Using Slurm on Pelle

This page describes how to use Slurm on Pelle.

What is Slurm?

See the general page about Slurm

What is Pelle?

See the general page about Pelle

See Slurm troubleshooting how to fix Slurm errors.

What about other clusters?

See the general page about Slurm at UPPMAX

Newer Slurm

• Slurm on Pelle have been upgraded to version 25.05.

• Several UPPMAX-specific Slurm changes from previous clusters have been removed, to make the config use more Slurm defaults. This makes the system easier to maintain and will behave more similar to clusters at other sites. Unfortunately this means that some extra changes to job scripts can be needed when moving from Rackham/Snowy.

Warning

• The max time limit for jobs is 10 days.
  • GPU jobs has a time limit of 2 days.

Info

• We recommend to replace the Slurm option -n (recommended in our documentation before), when allocating several cores, with -c (CPUs-per-task)
• This prevents the allocation to be spread among multiple nodes.
• If you, however, are using MPI, you should define the number of tasks with -n (number of tasks (in total)).
• The reason why -c often can/could be used interchangeably with -n is the default value of one core per task.

Quick start

Quick start for starting jobs on Pelle

Ways to start jobs

• Interactive

  • work interactively, starting programs and view data etcetera on a compute node
  • interactive -A uppmax202X-Y-ZZZ -c 2 -t 3:0:0
  • read more

• Batch system

  • Allocate much resources or long wall times and let job run by its own without interaction with you
  • batch <submit script>
  • read more
  Simple batch script

  #!/bin/bash
  #SBATCH -A uppmax2023-2-25
  #SBATCH -c 1   # number of threads
  #SBATCH -t 0:1:0 # 1 minute
  echo "Hello"
  

Wall times

• Specify the maximum time needed before slurm breaks the job.
• -t 10:0 10 minutes
• -t 10:0:0 10 hours
• -t 5-10:0:0 5 days and 10 hours

Resources

• What kind of computations do you need?
• Core jobs (Default):
  • -c <number of cores>
  • read more
• Node jobs:
  • -N <number of nodes>
  • read more
• Jobs using MPI:
  • -n <number of tasks>
• Large memory jobs:
  • -p fat
  • read more
• GPU (NVIDIA) jobs
  • T4 (36 nodes with 1 GPU each): interactive -A staff -t 1:0:0 -p haswell --gpus=t4
  • Faster L40s (4 nodes with 10 GPUs each): interactive -A staff -t 1:0:0 -p gpu --gpus=l40s:1
  • Superfast H100 (2 nodes with 2 GPUs each): interactive -A staff -t 1:0:0 -p gpu --gpus=h100:1
  • read more
• Intel Haswell nodes (with 16 cores per node)
  • -p haswell ...
  • read more

sbatch (and interactive) on Pelle

sbatch (and interactive) work the same as on the other clusters, the only difference is that some flags/options may be different, like partition name, see below.

Want to start an interactive session?

See how to start an interactive session on Pelle

Here it is shown how to submit a job with:

• Command-line Slurm parameters
• Slurm parameters in the script

Partitions on Pelle

Partition flag is either --partition or -p

Partition name	Description
pelle	(Default) Use one or more CPU cores
fat	Use a fat node with 2 or 3 TB memory, see below
gpu	GPU nodes, 2 types see below
haswell	Old Snowy/Irma nodes, half with GPUs (T4)

The pelle partition

The pelle partition is default so you can omit specifying -p or --partition.

This is the ordinary CPU nodes of Pelle.

Warning

• Time limit is 10 days on the CPU nodes.
• You may, if really needed, ask for more through the support support@uppmax.uu.se.

Info

The compute node CPUs have Simultaneous multithreading (SMT) enabled. Each CPU core runs two Threads. In Slurm the Threads are referred to as CPUs. Learn more here about SMT

No of nodes	CPUs	Cores Threads	Memory	Scratch	GPUs
80	AMD EPYC 9454P (Zen4) 2.75 GHz	48 96	768 GiB	1.7 TB	N/A

Much more cores per node compared to Rackham

• You can now have 96 parallel processes per node!
• Even more important that you not by mistake allocate a full node when needing just a part of it.
• A full node is 768 GB, compared to 128 GB on Rackham. That means less need for a "fat" partition allocation.

