Hi everybody.
I just started using lammps in cuda mode and I would like to report one “weird behaviour”
If I have a skin too short in neighbor
for example:
neighbor 2.0 multi
Most of the time I have the following error:
Test TpA
Cuda error: Cuda_Pair: before updateNmax failed in file ‘cuda_pair.cu’ in line 861 : unspecified launch failure.
And once I had:
ERROR on proc 0: Too many neighbor bins (…/neighbor.cpp:1576)
which lead me to increase skin to 10:
neighbor 10.0 multi
and it worked.
I also tried the following with:
neighbor 2.0 multi
I setted
package cuda override/bpa 1
and this time I get as error
Cuda error in file ‘cuda_wrapper.cu’ in line 180 : unspecified launch failure.
or more frequently:
ERROR on proc 0: Too many neighbor bins (…/neighbor.cpp:1576)
increasing the skin to 10 also solved the problem.
JM
Hi everybody.
I just started using lammps in cuda mode and I would like to report one
"weird behaviour"
If I have a skin too short in neighbor
for example:
neighbor 2.0 multi
Most of the time I have the following error:
Test TpA
Cuda error: Cuda_Pair: before updateNmax failed in file 'cuda_pair.cu' in
line 861 : unspecified launch failure.
And once I had:
ERROR on proc 0: Too many neighbor bins (../neighbor.cpp:1576)
which lead me to increase skin to 10:
neighbor 10.0 multi
and it worked.
have you double checked, whether this also happens on the CPU?
the issues you describe are more likely to be due to a bad model, or
an unusual model. setting the neighbor skin to such a large value is
likely to be very wasteful and inefficient. it would be important for
you to understand where this is coming from.
I also tried the following with:
neighbor 2.0 multi
I setted
package cuda override/bpa 1
and this time I get as error
Cuda error in file 'cuda_wrapper.cu' in line 180 : unspecified launch
failure.
or more frequently:
ERROR on proc 0: Too many neighbor bins (../neighbor.cpp:1576)
increasing the skin to 10 also solved the problem.
same issue, same problem, same suggestions.
axel.
I never had this probleme before with CPUs, with the exact same parameters.
Regarding my previous mail I have to correct something.
whatever the size of the skin it never worked with the option multi
and the error where:
Cuda error in file ‘cuda_wrapper.cu’ in line 180 : unspecified launch failure
with package cuda override/bpa 1
and
Test TpA
Cuda error: Cuda_Pair: before updateNmax failed in file ‘cuda_pair.cu’ in line 861 : unspecified launch failure.
otherwise.
And it worked by setting
neighbor 10.0 bin
I am working in lennard jonnes unit, Is my skin value to high ?
I never had this probleme before with CPUs, with the exact same parameters.
have you compile cuda support in single precision? single precision
calculations are much more sensitive to errors due to large forces (or
overflows).
Regarding my previous mail I have to correct something.
whatever the size of the skin it never worked with the option multi
i don't know the inner workings of the code, but from a cursory look
it seems as if USER-CUDA only supports a subset of the neighborlist.
what you you need the style "multi" for anyway?
and the error where:
Cuda error in file 'cuda_wrapper.cu' in line 180 : unspecified launch
failure
with package cuda override/bpa 1
and
Test TpA
Cuda error: Cuda_Pair: before updateNmax failed in file 'cuda_pair.cu' in
line 861 : unspecified launch failure.
otherwise.
And it worked by setting
neighbor 10.0 bin
I am working in lennard jonnes unit, Is my skin value to high ?
yes, by over an order of magnitude. unless you have a very, very unusual system.
axel.
Well before playing with cuda it was set to
neighbor 2.0 multi
I use multi because I have a polymer with bead size of 4 and other with size 1.
The one with size 4 segregate in one corner and so the density in that part is a lot smaller.