Frequently Asked Questions
Common questions about accounts, logging in, jobs, GPUs, tokens, storage and software on the ECE cluster. Type in the box to search.
1. Account & access
How do I get an account on the ECE cluster?
Send a request email to helpdesk@iiitd.ac.in with the approval of ECE Faculty/HOD
How do I change my password?
Log in to the cluster and run:
passwd
Enter your current password, then the new one twice. Nothing appears on screen while you type; this is normal.
Forgotten your password? send an email to helpdesk@iiitd.ac.in
Can I connect from home or from outside the campus?
The ECE cluster can be accessable from the IIIT Delhi network.
whether off-campus access is allowed through the institute VPN, Please visit the website https://it.iiitd.edu.in/static/VPN.pdf for more details and documentation
Can I share my account with a lab-mate?
No. Every user must have their own account. You are responsible for everything run from your account, so never share your password. Lab-mates can share files through the group's storage instead.
What happens to my account and data when I graduate or leave?
Accounts are disabled/deleted after leaving/graduating so copy everything you need to your own computer, or hand it to your supervisor, before you leave.
2. Login problems
I get “Connection timed out” or “No route to host”. What is wrong?
Your computer cannot reach the cluster. Check that:
- you are on the IIIT Delhi network (or connected to the VPN, if allowed);
- you typed the address correctly:
ssh username@192.168.4.14; - the cluster is not down for maintenance.
I get “Permission denied, please try again”.
The username or password is wrong, or your account is not active yet. Usernames are case-sensitive. If it still fails after three careful tries, contact the administrator rather than retrying many times.
I see “WARNING: REMOTE HOST IDENTIFICATION HAS CHANGED!”. What should I do?
This appears when the cluster's security key has changed, for example after maintenance. If the administrator has announced a change, remove the old key on your own computer and log in again:
ssh-keygen -R 192.168.4.14
This works in Linux/macOS Terminal and in Windows Command Prompt. MobaXterm asks you to accept the new key instead.
My connection dropped. Did my work stop?
- Batch jobs (
sbatch) keep running. Check them withsqueue -u $USERafter logging in again. - Interactive jobs (
srun) stop when the connection drops.
For anything longer than a few minutes, use sbatch.
3. Jobs & queues
Which queue should I use?
Pick the shortest queue that fits your job's run time. Shorter queues cost fewer tokens and allow more jobs at once.
| Queue | Default time | Maximum time | Jobs running per group | Token cost |
|---|---|---|---|---|
short | 6 hr | 12 hr | 3 | 0.5 per job |
medium | 12 hr | 36 hr | 2 | 1 per job |
long | 1 day | 3 days | 1 | 2 per job |
For GPU work use gpu-short or gpu-long. See GPUs. Full limits are on the Home page.
How do I submit a job?
Write a job script, for example myjob.sh:
#!/bin/bash #SBATCH --time=06:00:00 #SBATCH --job-name=testingjob #SBATCH --account=it #SBATCH --partition=gpu-short #SBATCH --qos=gpu-short #SBATCH --ntasks=1 #SBATCH --gres=gpu:1g.35gb:1 #SBATCH --mem=64G echo "Job $SLURM_JOB_ID on $(hostname), started $(date)" nvidia-smi # --- Python environment (see README for the one-time setup) --- module load anaconda3-2026 # Record GPU utilisation, memory, power and temperature every 60 s # (shows N/A for utilisation on a MIG slice - that is normal) python gpu_stress_train.py
Then submit it with sbatch myjob.sh. The Submitting jobs video walks through every step.
--partition and --qos.Why is my job stuck in “PD” (pending)?
Run squeue -u $USER and look at the last column (NODELIST(REASON)):
| Reason | What it means |
|---|---|
Resources | The nodes or GPUs you asked for are busy. The job starts when they free up. |
Priority | Other jobs are ahead of yours in the queue. |
A reason containing QOS or Grp and Jobs | Your group has reached its limit of running jobs for that queue. It starts when one of your group's jobs finishes. |
A reason containing Billing or Grp and Minutes | Your group's token allowance may be used up. See Tokens. |
ReqNodeNotAvail | The node you asked for (--nodelist) is down, reserved or under maintenance. |
Asking for less (fewer CPUs, less memory, shorter time, a MIG slice instead of a full GPU) usually makes a job start sooner.
My job was killed with “TIMEOUT” or “DUE TO TIME LIMIT”.
The job ran longer than its time limit. Ask for more time with #SBATCH --time=HH:MM:SS (or D-HH:MM:SS), up to the queue's maximum. If the job needs longer than the maximum, move to a longer queue or make your program save checkpoints so it can resume.
My job failed with “OUT_OF_MEMORY” or “oom-kill”.
The job used more memory than it asked for. Request more with #SBATCH --mem=32G, for example. To see how much a finished job really used:
sacct -j <jobid> --format=JobID,JobName,State,Elapsed,MaxRSS,ExitCode
MaxRSS is the peak memory. Ask for a little more than that next time.
How do I check my running and past jobs?
squeue -u $USER # jobs waiting or running now sacct -u $USER --starttime=today # today's jobs, including finished ones sacct -j <jobid> # one job's status and exit code scontrol show job <jobid> # full details of a waiting or running job scancel <jobid> # cancel a job
A job's printed output is in the file set by --output (for example myjob_12345.out) in the folder you submitted from.
