ECE Cluster Usage Policy

Institute ECE Cluster(4x CPUs compute nodes and 1 GPU node [4 × NVIDIA H200 )

IIIT Delhi | Version 1.0 | Sep 2026

ECE Cluster Job Submission Guide

CPU Queues Information

Settings CPU Queue Type
Short Medium Long
Default Time Limit 6 hr 12 hr 1 day
Maximum Time Limit 12 hr 36 hr 3 days
Max Job running per Group 3 2 1
Storage quota 4 TB per Group
Default CPU Cores 20/30 40/60 80/128
No. of available GPUs per job 1 1 1
Token Cost 0.5 per job 1 per job 2 per job
Settings GPU Queue
gpu-short gpu-long
Default Time Limit 6 hr 2 day
Max Job running per Group 2 1
Storage quota 4 TB per Group
No. of available GPUs per job 1 1
Token Cost 0.5 per job 1 per job
Queue Description
Short MIG-enabled (2x 4= 8 GPUs @ 35 GB each) GPUs (DGX H200)
Long Dedicated GPUs (2x 2H200=4 GPUs @141 GB)

CPU JOBs

1. Short queue

srun --account=it --partition=short --qos=short --ntasks=1 --pty hostname

If you need to run the job on a specific compute node.

srun --account=it --partition=short --qos=short --ntasks=1 --cpus-per-task=1 --nodelist=compute2 hostname

2. Medium Queue

srun --account=it --partition=medium --qos=medium --ntasks=1 --cpus-per-task=1 hostname

If you need to run the job on a specific compute node.

srun --account=it --partition=medium --qos=medium --ntasks=1 --cpus-per-task=1 --nodelist=compute2 hostname

3. Long Queue

srun --account=it --partition=long --qos=long --ntasks=1 --cpus-per-task=1 hostname 

If you need to run the job on a specific compute node.

srun --account=it --partition=long --qos=long --ntasks=1 --cpus-per-task=1 --nodelist=compute3 hostname 

GPU JOBs

4. Short queue

srun --account=it --partition=gpu-short --qos=gpu-short --gres=gpu:1g.35gb:1 --time=00:05:00 --pty nvidia-smi

5. Long Queue

srun --account=it --partition=gpu-long --qos=gpu-long --qos=gpu-long --ntasks=1 --gres=gpu:1 nvidia-smi 

Alternatively, you can use the job script and modify the parameters according to your requirements.

#!/bin/bash
#SBATCH --time=06:00:00
#SBATCH --job-name=Any_name
#SBATCH --account=GroupName
#SBATCH --partition=gpu-short
#SBATCH --qos=gpu-short
#SBATCH --ntasks=1
#SBATCH --gres=gpu:1g.35gb:1
#SBATCH --mem=64G

module load anaconda3-2026                  

# if you are using your own python environment then activate it here
# source $HOME/.venv/bin/activate  
python gpu_stress_train.py

Save the script as jobscript.slurm and submit using:

sbatch jobscript.slurm

Cluster commands and monitoring

  • sbatch jobscript.slurm                                                                                            ### for job submission
  • squeue                                                                                                                     ### To check running job status
  • sinfo                                                                                         ### View available partitions and compute nodes
  • scancle jobid                                                                                                         ### To cancle running job
  • sacctmgr show assoc where user=$USER format=Cluster,Account,User     ## to know your group/Account name
  • slurm-tokenctl status --account=it           ### Replace "it" with your group name to check your remaining and consumed token
  • srun --jobid= --overlap --pty nvidia-smi       ### A user may want to verify whether their submitted job is actually running on a GPU rather than only using CPU resources.
  • User Access Policy

  • Access to the cluster shall be provided only through faculty approval.
  • Each faculty member shall be provided one group account through the IT Team.
  • Individual user accounts shall be created only after approval from the respective faculty member.
  • Users shall not share passwords or login credentials.
  • The cluster shall be used only for academic and research activities.
  • Sudo access will not be given to any account.
  • Faculty will work with IT teams in deleting accounts when students graduate
  • Storage Policy

  • Each faculty group shall be allocated a 4 TB storage quota on the Parallel File System (PFS).
  • Users are responsible for:
  • (a) Removing unwanted data.
    (b)Maintaining project directories.
    (c)Taking backups of important data.
    (d)Avoiding unnecessary duplication of datasets.

    Token Policy

    To ensure the equitable utilisation of GPU resources, a token-based scheduling policy shall be implemented.

  • Every faculty group shall receive 36 tokens at the beginning of each calendar month.
  • Tokens shall reset automatically on the first day of every month.
  • Tokens shall not be carried forward to the next month.
  • Token consumption depends on the queue selected, as shown below:
  • Job Scheduling Policy

  • Jobs shall be scheduled using the Slurm Workload Manager.
  • Resources shall be allocated on a first-come, first-served basis, subject to queue priority and resource availability.
  • Jobs exceeding the allocated limits shall not be scheduled.
  • Prohibited Activities (Policy Violations)

    • Any job found running directly on the Master Node will be terminated immediately. An automated notification will be sent to the user informing them of the policy violation and advising them to submit jobs through the Slurm job scheduler.
    • Bypassing the GPU quota system.
    • Obtaining unauthorized administrative privileges.
    • Cryptocurrency mining.
    • Commercial or non-academic workloads.
    • Keeping GPUs idle intentionally.
    • Sharing login credentials.
    • Illegal or prohibited activities.

    Monitoring and Enforcement

    GPU jobs are logged, monitored and audited. Misuse may result in reduced quota, temporary suspension or permanent loss of access.

    Support and Exceptions

    • Additional GPU hours may be requested with Faculty approval.
    • User tiers are managed by administrators.
    • Contact: helpdesk@iiitd.ac.in
    Agreement

    By using the ECE Cluster you agree to comply with this policy.