Skip to content

skydocs.md

Latest commit

 

History

History
254 lines (167 loc) · 5.48 KB

skydocs.md

File metadata and controls

254 lines (167 loc) · 5.48 KB

🪐 Skyriver Orientation

Welcome! You now have access to our Slurm-based compute cluster.

Here's a quick guide to get you productive fast.

Cluster realtime usage and details available at : (Skyriver page)

📍 1. You are here

Username and Password will be sent through

ssh <username>@skyriver.nri.bcm.edu

The authenticity of host 'x.x.x.x (x.x.x.x)' can't be established.
ECDSA key fingerprint is asdfasdfo80sad8f7a9sd0f7a89sdf0987asdfo87a6sdf.
Are you sure you want to continue connecting (yes/no/[fingerprint])?

type yes here

and then you will be asked for password 
<enter password>
  • You’re logging into the head node: leia1
  • This node is for job submission
  • Do not run heavy computation on the head node

💻 2. Auto Compute Shell Prompt

  1. Termius (quick start) is good for quick SSH session
  2. VS Code is awesome with both terminal and IDE

When you SSH into the cluster or open a terminal:

Do you want to enter a compute shell? [Y/n]
  • Press Y or hit Enter to start a Slurm interactive shell on a worker node
  • Press N to stay on the head node for lightweight tasks (like file browsing)

Note: You will not be able to create srun session through this auto compute shell. If you plan to run srun, please press "N" and stay on head node.

⚙️ 3. Running Jobs

✅ Option A: Interactive Shell (good for exploration & development)

srun --pty --job-name=devshell --time=01:00:00 bash

Note: If you don’t specify a partition, jobs will run in the default queue (defq15). For GPU jobs, use --partition=a30q or --partition=a100q. For longer-running CPU jobs, use --partition=defq610.

✅ Option B: Batch Job (best for long-running or scripted jobs)

Create a file job.slurm:

#!/bin/bash
#SBATCH --job-name=myjob
#SBATCH --output=output.txt
#SBATCH --partition=defq15
#SBATCH --time=02:00:00
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=4
#SBATCH --mem=16G

module load conda
conda activate aim
python my_script.py

Submit it:

sbatch job.slurm

🧭 3.1. Understanding Queues

Different queues (also called partitions) are available for various workloads:

Queue Description Typical Usage
defq15 Default queue for general CPU jobs Standard analysis or short jobs (<15 hrs)
defq610 Long-duration CPU queue Multi-day runs or batch pipelines (>15 hrs)
a30q GPU queue with NVIDIA A30 GPUs Moderate GPU workloads (e.g., model training, inference)
a100q GPU queue with NVIDIA A100 GPUs High-performance GPU workloads (large-scale deep learning)

3.2 Using Visual studio code and wsl

Please utilize this helper script (skycode.sh)

Windows users can run bash scripts using wsl : (wsl usage)

✅ Example Job Submissions

CPU Job (Default Queue)

#!/bin/bash
#SBATCH --job-name=cpu_job
#SBATCH --output=cpu_output.txt
#SBATCH --partition=defq15
#SBATCH --time=04:00:00
#SBATCH --cpus-per-task=4
#SBATCH --mem=16G

python my_script.py

Long Job

#SBATCH --partition=defq610
#SBATCH --time=72:00:00

GPU Job (A30)

#SBATCH --partition=a30q
#SBATCH --gres=gpu:1
#SBATCH --time=08:00:00

GPU Job (A100)

#SBATCH --partition=a100q
#SBATCH --gres=gpu:1
#SBATCH --time=04:00:00

✅ Option C: Using Reservations (for dedicated time slots)

If you need a dedicated reservation (for example, for teaching sessions, workshops, or large-scale coordinated runs), please email the cluster admin with the following details:

  • Requested date and time window
  • Duration of the reservation
  • Number of nodes and resources needed
  • Purpose or event name

🕐 Please submit reservation requests at least one week prior to your desired slot. Reservations are approved on a case-by-case basis depending on cluster availability.

To check existing reservations:

scontrol show reservation

Once your reservation is confirmed, you can use it as follows:

Interactive session with reservation:

srun --pty --job-name=devres --time=01:00:00 --reservation=<reservation_name> bash

Batch job with reservation:

#SBATCH --reservation=<reservation_name>

Then submit as usual:

sbatch job.slurm

🧪 4. Monitor Jobs

squeue -u $USER     # Show your jobs
scancel <jobid>     # Cancel a job

Jobs from the auto-shell show up as shlurm by default.

🧰 5. Module System

To load software:

module avail
module load anaconda3/3.11

Use your own modulefiles:

module load use.own
module avail

📦 6. Conda Environments

Create and use your own:

conda create -n myenv python=3.9
conda activate myenv

🛑 7. What Not To Do

  • Don’t run jobs directly on the head node
  • Don’t leave idle compute shells open

🙋 Need Help?

Contact your cluster admin or run:

man srun
man sbatch

Or just ask in your internal support channel.