[petsc-users] Running CG with HYPRE AMG preconditioner in AMD GPUs

Vanella, Marcos (Fed) marcos.vanella at nist.gov
Tue Mar 19 16:06:51 CDT 2024


Hi Mark, thanks. I'll try your suggestions. So, I would keep -mat_type mpiaijkokkos but -vec_type hip as runtime options?
Thanks,
Marcos
________________________________
From: Mark Adams <mfadams at lbl.gov>
Sent: Tuesday, March 19, 2024 4:57 PM
To: Vanella, Marcos (Fed) <marcos.vanella at nist.gov>
Cc: PETSc users list <petsc-users at mcs.anl.gov>
Subject: Re: [petsc-users] Running CG with HYPRE AMG preconditioner in AMD GPUs

[keep on list]

I have little experience with running hypre on GPUs but others might have more.

1M dogs/node is not a lot and NVIDIA has larger L1 cache and more mature compilers, etc. so it is not surprising that NVIDIA is faster.
I suspect the gap would narrow with a larger problem.

Also, why are you using Kokkos? It should not make a difference but you could check easily. Just use -vec_type hip with your current code.

You could also test with GAMG, -pc_type gamg

Mark


On Tue, Mar 19, 2024 at 4:12 PM Vanella, Marcos (Fed) <marcos.vanella at nist.gov<mailto:marcos.vanella at nist.gov>> wrote:
Hi Mark, I run a canonical test we have to time our code. It is a propane fire on a burner within a box with around 1 million cells.
I split the problem in 4 GPUS, single node, both in Polaris and Frontier. I compiled PETSc with gnu and HYPRE being downloaded and the following configure options:


  *
Polaris:
$./configure COPTFLAGS="-O3" CXXOPTFLAGS="-O3" FOPTFLAGS="-O3" FCOPTFLAGS="-O3" CUDAOPTFLAGS="-O3" --with-debugging=0 --download-suitesparse --download-hypre --with-cuda --with-cc=cc --with-cxx=CC --with-fc=ftn --with-cudac=nvcc --with-cuda-arch=80 --download-cmake


  *
Frontier:
$./configure COPTFLAGS="-O3" CXXOPTFLAGS="-O3" FOPTFLAGS="-O3" FCOPTFLAGS="-O3" HIPOPTFLAGS="-O3" --with-debugging=0 --with-cc=cc --with-cxx=CC --with-fc=ftn --with-hip --with-hipc=hipcc --LIBS="-L${MPICH_DIR}/lib -lmpi ${PE_MPICH_GTL_DIR_amd_gfx90a} ${PE_MPICH_GTL_LIBS_amd_gfx90a}" --download-kokkos --download-kokkos-kernels --download-suitesparse --download-hypre --download-cmake

Our code was compiled also with gnu compilers and -O3 flag. I used latest (from this week) PETSc repo update. These are the timings for the test case:


  *   8 meshes + 1Million cells case, 8 MPI processes, 4 GPUS, 2 MPI Procs per GPU, 1 sec run time (~580 time steps, ~1160 Poisson solves):

System              Poisson Solver          GPU Implementation          Poisson Wall time (sec)         Total Wall time (sec)
Polaris             CG + HYPRE PC           CUDA                        80                              287
Frontier            CG + HYPRE PC           Kokkos + HIP                158                             401

It is interesting to see that the Poisson solves take twice the time in Frontier than in Polaris.
Do you have experience on running HYPRE AMG on these machines? Is this difference between the CUDA implementation and Kokkos-kernels to be expected?

I can run the case in both computers with the log flags you suggest. Might give more information on where the differences are.

Thank you for your time,
Marcos


________________________________
From: Mark Adams <mfadams at lbl.gov<mailto:mfadams at lbl.gov>>
Sent: Tuesday, March 5, 2024 2:41 PM
To: Vanella, Marcos (Fed) <marcos.vanella at nist.gov<mailto:marcos.vanella at nist.gov>>
Cc: petsc-users at mcs.anl.gov<mailto:petsc-users at mcs.anl.gov> <petsc-users at mcs.anl.gov<mailto:petsc-users at mcs.anl.gov>>
Subject: Re: [petsc-users] Running CG with HYPRE AMG preconditioner in AMD GPUs

You can run with -log_view_gpu_time to get rid of the nans and get more data.

You can run with -ksp_view to get more info on the solver and send that output.

-options_left is also good to use so we can see what parameters you used.

