Hi Pierre,
> I tried to rebuild the optimized PETSc library by changing several
options and ran:
mpirun -np 2 ./ex19 -cuda_show_devices -dm_mat_type aijcusp -dm_vec_type
cusp
-ksp_type fgmres -ksp_view -log_summary -pc_type none
-snes_monitor_short -snes_rtol 1.e-5
Options used:
--with-pthread=1 -O3 -> crash
--with-pthread=0 -O2 -> crash
--with-debugging=1 --with-pthread=1 -O2 -> OK
So --with-debugging=1 is the key to avoid the crash. Not
good for the performance of course...
thanks for the input. I get crashes even when using --with-debugging=1,
and valgrind spits out a couple of errors as soon as
-dm_vec_type cusp
is provided. I'll keep digging, the error is somewhere in the VecScatter
routines when using CUSP...
Best regards,
Karli
If it can helps,
Pierre
Previously, I had noticed strange behaviour when running the GPU code
with the threadComm package. It might be worth trying to disable that
code in the build to see if the problem persists?
-Paul
On Tue, Jan 14, 2014 at 9:19 AM, Karl Rupp <[email protected]
<mailto:[email protected]>> wrote:
Hi Pierre,
>> I could reproduce the problem and also get some uninitialized
variable
warnings in Valgrind. The debug version detects these
errors, hence
you only see the errors in the debug build. For the
optimized build,
chances are good that the computed values are either wrong
or may
become wrong in other environments. I'll see what I can do
when I'm
again at GPU machine tomorrow (parallel GPU debugging via
SSH is not
great...)
Sorry, I mean:
Parallel calculation on CPU or GPU run well with PETSc non
optimized library
Parallel calculation on GPU crashes with PETSc optimized
library (on CPU
it is OK)
The fact that it happens to run in one mode out of {debug,
optimized} but not in the other is at most a lucky coincidence,
but it still means that this is a bug we need to solve :-)
I could add that the "mpirun -np 1 ex19" runs well for all
builds on CPU
and GPU.
I see valgrind warnings in the vector scatter routines, which is
likely the reason why it doesn't work with multiple MPI ranks.
Best regards,
Karli
--
*Trio_U support team*
Marthe ROUX (Saclay)
Pierre LEDAC (Grenoble)