How do I get an interactive shell on a compute node?
srun --account=GroupName --partition=short --qos=short --ntasks=1 --cpus-per-task=2 --time=01:00:00 --pty bash
Your prompt moves to a compute node (for example compute1). Type exit when done so the resources are freed. The session ends if your connection drops.
Can I run my program directly on the login node?
No. The login node is shared by everyone and is only for editing files, copying data, compiling small programs and submitting jobs. Programs that run on the login node for a long time or use a lot of CPU or memory may be stopped without warning. Always use srun or sbatch.
How do I run the same script many times with different inputs?
Use a job array. Each task gets its own number in $SLURM_ARRAY_TASK_ID:
#SBATCH --array=1-10
python my_script.py --input data_${SLURM_ARRAY_TASK_ID}.txt
Each array task counts as a separate job for your group's job limit.
4. GPUs
How do I ask for a GPU?
Use a GPU queue and add --gres:
# one full NVIDIA H200 (about 141 GB memory) #SBATCH --partition=gpu-long #SBATCH --qos=gpu-long #SBATCH --gres=gpu:1 # one 35 GB MIG slice of an H200 #SBATCH --partition=gpu-short #SBATCH --qos=gpu-short #SBATCH --gres=gpu:1g.35gb:1
Should I use a full H200 or a MIG slice?
- MIG slice (
1g.35gb): testing, debugging, inference, and training small or medium models that fit in about 33 GB of GPU memory. Usually starts sooner. - Full H200: large models or batches that need more than a slice's memory, or the full GPU's speed.
Start on a slice. Move to a full GPU only if you run out of GPU memory.
Why does nvidia-smi show GPUs that are not mine?
nvidia-smi can list other GPUs on the node, but your job can only use the GPU or MIG slice SLURM gave it. With a MIG slice, your slice appears under “MIG devices”. Inside your program, use the default device (cuda:0 in PyTorch); SLURM maps it to your allocation.
Which CUDA version can I use?
The GPU node runs NVIDIA driver 595.71.05, which supports CUDA up to 13.2. Libraries built for older CUDA versions (for example a PyTorch build for CUDA 12.x) also work with this driver.
5. Tokens
What is a token?
Tokens are your group's allowance for running jobs. Each job costs tokens according to its queue: short 0.5, medium 1, long 2 per job. GPU queue costs are on the Home page. Tokens keep usage fair between groups.
How many tokens does my group get, and when do they reset?
Each group currently gets 36 tokens per month (shown as “Monthly” in the report at login).
Token will be renew every month
How do I check how many tokens my group has left?
The SLURM Token Usage Report appears every time you log in, with columns Monthly, Used, Remaining and Enabled for each account.
However you can run the command "slurm-tokenctl status --account=it" ### Replace "it" with your group name
What happens when my group runs out of tokens?
jobs stay pending or are rejected until the next reset
6. Storage & files
How much storage do we get, and how do I check usage?
Each group has a 4 TB storage quota, shared by all its members. Your group's usage is shown at login under “Disk quotas for group”.
To find what is using space in your home folder:
du -sh ~/* | sort -h
How do I copy files to and from the cluster?
Run these on your own computer:
scp myfile.txt username@192.168.4.14:~/ # upload a file scp username@192.168.4.14:~/result.txt . # download a file rsync -avP data/ username@192.168.4.14:~/data/ # upload a folder
Windows users can drag and drop in MobaXterm's left panel or in WinSCP. See the Windows video.
My large upload was interrupted. Do I have to start again?
No. Use rsync with -P and run the same command again. It skips files already copied and continues partly copied ones:
rsync -avP bigdata/ username@192.168.4.14:~/bigdata/
7. Software
What software is installed on the cluster?
module avail # list available software module load python # load one module list # show what is loaded
Put the same module load lines in your job script, before your program runs.
How do I install my own Python packages?
You cannot install system-wide (no sudo), but you can install into your own folder. A virtual environment is the simplest way:
python3 -m venv ~/myenv # create once source ~/myenv/bin/activate # use it (also add this line to job scripts) pip install numpy torch
Pre-installed anaconda3-2026 is available, please load the module by following command module load anaconda3-2026
Can I use Docker or containers?
Docker is generally not available on shared clusters for security reasons.
How do I ask for new software to be installed?
[Write an email to helpdesk@iiitd.ac.in] Include the software name, version and a link to it.
8. Rules & help
What is not allowed on the cluster?
- Running heavy work on the login node.
- Sharing accounts or passwords.
- Using the cluster for work unrelated to research or teaching, e.g. cryptocurrency mining.
- Trying to get around queue limits or other users' jobs.
Breaking the rules can lead to jobs being cancelled or the account being suspended. The full policy is on the Home page.
How do I report a problem?
Email [helpdesk@iiitd.ac.in] and include:
- your username and group (account);
- the job ID, if it is about a job;
- the exact command you ran or your job script;
- the full error message (copy the text rather than a photo, if you can);
- when it happened.
Clear reports get answered much faster.