How to request GPUs for batch jobs on ARC

From RCSWiki
Revision as of 21:56, 20 September 2023 by Lleung (talk | contribs) (Added categories)
Jump to navigation Jump to search

GPUs on ARC

Currently, there are three partitions on ARC that have usable CUDA GPUs, gpu-v100, gpu-a100, and bigmem. ARC's configuration keeps changing depending on addition of new hardware and retiring of the old hardware, so you have to check available hardware on ARC before starting a new set of computations.

  • First, check the available partitions with GPUs on ARC following this article:
How to find available partitions on ARC.
  • Then, check the resource limits for the partitions you are intended to use, as it is explained here:
How to find current limits on ARC.


The main partitions that contain GPUs on ARC at the moment of writing are gpu-v100 and gpu-a100. These partitions contain V100 and A100 nVidia GPUs, correspondingly. The maximum time limit for jobs on both the partitions is 24 hours.


Please note, that GPU partitions should only be used to run jobs that required GPUs. Any CPU-only computations have to be directed to CPU-only partitions.

Examples

Example 1, one GPU

Request 1 V100 GPU, 4 CPUs and 16 GB of RAM for 1 hour.

V100 GPUs can be got from the gpu-v100 partitions. This is the gpu-1.slurm jobs scripts that requests these resources:

#! /bin/bash
# ====================================
#SBATCH --job-name=GPU_test
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=4
#SBATCH --mem=16GB
#SBATCH --time=0-01:00:00
#SBATCH --gres=gpu:1
#SBATCH --partition=gpu-v100
# ====================================

# Check the GPUs with the nvidia-smi command.
nvidia-smi

The script only runs the nvidia-smi command to make sure that the GPU has been provided by SLURM.

Example 2, two GPUs on one node

Request 2 A100 GPU, 8 CPUs and 64 GB of RAM for 2 hour.

A100 GPUs can be got from the gpu-a100 partitions. This is the gpu-2.slurm jobs scripts that requests these resources:

#! /bin/bash
# ====================================
#SBATCH --job-name=2GPU_test
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=8
#SBATCH --mem=64GB
#SBATCH --time=0-02:00:00
#SBATCH --gres=gpu:2
#SBATCH --partition=gpu-a100
# ====================================

# Check the GPUs with the nvidia-smi command.
nvidia-smi

The script only runs the nvidia-smi command to make sure that the GPU has been provided by SLURM.

Example 3, four GPUs on two nodes

Request 4 V100 GPU for one job for 2 hours.

The V100 GPUs are in the gpu-v100 partition. As the arc.hardware shows, each node has 2 GPUs, not 4 as we want. Therefore, to get 4 GPUs we have to use 2 compute nodes from this partition.


Note that before proceeding with the job, one has to understand, that for this request to make sense, the code that is going to run in the job must support distributed execution, when computation spreads several nodes.

Even though, you may believe that the code can run on multiple nodes and can use multiple GPUs for performance benefit, it has to be tested and confirmed, as well as the performance benefit has to be measured. You have to confirm that the computation scales well with increasing number of nodes, as well as the number of GPUs.

Typically, this is done by running some test case using 1, 2, 4 GPUs and comparing the time that was required to obtain the result.

Even though the code itself may be capable of running in a distributed fashion, it may require some trying before the code works the way you expect it to run. Running the test case will show if you are doing it right, by providing actual performance numbers.


We still have to request the CPU part of the resource request, in this example, we request 8 CPUs and 64 GB of RAM on each node.

The GPUs are also requested per node, thus, the GPU request is --gres=gpu:2.


This is the gpu-4.slurm jobs scripts that requests these resources:

#! /bin/bash
# ====================================
#SBATCH --job-name=4GPU_test
#SBATCH --nodes=2
#SBATCH --ntasks=2
#SBATCH --cpus-per-task=8
#SBATCH --mem=64GB
#SBATCH --time=0-02:00:00
#SBATCH --gres=gpu:2
#SBATCH --partition=gpu-v100
# ====================================

# Report allocated nodes.
echo "Alocated nodes: $SLURM_JOB_NODELIST"

# Check the GPUs on each of the allocated nodes with the nvidia-smi command.
for NODE in `scontrol show hostnames`; do
    echo $NODE; ssh $NODE "nvidia-smi -L"
done

The script only runs the nvidia-smi command on each of the allocated nodes to make sure that all 4 GPU have been provided by SLURM.


The way you are going to start your distributed GPU computation depends on the software, and you will have to find this out.

Checking GPU utilization

Before running all your productions jobs you have to run several test jobs to make sure that the requested resources are properly used. You really have to make sure that the GPU(s) your job requests are actually used.

Software that often used with GPUs


Links

How-Tos