SCRP nodes utilizes the Slurm Workload Manager to allocate compute node resources.

Partitions

Compute nodes are grouped into partitions, which you can think of as queues:

Partition Nodes Resource per Node Limits
scrp* scrp-node-[1-4,6,8,9,12,14-18] CPU: 16-64 cores
RAM: 60-497GiB
GPU: 0-4		 1 node/job
32 CPU/job
large scrp-node-[1-4,6-8,12,14-18] CPU: 16-192 cores
RAM: 60GiB-1.5TiB		 4 GPU/user
a100 scrp-node-10,13,21 CPU: 64-128 cores
RAM: 1-2TiB
GPU: A100 80GB x 4		 32 CPU/job
512GB RAM/job
2 GPU/user
jrc scrp-node-11 CPU: 128 cores
RAM: 747GiB
GPU: A100 80GB x 2		 JRC members only
g2007 scrp-node-5 CPU: 128 cores
RAM: 747GiB
 g2007 members only

*Default partition.

The scrp partition is accessible to all users, while the large and a100 partitions is only accessible to faculty members and research postgraduate students. The jrc parititon is only accessible to JRC members.

Resource Limits

To check what resources you can use, type in a terminal:

qos

The QoS (for “Quality of Service”) field tells you the resource limits applicable to you:

QoS Logical CPUs GPUs Max. Job Duration
c4g1 4 1 1 Day
c16g1 16 1 5 Days
c32g4 128 4 5 Days
c32g8 128 8 5 Days
c16-long 16 1 30 Days

Default job duration is 1 day regardless of the QoS you use.

Compute Node Status

To check the status of the compute nodes, type:

scrp-info

Use Slurm’s sinfo if you want to see a different set of information.

Run a Job Immediately

Predefined Shortcuts

SCRP has two sets of predefined shortcuts to provide quick access:

  • compute-[1-4,8,16,32,64,128] requests the specified number of logical CPUs from a compute node. Each CPU comes with 2GB of RAM.
  • gpu-[1-4] requests 1-4 RTX 3060 GPUs from a compute node. Four logical CPUs and 24GB of RAM are allocated per GPU.

For example, to launch Stata with 16 logical CPUs and 32GB RAM, you can simply type:

compute-16 stata-mp

compute

For more control, you can use the compute command:

compute [options] [command]

Note:

  • Pseudo terminal mode is enabled. X11 forwarding is also enabled when a valid display is present.
  • Jobs requesting more than 32 cores are routed to the large partition, while jobs requesting A100 GPUs are routed to the a100 parititon.
  • Automatically scales CPU core count and memory with GPU type unless user specify them.
    • RTX 3060: four CPU cores and 24GB RAM per GPU
    • RTX 3090: eight CPU cores and 48GB RAM per GPU
    • A100: eight CPU cores and 160GB RAM per GPU

Options:

  • -c requests a specific number of logical CPUs. Defaults to four CPU cores.
  • --mem requests a specific amount of memory.
  • --gpus-per-task request GPUs in one of the following formats:
    • number
    • model
    • model:number
  • -t sets the maximum running time of the job in one of the following format:
    • minutes
    • minutes:seconds
    • hour:minutes:seconds
    • day-hours
    • day-hours:minutes:seconds
  • -p requests a specific partition. Defaults to ‘scrp’.
  • -q requests a specific quality of service.
  • -w Request specify nodes.
  • --account requests resource under the specified account.
  • --constraint additional restrictions. Valid values: a100 or h100.
  • -z Print the generated srun command.
  • command The command to run. Defaults to starting a bash shell.

Examples:

  1. To request 16 CPUs and 40G of memory:

    compute -c 16 --mem=40G stata-mp

  2. To request 8 CPUs, 40G of memory and one RTX 3090 GPU:

    compute -c 8 --mem=40G --gpus-per-task=rtx3090 python

  3. If you need 1TB memory, you will need to request the large memory node. Since that node is under the separate large partition, you need to specify the -p option:

    compute -c 16 --ntasks=1 --mem 1000G -p large stata-mp

  4. To access NVIDIA A100 GPUs:

    compute --gpus-per-task=a100 -c 16 --mem=160G python my_script.py

  1. For accounting reasons, H100 GPUs are also named a100 on the cluster. To specifically request H100 GPUs, use the --constraint option:

    compute --gpus-per-task=a100 --constraint=h100 -c 16 --mem=160G python my_script.py

  2. Faculty members can run jobs longer than five days by requesting the QoS c16-long. The maximum job duration is 30 days.

    compute -t 30-0 -c 16 --mem=40G stata-mp

srun

Finally, for maximum flexibility, you can use Slurm’s srun command. All the shortcuts above utilize srun underneath.

srun [options] command

Common Options:

  • --pty pseudo terminal mode. Required for any interactive software.
  • --x11 x11 forwarding. Required for software with GUI.
  • -c requests a specific number of logical CPUs.
  • --ntasks specifies the number of parallel tasks to run. Specifying ths option is generally not necessary unless you are running an MPI job or setting --hint=nomultithread. For the latter, setting --ntasks=1 prevent Slurm from erroneously starting multiple identical tasks.
  • --mem requests a specific amount of memory.
  • --gpus-per-task request GPUs in one of the following formats:
    • number
    • model
    • model:number
  • -t sets the maximum running time of the job in one of the following format:
    • minutes
    • minutes:seconds
    • hour:minutes:seconds
    • day-hours
    • day-hours:minutes:seconds
  • -p requests a specific partition.
  • -q requests a specific quality of service.
  • --account requests resource under the specified account.
  • --constraint additional restrictions. Valid values: a100 or h100.