The last 100 in this row:

KSPSolve            1197 0.0 2.0291e+02 0.0 2.55e+11 0.0 3.9e+04 8.0e+04 3.1e+04 12 100 100 100 49  12 100 100 100 98  2503    -nan      0 1.80e-05    0 0.00e+00  100

tells us that all the flops were logged on GPUs.

You do need at least 100K equations per GPU to see speedup, so don't worry about small problems.

Mark




On Tue, Mar 5, 2024 at 12:52 PM Vanella, Marcos (Fed) via petsc-users <petsc-users at mcs.anl.gov<mailto:petsc-users at mcs.anl.gov>> wrote:
Hi all, I compiled the latest PETSc source in Frontier using gcc+kokkos and hip options: ./configure COPTFLAGS="-O3" CXXOPTFLAGS="-O3" FOPTFLAGS="-O3" FCOPTFLAGS="-O3" HIPOPTFLAGS="-O3" --with-debugging=0
ZjQcmQRYFpfptBannerStart
This Message Is From an External Sender
This message came from outside your organization.

ZjQcmQRYFpfptBannerEnd
Hi all, I compiled the latest PETSc source in Frontier using gcc+kokkos and hip options:

./configure COPTFLAGS="-O3" CXXOPTFLAGS="-O3" FOPTFLAGS="-O3" FCOPTFLAGS="-O3" HIPOPTFLAGS="-O3" --with-debugging=0 --with-cc=cc --with-cxx=CC --with-fc=ftn --with-hip --with-hipc=hipcc --LIBS="-L${MPICH_DIR}/lib -lmpi ${PE_MPICH_GTL_DIR_amd_gfx90a} ${PE_MPICH_GTL_LIBS_amd_gfx90a}" --download-kokkos --download-kokkos-kernels --download-suitesparse --download-hypre --download-cmake

and have started testing our code solving a Poisson linear system with CG + HYPRE preconditioner. Timings look rather high compared to compilations done on other machines that have NVIDIA cards. They are also not changing when using more than one GPU for the simple test I doing.
Does anyone happen to know if HYPRE has an hip GPU implementation for Boomer AMG and is it compiled when configuring PETSc?

Thanks!

Marcos


PS: This is what I see on the log file (-log_view) when running the case with 2 GPUs in the node:


------------------------------------------------------------------ PETSc Performance Summary: ------------------------------------------------------------------

/ccs/home/vanellam/Firemodels_fork/fds/Build/mpich_gnu_frontier/fds_mpich_gnu_frontier on a arch-linux-frontier-opt-gcc named frontier04119 with 4 processors, by vanellam Tue Mar  5 12:42:29 2024
Using Petsc Development GIT revision: v3.20.5-713-gabdf6bc0fcf  GIT Date: 2024-03-05 01:04:54 +0000

                         Max       Max/Min     Avg       Total
Time (sec):           8.368e+02     1.000   8.368e+02
Objects:              0.000e+00     0.000   0.000e+00
Flops:                2.546e+11     0.000   1.270e+11  5.079e+11
Flops/sec:            3.043e+08     0.000   1.518e+08  6.070e+08
MPI Msg Count:        1.950e+04     0.000   9.748e+03  3.899e+04
MPI Msg Len (bytes):  1.560e+09     0.000   7.999e+04  3.119e+09
MPI Reductions:       6.331e+04   2877.545

Flop counting convention: 1 flop = 1 real number operation of type (multiply/divide/add/subtract)
                            e.g., VecAXPY() for real vectors of length N --> 2N flops
                            and VecAXPY() for complex vectors of length N --> 8N flops

Summary of Stages:   ----- Time ------  ----- Flop ------  --- Messages ---  -- Message Lengths --  -- Reductions --
                        Avg     %Total     Avg     %Total    Count   %Total     Avg         %Total    Count   %Total
 0:      Main Stage: 8.3676e+02 100.0%  5.0792e+11 100.0%  3.899e+04 100.0%  7.999e+04      100.0%  3.164e+04  50.0%

