Accessing Kempner GPUs by All FASRC Users

1.5. Accessing Kempner GPUs by All FASRC Users#

Priority access to the Kempner AI cluster is granted based on an ongoing, direct relationship with the Kempner Institute and is intended exclusively to support mission-aligned research.

However, the Kempner Institute and FAS Research Computing (FASRC) want to ensure that GPU resources are fully utilized. Thus, if any resources on the Kempner AI cluster (kempner, kempner_h100, or kempner_requeue partitions) are not being used by a Kempner user, it is accessible to all users in the FASRC Cannon cluster environment through the gpu_requeue partition. The gpu_requeue partition has a preemption policy, meaning that jobs are preempted by SLURM when a high-priority job is submitted.

Conversely, some Kempner users also have access to other FASRC GPUs via the gpu_requeue partition. You can see the full list of available GPUs on the FASRC website.

1.5.1. Interactive Jobs#

Use the following command to request an interactive session with H100 GPUs in the FASRC Cannon cluster environment,

salloc -p gpu_requeue -C h100  -N 1 -n 24 -t 0-01:00 --mem=375G --gres=gpu:1

Use the following command to request an interactive session with A100 GPUs in the FASRC Cannon cluster environment,

salloc -p gpu_requeue -C a100  -N 1 -n 16 -t 0-01:00 --mem=250G --gres=gpu:1

1.5.2. Batch Jobs#

Use the following SLURM submission script to submit a batch job using H100 GPUs on the gpu_requeue partition.

#!/bin/bash

#SBATCH -J job_name                # job name
#SBATCH -p gpu_requeue             # partition (queue)
#SBATCH -N 1                       # number of nodes
#SBATCH -n 24                      # number of cores , A100: 64 cores, H100: 96 cores
#SBATCH --gres=gpu:1               # Number of GPUs per node
#SBATCH --mem 375G                 # Memory per node, H100: 1.5 TB, A100: 1 TB RAM
#SBATCH -t 00-07:00                # time (D-HH:MM)
#SBATCH --constraint=h100          # Choose H100 GPU or a100 for A100 GPU
#SBATCH -o job.%N.%j.out           # STDOUT
#SBATCH -e job.%N.%j.err           # STDERR

# Code commands to run your program
date
# python my_code.py

Use the following submission script For A100 GPUs,

#!/bin/bash

#SBATCH -J job_name                # job name
#SBATCH -p gpu_requeue             # partition (queue)
#SBATCH -N 1                       # number of nodes
#SBATCH -n 16                      # number of cores , A100: 64 cores, H100: 96 cores
#SBATCH --gres=gpu:1               # Number of GPUs per node
#SBATCH --mem 250G                 # Memory per node, H100: 1.5 TB, A100: 1 TB RAM
#SBATCH -t 00-07:00                # time (D-HH:MM)
#SBATCH --constraint=a100          # Choose H100 GPU or a100 for A100 GPU
#SBATCH -o job.%N.%j.out           # STDOUT
#SBATCH -e job.%N.%j.err           # STDERR

# Code commands to run your program
date
# python my_code.py
Contents