The GPU supports double precision and I didn't explicitly tell PETSc to use
float when compiling, so
I guess it uses double? What's the easiest way to check?
Barry, running -ksp_view shows that the solver options are the same for CPU and
GPU. The only
difference is the coarse grid solver for gamg ("the package used to perform
factorization:") which
is petsc for CPU and cusparse for GPU. I tried forcing the GPU to use petsc via
-fieldsplit_0_mg_coarse_sub_pc_factor_mat_solver_type, but then ksp failed to
converge
even on the first topology optimization iteration.
-ksp_view also shows differences in the eigenvalues from the Chebyshev
smoother. For example,
GPU:
Down solver (pre-smoother) on level 2 -------------------------------
KSP Object: (fieldsplit_0_mg_levels_2_) 1 MPI process
type: chebyshev
eigenvalue targets used: min 0.109245, max 1.2017
eigenvalues provided (min 0.889134, max 1.09245) with
CPU:
eigenvalue targets used: min 0.112623, max 1.23886
eigenvalues provided (min 0.879582, max 1.12623)
But I guess such differences are expected?
/Carl-Johan
From: Matthew Knepley <knep...@gmail.com>
Sent: den 30 oktober 2022 22:00
To: Barry Smith <bsm...@petsc.dev>
Cc: Carl-Johan Thore <carl-johan.th...@liu.se>; petsc-users@mcs.anl.gov
Subject: Re: [petsc-users] KSP on GPU
On Sun, Oct 30, 2022 at 3:52 PM Barry Smith
<bsm...@petsc.dev<mailto:bsm...@petsc.dev>> wrote:
In general you should expect similar but not identical conference behavior.
I suggest running with all the monitoring you can.
-ksp_monitor_true_residual -fieldsplit_0_monitor_true_residual
-fieldsplit_1_monitor_true_residual and compare the various convergence between
the CPU and GPU. Also run with -ksp_view and check that the various solver
options are the same (they should be).
Is the GPU using float or double?
Matt
Barry
On Oct 30, 2022, at 11:02 AM, Carl-Johan Thore via petsc-users
<petsc-users@mcs.anl.gov<mailto:petsc-users@mcs.anl.gov>> wrote:
Hi,
I'm solving a topology optimization problem with Stokes flow discretized by a
stabilized Q1-Q0 finite element method
and using BiCGStab with the fieldsplit preconditioner to solve the linear
systems. The implementation
is based on DMStag, runs on Ubuntu via WSL2, and works fine with PETSc-3.18.1
on multiple CPU cores and the following
options for the preconditioner:
-fieldsplit_0_ksp_type preonly \
-fieldsplit_0_pc_type gamg \
-fieldsplit_0_pc_gamg_reuse_interpolation 0 \
-fieldsplit_1_ksp_type preonly \
-fieldsplit_1_pc_type jacobi
However, when I enable GPU computations by adding two options -
...
-dm_vec_type cuda \
-dm_mat_type aijcusparse \
-fieldsplit_0_ksp_type preonly \
-fieldsplit_0_pc_type gamg \
-fieldsplit_0_pc_gamg_reuse_interpolation 0 \
-fieldsplit_1_ksp_type preonly \
-fieldsplit_1_pc_type jacobi
- KSP still works fine the first couple of topology optimization iterations but
then
stops with "Linear solve did not converge due to DIVERGED_DTOL ..".
My question is whether I should expect the GPU versions of the linear solvers
and pre-conditioners
to function exactly as their CPU counterparts (I got this impression from the
documentation),
in which case I've probably made some mistake in my own code, or whether there
are other/additional
settings or modifications I should use to run on the GPU (an NVIDIA Quadro
T2000)?
Kind regards,
Carl-Johan
--
What most experimenters take for granted before they begin their experiments is
infinitely more interesting than any results to which their experiments lead.
-- Norbert Wiener
https://www.cse.buffalo.edu/~knepley/<https://eur01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fwww.cse.buffalo.edu%2F~knepley%2F&data=05%7C01%7Ccarl-johan.thore%40liu.se%7C8113f968516a44cce6ba08dabab9b28e%7C913f18ec7f264c5fa816784fe9a58edd%7C0%7C0%7C638027603971893036%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=gQ45cdsgo404NeBzZ7e6c5zNXhYOy39ZPfZzqtGDaDk%3D&reserved=0>