------------------------------------------------------------------------------------------------------------------------
See the 'Profiling' chapter of the users' manual for details on interpreting output.
Phase summary info:
   Count: number of times phase was executed
   Time and Flop: Max - maximum over all processors
                  Ratio - ratio of maximum to minimum over all processors
   Mess: number of messages sent
   AvgLen: average message length (bytes)
   Reduct: number of global reductions
   Global: entire computation
   Stage: stages of a computation. Set stages with PetscLogStagePush() and PetscLogStagePop().
      %T - percent time in this phase         %F - percent flop in this phase
      %M - percent messages in this phase     %L - percent message lengths in this phase
      %R - percent reductions in this phase
   Total Mflop/s: 10e-6 * (sum of flop over all processors)/(max time over all processors)
   GPU Mflop/s: 10e-6 * (sum of flop on GPU over all processors)/(max GPU time over all processors)
   CpuToGpu Count: total number of CPU to GPU copies per processor
   CpuToGpu Size (Mbytes): 10e-6 * (total size of CPU to GPU copies per processor)
   GpuToCpu Count: total number of GPU to CPU copies per processor
   GpuToCpu Size (Mbytes): 10e-6 * (total size of GPU to CPU copies per processor)
   GPU %F: percent flops on GPU in this event
------------------------------------------------------------------------------------------------------------------------
Event                Count      Time (sec)     Flop                              --- Global ---  --- Stage ----  Total   GPU    - CpuToGpu -   - GpuToCpu - GPU
                   Max Ratio  Max     Ratio   Max  Ratio  Mess   AvgLen  Reduct  %T %F %M %L %R  %T %F %M %L %R Mflop/s Mflop/s Count   Size   Count   Size  %F
---------------------------------------------------------------------------------------------------------------------------------------------------------------

--- Event Stage 0: Main Stage

BuildTwoSided       1201 0.0   nan nan 0.00e+00 0.0 2.0e+00 4.0e+00 6.0e+02  0  0  0  0  1   0  0  0  0  2  -nan    -nan      0 0.00e+00    0 0.00e+00  0
BuildTwoSidedF      1200 0.0   nan nan 0.00e+00 0.0 0.0e+00 0.0e+00 6.0e+02  0  0  0  0  1   0  0  0  0  2  -nan    -nan      0 0.00e+00    0 0.00e+00  0
MatMult            19494 0.0   nan nan 1.35e+11 0.0 3.9e+04 8.0e+04 0.0e+00  7 53 100 100  0   7 53 100 100  0  -nan    -nan      0 1.80e-05    0 0.00e+00  100
MatConvert             3 0.0   nan nan 0.00e+00 0.0 0.0e+00 0.0e+00 1.5e+00  0  0  0  0  0   0  0  0  0  0  -nan    -nan      0 0.00e+00    0 0.00e+00  0
MatAssemblyBegin       2 0.0   nan nan 0.00e+00 0.0 0.0e+00 0.0e+00 1.0e+00  0  0  0  0  0   0  0  0  0  0  -nan    -nan      0 0.00e+00    0 0.00e+00  0
MatAssemblyEnd         2 0.0   nan nan 0.00e+00 0.0 4.0e+00 2.0e+04 3.5e+00  0  0  0  0  0   0  0  0  0  0  -nan    -nan      0 0.00e+00    0 0.00e+00  0
VecTDot            41382 0.0   nan nan 4.14e+10 0.0 0.0e+00 0.0e+00 2.1e+04  0 16  0  0 33   0 16  0  0 65  -nan    -nan      0 0.00e+00    0 0.00e+00  100
VecNorm            20691 0.0   nan nan 2.07e+10 0.0 0.0e+00 0.0e+00 1.0e+04  0  8  0  0 16   0  8  0  0 33  -nan    -nan      0 0.00e+00    0 0.00e+00  100
VecCopy             2394 0.0   nan nan 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00  0  0  0  0  0   0  0  0  0  0  -nan    -nan      0 0.00e+00    0 0.00e+00  0
VecSet             21888 0.0   nan nan 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00  0  0  0  0  0   0  0  0  0  0  -nan    -nan      0 0.00e+00    0 0.00e+00  0
VecAXPY            38988 0.0   nan nan 3.90e+10 0.0 0.0e+00 0.0e+00 0.0e+00  0 15  0  0  0   0 15  0  0  0  -nan    -nan      0 0.00e+00    0 0.00e+00  100
VecAYPX            18297 0.0   nan nan 1.83e+10 0.0 0.0e+00 0.0e+00 0.0e+00  0  7  0  0  0   0  7  0  0  0  -nan    -nan      0 0.00e+00    0 0.00e+00  100
VecAssemblyBegin    1197 0.0   nan nan 0.00e+00 0.0 0.0e+00 0.0e+00 6.0e+02  0  0  0  0  1   0  0  0  0  2  -nan    -nan      0 0.00e+00    0 0.00e+00  0
VecAssemblyEnd      1197 0.0   nan nan 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00  0  0  0  0  0   0  0  0  0  0  -nan    -nan      0 0.00e+00    0 0.00e+00  0
VecScatterBegin    19494 0.0   nan nan 0.00e+00 0.0 3.9e+04 8.0e+04 0.0e+00  0  0 100 100  0   0  0 100 100  0  -nan    -nan      0 1.80e-05    0 0.00e+00  0
VecScatterEnd      19494 0.0   nan nan 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00  0  0  0  0  0   0  0  0  0  0  -nan    -nan      0 0.00e+00    0 0.00e+00  0
SFSetGraph             1 0.0   nan nan 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00  0  0  0  0  0   0  0  0  0  0  -nan    -nan      0 0.00e+00    0 0.00e+00  0
SFSetUp                1 0.0   nan nan 0.00e+00 0.0 4.0e+00 2.0e+04 5.0e-01  0  0  0  0  0   0  0  0  0  0  -nan    -nan      0 0.00e+00    0 0.00e+00  0
SFPack             19494 0.0   nan nan 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00  0  0  0  0  0   0  0  0  0  0  -nan    -nan      0 1.80e-05    0 0.00e+00  0
SFUnpack           19494 0.0   nan nan 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00  0  0  0  0  0   0  0  0  0  0  -nan    -nan      0 0.00e+00    0 0.00e+00  0
KSPSetUp               1 0.0   nan nan 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00  0  0  0  0  0   0  0  0  0  0  -nan    -nan      0 0.00e+00    0 0.00e+00  0
KSPSolve            1197 0.0 2.0291e+02 0.0 2.55e+11 0.0 3.9e+04 8.0e+04 3.1e+04 12 100 100 100 49  12 100 100 100 98  2503    -nan      0 1.80e-05    0 0.00e+00  100
PCSetUp                1 0.0   nan nan 0.00e+00 0.0 0.0e+00 0.0e+00 1.5e+00  0  0  0  0  0   0  0  0  0  0  -nan    -nan      0 0.00e+00    0 0.00e+00  0
PCApply            20691 0.0   nan nan 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00  5  0  0  0  0   5  0  0  0  0  -nan    -nan      0 0.00e+00    0 0.00e+00  0
---------------------------------------------------------------------------------------------------------------------------------------------------------------

