General QuickStart Guide for ASPIRE 2A:  https://help.nscc.sg/wp-content/uploads/2024/05/ASPIRE2A-General-Quickstart-Guide.pdf 
General QuickStart Guide for ASPIRE 2A+: https://help.nscc.sg/wp-content/uploads/ASPIRE2A_Plus_General_Quick_Start_Guide.pdf 
 

GPU Server Access and Job Submission Guide
This guide explains how to connect to the H100 GPU server through VS Code Remote SSH and how to submit GPU jobs using qsub.

1. SSH Access Path
The connection path is:
Local PC → Mac mini → GPU server → H100 GPU server
Manual SSH steps:
ssh -X ecteam@eyecodegpus.ddns.net
ssh -X minh@gpuserver3
ssh smrvmdo@aspire2p.nus.edu.sg
Do not store passwords inside this README or inside SSH config files.

2. VS Code Remote SSH Setup
Open your SSH config file:
code ~/.ssh/config
Add the following configuration:
Host mac-mini
    HostName eyecodegpus.ddns.net
    User ecteam
    ForwardX11 yes
    ServerAliveInterval 60

Host gpuserver3
    HostName gpuserver3
    User minh
    ProxyJump mac-mini
    ForwardX11 yes
    ServerAliveInterval 60

Host h100
    HostName aspire2p.nus.edu.sg
    User smrvmdo
    ProxyJump gpuserver3
    ServerAliveInterval 60
Then connect in VS Code:
Remote-SSH: Connect to Host...
Choose:
h100
VS Code will connect through:
mac-mini → gpuserver3 → h100

3. Test SSH Connection from Terminal
Before using VS Code, test the connection from terminal:
ssh -J ecteam@eyecodegpus.ddns.net,minh@gpuserver3 smrvmdo@aspire2p.nus.edu.sg
If this works, VS Code Remote SSH should also work.

4. Request Interactive GPU Session
To request an interactive session with 4 GPUs:
qsub -I -q normal -P 11714283 -l select=1:ncpus=56:ngpus=4 -l walltime=12:00:00
For 2 GPUs:
qsub -I -q normal -P 11714283 -l select=1:ncpus=28:ngpus=2 -l walltime=12:00:00
After the interactive job starts, you can run commands directly on the allocated compute node.
Example:
cd /path/to/project
conda activate your_env
python train.py

5. Submit a Batch Job
Create a job script, for example:
nano run_job.pbs
Example run_job.pbs:
#!/bin/bash
#PBS -N my_gpu_job
#PBS -q normal
#PBS -P 11714283
#PBS -l select=1:ncpus=56:ngpus=4
#PBS -l walltime=12:00:00
#PBS -j oe

cd $PBS_O_WORKDIR

source ~/.bashrc
conda activate your_env

python train.py
Submit the job:
qsub run_job.pbs

6. Useful PBS Commands
Check your running jobs:
qstat -u $USER
Check all jobs:
qstat
Delete/cancel a job:
qdel <job_id>
Example:
qdel 123456
Check job output after it finishes:
ls
cat my_gpu_job.o*

7. Example Training Job
Example project structure:
project/
├── train.py
├── run_job.pbs
└── README.md
Example run_job.pbs:
#!/bin/bash
#PBS -N train_yolo
#PBS -q normal
#PBS -P 11714283
#PBS -l select=1:ncpus=56:ngpus=4
#PBS -l walltime=12:00:00
#PBS -j oe

cd $PBS_O_WORKDIR

source ~/.bashrc
conda activate yolo_env

python train.py --batch 64 --epochs 100
Submit:
qsub run_job.pbs

8. Notes
Use qsub -I only when you need an interactive terminal session.
Use qsub run_job.pbs when you want to submit a training script and let it run in the queue.
Do not run heavy GPU training directly on the login node.
Do not put passwords in ~/.ssh/config.
If VS Code SSH fails, first test the SSH command in terminal.