Introduction

We have prepared a hal-login3 login4 machine as a login node so that users can request computational resources from hal-dgx and overdrive via slurm. This is the only way to access the DGX and overdrive nodes. 

• hal-dgx is a NVIDIA DGX A100 machine: https://www.nvidia.com/en-us/data-center/dgx-a100/
• overdrive is a NVIDIA Arm HPC Developer Kit machine: https://developer.nvidia.com/arm-hpc-devkit

How to login hal-login3login4

ssh <user_id>@hal-login3login4.ncsa.illinois.edu

Type sinfo to check the existing partitions

[dmu@hal-login3login4 ~]$ sinfo
PARTITION AVAIL  TIMELIMIT  NODES  STATE NODELIST
arm          up 15-00:00:0      1   idle overdrive
x86*         up 15-00:00:0      1   idle hal-dgx

Note: hal-login3 login4 has no a shared file system. Therefore, you can not find the same layout among these three machines. For example, users can not see their data and/or result on hal-dgx from hal-login3. user need to start a job to login to hal-dgx to access them./home system and shared /projects now with hal-dgx

Rules

1. the maximum wall time for each job is 48 hours
2. the maximum GPU one user can request is 4x.

Access to hal-dgx

You need to submit an interactive job and/or batch script to request some resources to run your jobs.

1. Interactive

Request 1x GPU along with 32x CPU cores for 4 hours

...

srun --partition=x86 --time=24:00:00 --nodes=1 --ntasks-per-node=128 --sockets-per-node=4 --cores-per-socket=16 --threads-per-core=2 --mem-per-cpu=4000 --wait=0 --export=ALL --gres=gpu:a100:4 --pty /bin/bash

2. Batch script

#!/bin/bash
#SBATCH --job-name="example"
#SBATCH --output="example.%j.%N.out"
#SBATCH --partition=x86
#SBATCH --time=1:00:00
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=32
#SBATCH --sockets-per-node=1
#SBATCH --cores-per-socket=16
#SBATCH --threads-per-core=2
#SBATCH --mem-per-cpu=4000
#SBATCH --gres=gpu:a100:1
#SBATCH --export=ALL

cd ~

echo STARTING `date`

srun hostname

3. Access Data and/or Result with sftp

1. Log on to hal-login3login4, start an interactive job with 1 CPU core

...

sftp hal-dgx.ncsa.illinois.edu

Access to overdrive

You need to submit an interactive job and/or batch script to request some resources to run your jobs.

1. Interactive

Request 1x GPU along with 40x CPU cores for 4 hours

...

srun --partition=arm --time=4:00:00 --nodes=1 --ntasks-per-node=80 --sockets-per-node=1 --cores-per-socket=80 --threads-per-core=1 --mem-per-cpu=3200 --wait=0 --export=ALL --gres=gpu:a100:2 --pty /bin/bash

2. Batch script

#!/bin/bash
#SBATCH --job-name="example"
#SBATCH --output="example.%j.%N.out"
#SBATCH --partition=arm
#SBATCH --time=1:00:00
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=40
#SBATCH --sockets-per-node=1
#SBATCH --cores-per-socket=40
#SBATCH --threads-per-core=1
#SBATCH --mem-per-cpu=3200
#SBATCH --gres=gpu:a100:1
#SBATCH --export=ALL

cd ~

echo STARTING `date`

srun hostname

3. Access Data and/or Result with sftp

1. Log on to hal-login3login4, start an interactive job with 1 CPU core

...