Examples with core jobs

Here is the minimal use for one core:

sbatch -A [project_code] [script_filename]

For example:

sbatch -A staff my_script.sh

To specify multiple cores, use --cpu-per-tasks (or -c) like this:

sbatch -A [project_code] -c [number_of_cores] [script_filename]

For example:

sbatch -A staff -c 2 my_script.sh

Here, two cores are used.

What is the relation between ntasks and number of cores?

Agreed, the flag ntasks only indicates the number of tasks. However, by default, the number of tasks per core is set to one. One can make this link explicit by using:

sbatch -A [project_code] --partition --ntasks [number_of_cores] --ntasks-per-core 1 [script_filename]

This is especially important if you might adjust core usage of the job to be something less than a full node.

• One task, using two threads: --ntasks=1 --cpus-per-task=2
• Two tasks, using one thread each: --ntasks=2 --cpus-per-task=1

Examples with node jobs

• On Rackham you have used -p node or -p core to specify node/core jobs.

• This is not used on Pelle. Instead Slurm's standard options is used to specify the job requirements.

• Example to request 2 full nodes: --nodes=2 --exclusive

Job memory specification

• Currently you do not have to request additional CPUs to get additional memory.
• You can use all Slurm options
  • --mem
  • --mem-per-cpu

The fat partition

With the fat partition you reach compute nodes with more memory.

Warning

• Time limit is 10 days on the fat nodes.
• You may, if really needed, ask for more through the support support@uppmax.uu.se.

No of nodes	CPUs	Cores Threads	Memory	Scratch	GPUs
1	AMD EPYC 9454P (Zen4) 2.75 GHz	48 96	2 TiB	6.9 TB	N/A
1	AMD EPYC 9454P (Zen4) 2.75 GHz	48 96	3 TiB	6.9 TB	N/A

Note

Jobs on these nodes always allocate the entire node with all cores.

• To allocate 2 TB: -p fat -C 2TB

  • Example: interactive -A staff -t 1:0:0 -p fat -C 2TB

• To allocate 3 TB: -p fat -C 3TB

  • Example: interactive -A staff -t 1:0:0 -p fat -C 3TB

The gpu partition

With the gpu partition you reach the nodes with GPUs.

Warning

• Time limit is 2 days.
• You may, if really needed, ask for more through the support support@uppmax.uu.se.
• The T4 GPUs in the Haswell partition can be booked for 10 days.

No of nodes	CPUs	Cores Threads	Memory	Scratch	GPUs
4	2x AMD EPYC 9124 (Zen4) 3 GHz	2 x 16 2 x 32	384 GiB	6.9 TB	10x NVIDIA L40s
2	2x AMD EPYC 9124 (Zen4) 3 GHz	2 x 16 2 x 32	384 GiB	6.9 TB	2x NVIDIA H100

To avoid long waiting times we primarily recommend to allocate the default L40s and just one of them. Or, if you can use one of the older T4 GPUs, those are in greater supply.

• To allocate L40s: -p gpu --gpus=l40s:<number of GPUs>

  • Example with 1 GPU: interactive -A staff -t 1:0:0 -p gpu --gpus=l40s:1
  • Example with the maximum 10 GPUs: interactive -A staff -t 1:0:0 -p gpu --gpus=l40s:10

• To allocate H100: -p gpu --gpus=h100:<number of GPUs>

  • Example with 1 GPU: interactive -A staff -t 1:0:0 -p gpu --gpus=h100:1
  • Example with both GPUs: interactive -A staff -t 1:0:0 -p gpu --gpus=h100:2
  • There are just 2 nodes with 2 each of these, so queues may become long if demand is high.

• Currently you do not have to request additional CPUs to get additional memory.

• You can use all Slurm options
  • --mem
  • --mem-per-cpu
  • --mem-per-gpu to specify memory requirements.

The haswell partition

This partition holds older hardware that may be in less demand.

