SLURM-Job Workflow

SLURM-Job Workflow#

In this section we explain the workflow from two complementary perspectives:

  1. Platform-level view — what happens when you submit a job from DesignSafe

  2. Scheduler-level view — how SLURM actually queues, schedules, and runs the job

Where Job Submission Happens

SLURM job submission commands (such as sbatch) are issued from a login or submission node, not from compute nodes. When using DesignSafe, this step is handled automatically for you. When working directly on a TACC system, you typically access a login node via SSH to prepare files and submit jobs. A dedicated section later in this book covers job submission environments and workflows in detail.

1. Platform-level view: SLURM-Job Submission Workflow#

How Are Jobs Run for DesignSafe on TACC? – Big Picture

TACC (Texas Advanced Computing Center) provides the high-performance computing (HPC) resources that DesignSafe uses to run jobs. Specifically, jobs submitted from DesignSafe are executed on TACC’s systems — typically Stampede3 or similar clusters.

Step-by-step Process:

  1. Log into DesignSafe

    • You access it through https://www.designsafe-ci.org

    • You don’t need to interact with TACC directly unless you want to.

  2. Use the Workspace

    • DesignSafe has an interactive workspace:

      • Jupyter Notebooks

      • MATLAB

      • SimCenter tools

      • Other scientific apps like OpenSees, ADCIRC, etc.

    • These are all powered on the backend by TACC resources.

  3. Submit Job When you submit a job via DesignSafe:

    • The job is packaged with:

      • Your script

      • Input data

      • Parameters (like CPU cores, memory, wall time)

    • DesignSafe translates this into a SLURM job (TACC uses SLURM as its job scheduler).

    • The job is then sent to a TACC queue (often the designsafe or community queue on Stampede3).

  4. Execution on TACC

    • TACC runs the job exactly like any HPC job using SLURM.

    • The job runs on actual compute nodes (on Stampede3, for example).

    • Output files are saved in your DesignSafe project or data workspace.

  5. Results

    • Once complete, results are sent back to DesignSafe automatically.

    • You can view, download, or analyze the data directly in the platform.

1. Scheduler-level view: SLURM-Job Submission Workflow#

How does SLURM manage your job

  1. Job Submission

You submit a job with a job script that specifies:

  • Number of nodes and cores (e.g., 32 cores on 4 nodes – –ntasks=32, –nodes=4)

  • Maximum runtime (e.g., 4 hours – –time=04:00:00)

  • Memory requirements (e.g., –mem=16G)

  • Partition (queue) to submit to (NODE TYPE – –partition=skx)

  • Allocation (project)

  • Submit with: sbatch job_script.sh

  1. Job Enters the Queue

  • Your job enters the SLURM queue and is assigned a priority.

  • You can monitor the queue status using SLURM commands (e.g., squeue, sacct, etc.)

  1. SLURM Schedules the Job SLURM decides when and where to run your job based on:

  • Requested resources: Number of nodes, memory, and runtime.

  • Queue priority: System policies may prioritize shorter/smaller jobs.

  • Current system load: Jobs may wait until required nodes become free.

  1. Job Execution

  • Once sufficient resources are available, SLURM starts the job.

  • Your job runs on the assigned compute nodes for the allocated time.

  1. Job Completion When the job finishes:

  • Output files and logs (e.g., SLURM-.out) are generated.

  • You can check results and performance.

Contents