Object Type          Creations   Destructions. Reports information only for process 0.

--- Event Stage 0: Main Stage

              Matrix     7              3
              Vector     7              1
           Index Set     2              2
   Star Forest Graph     1              0
       Krylov Solver     1              0
      Preconditioner     1              0
========================================================================================================================
Average time to get PetscTime(): 3.01e-08
Average time for MPI_Barrier(): 3.8054e-06
Average time for zero size MPI_Send(): 7.101e-06
#PETSc Option Table entries:
-log_view # (source: command line)
-mat_type mpiaijkokkos # (source: command line)
-vec_type kokkos # (source: command line)
#End of PETSc Option Table entries
Compiled without FORTRAN kernels
Compiled with full precision matrices (default)
sizeof(short) 2 sizeof(int) 4 sizeof(long) 8 sizeof(void*) 8 sizeof(PetscScalar) 8 sizeof(PetscInt) 4
Configure options: COPTFLAGS=-O3 CXXOPTFLAGS=-O3 FOPTFLAGS=-O3 FCOPTFLAGS=-O3 HIPOPTFLAGS=-O3 --with-debugging=0 --with-cc=cc --with-cxx=CC --with-fc=ftn --with-hip --with-hipc=hipcc --LIBS="-L/opt/cray/pe/mpich/8.1.23/ofi/gnu/9.1/lib -lmpi -L/opt/cray/pe/mpich/8.1.23/gtl/lib -lmpi_gtl_hsa" --download-kokkos --download-kokkos-kernels --download-suitesparse --download-hypre --download-cmake
-----------------------------------------
Libraries compiled on 2024-03-05 17:04:36 on login08
Machine characteristics: Linux-5.14.21-150400.24.46_12.0.83-cray_shasta_c-x86_64-with-glibc2.3.4
Using PETSc directory: /autofs/nccs-svm1_home1/vanellam/Software/petsc
Using PETSc arch: arch-linux-frontier-opt-gcc
-----------------------------------------

Using C compiler: cc  -fPIC -Wall -Wwrite-strings -Wno-unknown-pragmas -Wno-lto-type-mismatch -Wno-stringop-overflow -fstack-protector -fvisibility=hidden -O3
Using Fortran compiler: ftn  -fPIC -Wall -ffree-line-length-none -ffree-line-length-0 -Wno-lto-type-mismatch -Wno-unused-dummy-argument -O3
-----------------------------------------

