by 0x4896979: Kokkos::Impl::DeepCopyCuda(void*, void const*,
> unsigned long) (Kokkos_CudaSpace.cpp:62)
> ==978424==  Address 0xcb722e1c is 12 bytes after a block of size 1,664 in
> arena "client"
> ==978424==
>       Residual norms for rap_mg_coarse_ solve.
>       0 KSP Residual norm 1.118121970641e+01
>       1 KSP Residual norm 3.575439993035e-01
> [0]PETSC ERROR: --------------------- Error Message
> --------------------------------------------------------------
> [0]PETSC ERROR: Diverged due to indefinite preconditioner, beta
> -0.00553023, betaold 5.27744
>
>
> ksp/ex55:
>
>
>     PC pc;
>   PetscCall(KSPGetPC(ksp, &pc));
>   PC_MG   *mg      = (PC_MG *)pc->data;
>   PC_MG_Levels **mglevels = mg->levels;
>   Mat P = mglevels[mg->nlevels-1]->interpolate;
>   PetscCall(MatViewFromOptions(mglevels[mg->nlevels-1]->A, NULL,
> "-rap_mat_view"));
>   PetscCall(MatViewFromOptions(Amat, NULL, "-rap_mat_view"));
>   KSP ksp2;
>   PetscCall(KSPCreate(PETSC_COMM_WORLD, &ksp2));
>   PetscCall(KSPSetOptionsPrefix(ksp2, "rap_"));
>   PetscCall(KSPSetFromOptions(ksp2));
>   PetscCall(KSPGetPC(ksp2, &pc));
>   PetscCall(KSPSetOperators(ksp2, Amat, Amat));
>   PetscCall(PCMGSetGalerkin(pc, PC_MG_GALERKIN_PMAT));
>   PetscCall(PCMGSetInterpolation(pc, 1, P));
>   PetscCall(VecSet(bb, 1.0));
>   PetscCall(PetscLogStagePush(stage[1]));
>   PetscCall(KSPSolve(ksp2, bb, xx));
>   PetscCall(PetscLogStagePop());
>   PetscCall(PetscFinalize());
>   exit(12);
>
>
>   PetscCall(PetscLogStagePush(stage[1])); // original ex55 code
>   PetscCall(KSPSolve(ksp, bb, xx));
>
>
> *$ srun -n 4 valgrind --tool=memcheck --leak-check=no ./ex55 -ne 3
> -pc_type gamg -rap_pc_type mg -rap_ksp_monitor -rap_mg_levels_pc_type
> jacobi -rap_mg_coarse_pc_type pbjacobi -rap_mg_coarse_ksp_monitor
> -options_left -rap_pc_mg_levels 2 -rap_mg_coarse_ksp_type cg -mat_type
> aijkokkos -fp_trap -ksp_monitor -rap_ksp_viewxx -info :pc,dm
> -rap_mg_coarse_ksp_error_if_not_converged *
>
>
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20250218/28fb75c2/attachment-0001.html>

More information about the petsc-dev mailing list