...
...
Primjeri
| Tip |
|---|
When you define the value of the ncpus variable in the header of the PBS script, the same value of the Kad u zaglavlju PBS skripte definirate vrijednost varijable |
CPU
MPI + OpenMP
Since the application supports hybrid parallelization, you can split MPI processes into OpenMP threadsBudući da aplikacija podržava hibridnu paralelizaciju, MPI procese možete podijeliti na OpenMP threadove.
| Tip |
|---|
| GROMACS preporučuje između 2 i 8 threadova po MPI procesu. |
GROMACS recommends between 2 and 8 threads per MPI process.U primjeru niže, aplikacija će stvoriti 8 MPI procesa, podijeljenih u 4 OpenMP threada
| Code Block | ||||||
|---|---|---|---|---|---|---|
| ||||||
#PBS -q cpu
#PBS -l select=8:ncpus=4
cd ${PBS_O_WORKDIR}
module load scientific/gromacs/2022.5-gnu
mpiexec -d ${OMP_NUM_THREADS} --cpu-bind depth gmx mdrun -pin on -v -deffnm md |
MPI
If you do not want to divide the application into OpenMP threads, you can use parallelization exclusively at the MPI process level.
Ako aplikaciju ne želite dijeliti u OpenMP threadove, možete koristiti paralelizaciju isključivo na razini MPI procesa.
U primjeru niže, aplikacija će pokrenuti 32 MPI procesaIn the example below, the application will launch 32 MPI processes.
| Code Block | ||||||
|---|---|---|---|---|---|---|
| ||||||
#PBS -q cpu
#PBS -l select=32
cd ${PBS_O_WORKDIR}
module load scientific/gromacs/2022.5-gnu
mpiexec gmx mdrun -pin on -v -deffnm md |
OpenMP
If you want to split the application exclusively into OpenMP threads, you must request one computer node, since in this case the application works with shared memoryAko aplikaciju želite dijeliti isključivo u OpenMP threadove, morate zatražiti jedan računalni čvor, budući da u ovom slučaju aplikacija radi s dijeljenom memorijom.
| Tip |
|---|
GROMACS will get the će vrijednost |
U primjeru niže, aplikacija će se pokrenuti s 32 OpenMP threadaIn the example below, the application will run with 32 OpenMP threads.
| Code Block | ||||||
|---|---|---|---|---|---|---|
| ||||||
#PBS -q cpu
#PBS -l ncpus=32
cd ${PBS_O_WORKDIR}
module load scientific/gromacs/2022.5-gnu
gmx mdrun -pin on -v -deffnm md |
GPU
Since the application supports hybrid parallelization, you can split MPI processes into OpenMP threadsAplikacija podržava rad s grafičkim procesorom, odnosno rad s više grafičkih procesora.
| Warning |
|---|
GPU iskorištenje se u slučaju korištenja više grafičkih procesora kreće se do maksimalno 50% (kod korištenja 2 GPU-a) u većem broju testiranih primjera. Iz tog razloga, razmotrite korištenje jednog grafičkog procesora i više procesorskih jezgri. GPU iskorištenje se u slučaju korištenja 1 GPU + 8 CPU kreće i do utilization in the case of using multiple graphics processors reaches a maximum of 50% (when using 2 GPUs) in a large number of tested examples.For this reason, consider using a single GPU and multiple CPU cores. GPU utilization in the case of using 1 GPU + 8 CPUs reaches up to 85%. |
Single GPU
In the example below, the application will use one graphics processor and eight processor coresU primjeru niže, aplikacija će koristiti jedan grafički procesor te osam procesorskih jezgara.
| Code Block | ||||||
|---|---|---|---|---|---|---|
| ||||||
#PBS -q gpu
#PBS -l select=1:ngpus=1:ncpus=8
cd ${PBS_O_WORKDIR}
module load scientific/gromacs/2022.5-cuda
gmx mdrun -pin on -v -deffnm md |
Multi GPU
When calling an application with Prilikom pozivanja aplikacije s mpiexec, each MPI process is assigned to one GPU (via the svaki MPI proces dodijeli se jednom grafičkom procesoru (putem varijable CUDA_VISIBLE_DEVICES variable), and each a svaka OpenMP thread nit (via the putem varijable OMP_NUM_THREADS variable) to one processor core.In the example below, the application will use two graphics processors and four processor cores for each graphics processor) jednoj procesorskoj jezgri.
U primjeru niže, aplikacija će koristiti dva grafička procesora te po četiri procesorske jezgre za svaki grafički procesor.
| Code Block | ||||||
|---|---|---|---|---|---|---|
| ||||||
#PBS -q gpu
#PBS -l select=2:ngpus=1:ncpus=4
cd ${PBS_O_WORKDIR}
module load scientific/gromacs/2022.5-cuda
mpiexec -d ${OMP_NUM_THREADS} --cpu-bind depth gmx mdrun -pin on -v -deffnm md |
...