No of nodes	CPUs	Cores = Threads	Memory	Scratch	GPUs
36	2x Xeon E5-2630 v3 2.4 GHz (Haswell)	2 x 16	256 GiB	1.8 TB	N/A
34	2x Xeon E5-2630 v3 2.4 GHz (Haswell)	2 x 16	256 GiB	1.8 TB	NVIDIA T4

• Example with 8 cores (1 socket): interactive -A staff -t 1:0:0 -p haswell -c 8

• Example with a T4 GPU: interactive -A staff -t 1:0:0 -p haswell --gpus=t4

  • If you ask for several T4 GPUs, you do get them, but they are spread across different nodes with just one GPU per node.

  sbatch -A staff -t 1:0:0 -p haswell --gpus=t4:16
  

sbatch a script with command-line Slurm parameters

The minimal command to use sbatch with command-line Slurm parameters is:

sbatch -A [project_code] [script_filename]

where [project_code] is the project code, and [script_filename] the name of a bash script, for example:

sbatch -A uppmax2023-2-25 my_script.sh

Forgot your Rackham project?

One can go to the SUPR NAISS pages to see one's projects,

Example of the Rackham project called 'UPPMAX 2023/2-25'

On the SUPR NAISS pages, projects are called 'UPPMAX [year]/[month]-[day]', for example, 'UPPMAX 2023/2-25'. The UPPMAX project name, as to be used on Rackham, has a slightly different name: the account name to use on Rackham is uppmax[year]-[month]-[day], for example, uppmax2023-2-25

What is in the script file?

The script file my_script.sh is a minimal example script. Such a minimal example script could be:

#!/bin/bash
echo "Hello"

Again, what is shown here is a minimal use of sbatch.

sbatch a script with Slurm parameters in script

The minimal command to use sbatch with Slurm parameters in the script:

sbatch [script_filename]

where [script_filename] the name of a bash script, for example:

sbatch my_script.sh

The script must contain at least the following lines:

#SBATCH -A [project_code]

where [project_code] is the project code, for example:

#SBATCH -A uppmax2023-2-25

Forgot your Pelle project?

One can go to the SUPR NAISS pages to see one's projects,

Example of the Rackham project called 'UPPMAX 2023/2-25'

On the SUPR NAISS pages, projects are called 'UPPMAX [year]/[month]-[day]', for example, 'UPPMAX 2023/2-25'. The UPPMAX project name, as to be used on Rackham, has a slightly different name: the account name to use on Rackham is uppmax[year]-[month]-[day], for example, uppmax2023-2-25

A full example script would be:

#!/bin/bash
#SBATCH -A uppmax2023-2-25
echo "Hello"

SMT

The compute node CPUs have Simultaneous multithreading (SMT) enabled. Each CPU core runs two Threads. In Slurm the Threads are referred to as CPUs.

Different jobs are never allocated to the same CPU core. The smallest possible job always gets one Core with two Threads (CPUs).

Jobs requesting multiple tasks or cpus gets threads by default.

Some examples:

• --ntasks=2 - one core, two threads
• --ntasks=1 --cpus-per-task=4 - two cores, four threads
• --ntasks=2 --cpus-per-task=3 - three cores, six threads.

One thread per core to avoid SMT

If you suspect SMT degrades the performance of your jobs, you can you can specify --threads-per-core=1 in your job.

Same examples as before but with --threads-per-core=1:

• --ntasks=2 --threads-per-core=1 - two cores, (4 threads, 2 used)
• --ntasks=1 --cpus-per-task=4 --threads-per-core=1 - 4 cores (8 threads, 4 unused)
• --ntasks=2 --cpus-per-task=3 --threads-per-core=1 - 6 cores (12 threads, 6 unused)

When doing this you should launch your tasks using srun to ensure your processes gets pinned to the correct CPUs (threads), one per core.

Again, what is shown here is a minimal use of sbatch.

Per-job scratch directories

Each job get job-specific private temporary directories on the compute nodes. This means that directories like /tmp, /scratch, /dev/shm start empty for each job. This is implemented using Slurm's job_container/tmpfs plugin.

To get the path for node local scratch in job scripts you should still use the variable $TMPDIR or the equivalent $SNIC_PATH.

More allocated memory can be used (/dev/shm size)

The size of /dev/shm have been increased so all allocated memory can be used, up to max 99% of a compute nodes total memory. On previous clusters the limit was 50% of total memory.