Using include paths: -I/autofs/nccs-svm1_home1/vanellam/Software/petsc/include -I/autofs/nccs-svm1_home1/vanellam/Software/petsc/arch-linux-frontier-opt-gcc/include -I/autofs/nccs-svm1_home1/vanellam/Software/petsc/arch-linux-frontier-opt-gcc/include/suitesparse -I/opt/rocm-5.4.0/include
-----------------------------------------

Using C linker: cc
Using Fortran linker: ftn
Using libraries: -Wl,-rpath,/autofs/nccs-svm1_home1/vanellam/Software/petsc/arch-linux-frontier-opt-gcc/lib -L/autofs/nccs-svm1_home1/vanellam/Software/petsc/arch-linux-frontier-opt-gcc/lib -lpetsc -Wl,-rpath,/autofs/nccs-svm1_home1/vanellam/Software/petsc/arch-linux-frontier-opt-gcc/lib -L/autofs/nccs-svm1_home1/vanellam/Software/petsc/arch-linux-frontier-opt-gcc/lib -Wl,-rpath,/opt/rocm-5.4.0/lib -L/opt/rocm-5.4.0/lib -Wl,-rpath,/opt/cray/pe/mpich/8.1.23/ofi/gnu/9.1/lib -L/opt/cray/pe/mpich/8.1.23/ofi/gnu/9.1/lib -Wl,-rpath,/opt/cray/pe/mpich/8.1.23/gtl/lib -L/opt/cray/pe/mpich/8.1.23/gtl/lib -Wl,-rpath,/opt/cray/pe/libsci/22.12.1.1/GNU/9.1/x86_64/lib<https://urldefense.us/v3/__http://22.12.1.1/GNU/9.1/x86_64/lib__;!!G_uCfscf7eWS!aFbqY6mG8CYI7pOoYJmZ-HjqKHrVjfZmct6pBKsFlSH9QcnbP3D9Y7KXN29dYiioBs6nwMBVVlOT1Jz2OwxWqLWCqlJgQXrP$ > -L/opt/cray/pe/libsci/22.12.1.1/GNU/9.1/x86_64/lib<https://urldefense.us/v3/__http://22.12.1.1/GNU/9.1/x86_64/lib__;!!G_uCfscf7eWS!aFbqY6mG8CYI7pOoYJmZ-HjqKHrVjfZmct6pBKsFlSH9QcnbP3D9Y7KXN29dYiioBs6nwMBVVlOT1Jz2OwxWqLWCqlJgQXrP$ > -Wl,-rpath,/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/gcc-12.2.0/darshan-runtime-3.4.0-ftq5gccg3qjtyh5xeo2bz4wqkjayjhw3/lib -L/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/gcc-12.2.0/darshan-runtime-3.4.0-ftq5gccg3qjtyh5xeo2bz4wqkjayjhw3/lib -Wl,-rpath,/opt/cray/pe/dsmml/0.2.2/dsmml/lib -L/opt/cray/pe/dsmml/0.2.2/dsmml/lib -Wl,-rpath,/opt/cray/pe/pmi/6.1.8/lib -L/opt/cray/pe/pmi/6.1.8/lib -Wl,-rpath,/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/lib64 -L/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/lib64 -Wl,-rpath,/opt/cray/pe/gcc/12.2.0/snos/lib/gcc/x86_64-suse-linux/12.2.0 -L/opt/cray/pe/gcc/12.2.0/snos/lib/gcc/x86_64-suse-linux/12.2.0 -Wl,-rpath,/opt/cray/pe/gcc/12.2.0/snos/lib64 -L/opt/cray/pe/gcc/12.2.0/snos/lib64 -Wl,-rpath,/opt/rocm-5.4.0/llvm/lib -L/opt/rocm-5.4.0/llvm/lib -Wl,-rpath,/opt/cray/pe/gcc/12.2.0/snos/lib -L/opt/cray/pe/gcc/12.2.0/snos/lib -lHYPRE -lspqr -lumfpack -lklu -lcholmod -lamd -lkokkoskernels -lkokkoscontainers -lkokkoscore -lkokkossimd -lhipsparse -lhipblas -lhipsolver -lrocsparse -lrocsolver -lrocblas -lrocrand -lamdhip64 -lmpi -lmpi_gtl_hsa -ldarshan -lz -ldl -lxpmem -lgfortran -lm -lmpifort_gnu_91 -lmpi_gnu_91 -lsci_gnu_82_mpi -lsci_gnu_82 -ldsmml -lpmi -lpmi2 -lgfortran -lquadmath -lpthread -lm -lgcc_s -lstdc++ -lquadmath -lmpi -lmpi_gtl_hsa
-----------------------------------------
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20240319/ab6371f8/attachment-0001.html>


More information about the petsc-users mailing list