See Slurm’s srun documentation for additional options.

Examples:

  1. To request two logical CPUs, 10GB of RAM and one NVIDIA RTX 3090 GPU for 10 hours of interactive use, type:

    srun --pty -t 600 -c 2 --mem 10G --gpus-per-task=rtx3090:1 bash

  2. If you need 1TB memory, you will need to request the large memory node. Since that node is under the separate large partition, you need to specify the -p option:

    srun --pty -c 16 --ntasks=1 --mem 1000G -p large stata-mp

  3. To access NVIDIA A100 GPUs, you need to specify the a100 partition, plus your desired number of CPU cores and memory:

    srun -p a100 --gpus-per-task=1 -c 16 --mem=160G python my_script.py

  4. For accounting reasons, H100 GPUs are also named a100 on the cluster. To specifically request H100 GPUs, use the --constraint option:

    srun -p a100 --gpus-per-task=1 --constraint=h100 -c 16 --mem=160G python my_script.py

  5. Faculty members and research postgraduate students can run jobs longer than five days by requesting the QoS c16-long. The maximum job duration is 30 days.

    srun --pty -t 30-0 -c 16 -q c16-long Rscript my_code.r

If Slurm is unable to allocate the resources you request, you will be put on a queue until resources are available or you terminate the command.

Run a Job in Batch Mode

sbatch allows you to submit a job without the need to wait for it to execute.

Run a Single Task

For example, to run Stata in batch mode, we will need the following batch script:

#!/bin/bash
#SBATCH --job-name=my_job
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=2

stata-mp -b do file_path

Then submit your job:

sbatch [file_path_to_batch_script]

Output of your job will be saved in a file named slurm-[jobid].out in the directory you called sbatch.

Run Multiple Tasks in Parallel within the Same Job

The Suboptimal Way

You can run multiple tasks in parallel by adding the & symbol behind the command for each task, followed by the wait command at the end of the script:

#!/bin/bash
#SBATCH --job-name=my_job
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=2
#SBATCH --mem=4G

stata-mp -b do file_path_1 &
stata-mp -b do file_path_2 &
wait

The batch script above will request 2 logical CPUs and 4GB of RAM, to be shared by all the tasks. Note that each task has full access to all allocated CPU cores and memory. This could result in unexpected slowdown if multiple tasks try to take advantage of all the available resources. In this example, both Stata instances will have access to the same four logical CPUs and 4GB of memory.

The Optimal Way

To allocate exclusive resource for each task, you need to run each task with a srun command inside the Batch file. You will need to specify:

  • -n, -c and --mem as special #SBATCH comments.
  • -n, --mem and --exclusive for each srun command.

Example:

#!/bin/bash
#SBATCH --job-name=my_job
#SBATCH -n 2
#SBATCH -c 4
#SBATCH --mem=8G

srun -n 1 --mem=4G --exclusive stata-mp -b do file_path_1 &
srun -n 1 --mem=4G --exclusive stata-mp -b do file_path_2 &
wait

The batch script above will requests 8 logical CPUs and 8GB of RAM. Each task will get 4 logical CPUs and 4GB of RAM for its exclusive use.

Explanation of the #SBATCH comments:

  • #SBATCH -n 2 specifies that we want to run 2 tasks in parallel.
  • #SBATCH -c 4 means 4 logical CPUS per task.
  • #SBATCH --mem=8G means 8GB of RAM across all tasks.

Explanation of the srun options:

  • -n 1 means run a single task. If you leave this option out, the task will run as many times as what you have specified in #SBATCH -n.
  • --mem=4G means 4GB of RAM for this task. The value you specify in #SBATCH --mem must be greater or equal to the sum of all the values you specify in srun --mem. If that is not true, or if you leave this option out, tasks will not necessarily run in parallel due to not having enough memory to share.
  • --exclusive specifies that each task should get exclusive access to the logical CPUs it has access to. If you leave this option out, all tasks will have access the same set of logical CPUs.

Job Status

Use scrp-queue to check the status of your job:

# Your running jobs
scrp-queue

# Completed jobs
scrp-queue -t COMPLETED

If you want to see a different set of information, you can use Slurm’s squeue instead.

Cancel a Job

You can use scancel to cancel a job before it completes. First find the ID of your running job with squeue, then:

scancel job_id

Further Information

See Slurm’s sbatch documentation for additional options.