Compute !

Contents view

Global view

Launching a computation on the platform means submitting a “job” in the queue among those available. This involves the following procedure:

  1. Cluster connexion
  2. Data transfert
  3. BATCH script creation
  4. Job submission

Context

Nodes | Waiting queues (Partitions)

Monocore job sample : monocore.slurm

Requesting a computation core on a node and 5 MB for 10 minutes. Sending an email at each stage of the job’s life.

Create a sbatch file here named monocore.slurm

#!/bin/bash                                                                                                      
#SBATCH --nodes=1
#SBATCH --partition=short
#SBATCH --ntasks-per-node=1
#SBATCH --time=00:10:00
#SBATCH --mail-type=ALL
#SBATCH --job-name=my_serial_job
#SBATCH --output=job_seq-%j.out
#SBATCH --mail-user=your.mail@your.domain
#SBATCH --mem=5M
time sleep 30
hostname

Job submission

sbatch monocore.slurm

MPI job sample

hello_mpi.c

hello_mpi.c
#include <mpi.h>
#include <stdio.h>
#include <unistd.h>

int main(int argc, char** argv) {
    MPI_Init(NULL, NULL);
    int world_size, world_rank;
    char hostname[256];

    MPI_Comm_size(MPI_COMM_WORLD, &world_size);
    MPI_Comm_rank(MPI_COMM_WORLD, &world_rank);
    gethostname(hostname, 256);

    printf("Hello from row %d on the machine %s (Total: %d processes)\n",
           world_rank, hostname, world_size);

    MPI_Finalize();
    return 0;
}

Compiling your code

ml libs/ompi gcc libs/ucx
mpicc hello_mpi.c -o hello_mpi

jobMPI.slurm

#!/bin/bash
#SBATCH --nodes=2
#SBATCH --partition=normal-amd  ## also works with --partition=normal
#SBATCH --ntasks-per-node=16
#SBATCH --time=00:10:00
#SBATCH --job-name=my_mpi_job
#SBATCH --output=mpi_job-%j.out
#SBATCH --mem=2G
#SBATCH --mail-type=ALL
#SBATCH --mail-user=your.mail@domain
ml libs/ompi gcc libs/ucx

# You can force the display of UCX information to verify that it is functioning
#export UCX_LOG_LEVEL=info

echo "Lancement avec srun :"
srun ./hello_mpi

Job submission

sbatch jobompi.slurm

OpenMP sample

omp.cc

#include <iostream>
#include <omp.h>

int main() {
    // This directive instructs the compiler to parallelize the section
    #pragma omp parallel
    {
        int id = omp_get_thread_num();
        int total = omp_get_num_threads();
        
        #pragma omp critical
        std::cout << "Thread " << id << " of " << total << " is ready !" << std::endl;
    }
    return 0;
}

Compiling your code

ml gcc
g++ -O3 -fopenmp omp.cc -o omp

job_openMP.slurm

#!/bin/bash                                                                                                 

#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=8
#SBATCH --time=04:00:00
#SBATCH --job-name=my_openmp_job
#SBATCH --mem=16M

export OMP_NUM_THREADS=$SLURM_CPUS_PER_TASK
./omp

Job submission

sbatch job_openMP.slurm

GPU usage

To use GPUs, specify the following parameter --gres=gpu:X with X the number of GPUs.
Here sbatch script “mon_script.sh” to ask 2 GPUs and 28 cores (bigpu).

#!/bin/sh
#SBATCH --job-name=tensor 
#SBATCH --partition=bigpu 
#SBATCH --gres=gpu:2 
#SBATCH --time=0:10:00 
#SBATCH --mail-type=ALL 
#SBATCH --output=job-%j.out 
#SBATCH --mem=60G 
#SBATCH --nodes=1 
#SBATCH --ntasks-per-node=28

hostname
python hello.py
  • To submit the job, use the following command :
sbatch mon_script.sh
  • Interactif sample :
    srun --ntasks=1 --mem=4G --gres=gpu:1 --time=1:00:00 --partition=bigpu --pty /bin/bash
  • nvidia-smi command shows you GPUs usage.