Memory leak in GPU build?
Posted: Wed Mar 08, 2017 11:57 pm
The memory used by Vasp running on GPUs grows as an optimisation runs.
I am doing relatively simple calculations, PBE+U on a 215 atom cell, ionic relaxation with IBRION=1 & ISIF=0. Running on 2 K80 GPUs tied to 6 CPUs, 6 MPI processes.
The queuing system imposes strict memory limits. If I give it:
30 GB the job crashes out of memory at the start of the 2nd geometry.
38 GB the job crashes out of memory in the scf of the 7th geometry.
45 GB the job crashes out of memory in the scf of the 11th geometry.
The memory use is clearly growing as the optimisation proceeds. Running the same job on 28 CPU cores gives me stable memory use of 25 GB, so this seems restricted to the GPU features. Any other similar reports?
(This looks like a bug to me, so I'm posting it here. I guess it could equally have gone in the GPU forum.)
I am doing relatively simple calculations, PBE+U on a 215 atom cell, ionic relaxation with IBRION=1 & ISIF=0. Running on 2 K80 GPUs tied to 6 CPUs, 6 MPI processes.
The queuing system imposes strict memory limits. If I give it:
30 GB the job crashes out of memory at the start of the 2nd geometry.
38 GB the job crashes out of memory in the scf of the 7th geometry.
45 GB the job crashes out of memory in the scf of the 11th geometry.
The memory use is clearly growing as the optimisation proceeds. Running the same job on 28 CPU cores gives me stable memory use of 25 GB, so this seems restricted to the GPU features. Any other similar reports?
(This looks like a bug to me, so I'm posting it here. I guess it could equally have gone in the GPU forum.)