Skip to content

Introduction to compute nodes


  • This is a short introduction in how to reach the calculation/compute/worker nodes
  • We will cover
    • queue system
    • allocation of the compute nodes
    • batch job scripts
    • interactive session
    • job efficiency

The compute nodes

When you are logged in, you are on a login node. There are two types of nodes:

Type Purpose
Login node Start jobs for worker nodes, do easy things
Worker node Do hard calculations, either from scripts of an interactive session

Bianca contains hundreds of nodes, each of which is isolated from each other and the Internet.

As Bianca is a shared resources, there are rules to use it together in fair way:

  • The login node is only for easy things, such as moving files, starting jobs or starting an interactive session
  • The worker nodes are for harder things, such as running a script or running an interactive session.

  graph TB

  Node1 -- interactive --> SubGraph2Flow
  Node1 -- sbatch --> SubGraph2Flow
  subgraph "Snowy"
  SubGraph2Flow(calculation nodes) 

        thinlinc -- usr-sensXXX + 2FA + VPN ----> SubGraph1Flow
        terminal -- usr --> Node1
        terminal -- usr-sensXXX + 2FA + VPN ----> SubGraph1Flow
        Node1 -- usr-sensXXX + 2FA + no VPN ----> SubGraph1Flow

        subgraph "Bianca"
        SubGraph1Flow(Bianca login) -- usr+passwd --> private(private cluster)
        private -- interactive --> calcB(calculation nodes)
        private -- sbatch --> calcB

        subgraph "Rackham"
        Node1[Login] -- interactive --> Node2[calculation nodes]
        Node1 -- sbatch --> Node2

Slurm, sbatch, the job queue

  • Problem: 1000 users, 300 nodes, 5000 cores
  • We need a queue:

  • Slurm is a job scheduler


  • Work interactively with your data or development
    • Run an Interactive session
    • $ interactive <flags> ...
  • If you don't need any live interaction with your workflow/analysis/simulation
    • Send your job to the slurm job batch (sbatch)
    • $ sbatch <flags> <program> or
    • $ sbatch <job script>


  • Job = what happens during booked time
  • Described in a script file or
  • Described in the command-line (priority over script)
  • The definitions of a job:
    • Slurm parameters (flags)
    • Load software modules
    • (Navigate in file system)
    • Run program(s)
    • (Collect output)
  • ... and more

Slurm parameters

  • 1 mandatory setting for jobs:
    • Which compute project? (-A)
  • 3 settings you really should set:
    • Type of queue? (-p)
      • core, node, (for short development jobs and tests: devcore, devel)
    • How many cores? (-n)
      • up to 16 for core job
    • How long at most? (-t)
  • If in doubt:
    • -p core
    • -n 1
    • -t 10-00:00:00

The queue


  • x-axis: cores, one thread per core
  • y-axis: time

  • Easiest to schedule single-threaded, short jobs

Image Image

  • Left: 4 one-core jobs can run immediately (or a 4-core wide job).

    • The jobs are too long to fit in core number 9-13.
  • Right: A 5-core job has to wait.

    • Too long to fit in cores 9-13 and too wide to fit in the last cores.

To think about

  • Where should it run? (-p node or -p core)
  • Use a whole node or just part of it?
    • 1 node = 16 cores
    • 1 hour walltime = 16 core hours = expensive
      • Waste of resources unless you have a parallel program or need all the memory, e.g. 128 GB per node
  • Default value: core

Interactive jobs

  • Most work is most effective as submitted jobs, but e.g. development needs responsiveness
  • Interactive jobs are high-priority but limited in -n and -t
  • Quickly give you a job and logs you in to the compute node
  • Require same Slurm parameters as other jobs
  • Log in to compute node
    • $ interactive ...
  • Logout with <Ctrl>-D or logout

Try interactive and run RStudio

We recommend using at least two cores for RStudio, and to get those resources, you must should start an interactive job.


Use ThinLinc

  • Start interactive session on compute node (2 cores)
  • If you already have an interactive session going on use that.
    • If you don't find it, do squeue
    • find your session, ssh to it, like: ssh sens2023598-b9
$ interactive -A sens2023598 -p core -n 2 -t 60:00
  • Once the interactive job has begun you need to load needed modules, even if you had loaded them before in the login node
  • You can check which node you are on?

    $ hostname

  • If the name before is ending with bXX you are on a compute node!

  • The login node has sens2023598-bianca
  • You can also probably see this information in your prompt, like: [bjornc@sens2023598-b9 ~]$

  • Load an RStudio module and an R_packages module (if not loading R you will have to stick with R/3.6.0) and run "rstudio" from there.

    $ ml R_packages/4.2.1

    $ ml RStudio/2022.07.1-554

  • Start rstudio, keeping terminal active (&)

$ rstudio &

  • Slow to start?
  • Depends on:

    • number of packages
    • if you save a lot of data in your RStudio workspace, to be read during start up.
  • Quit RStudio!

  • Log out from interactive session with <Ctrl>-D or logout

Job scripts (batch)

  • Write a bash script called
    • You can be in your ~ folder
  • Make first line be #!/bin/bash in the top line
  • Add also before the rest of the commands the the keywords #SBATCH
  • # will be ignored by bash and can run as an ordinary bash script
  • if running the script with the command sbatch <script> the #SBATCH lines will be interpreted as slurm flags

A simple job script template


#SBATCH -A sens2023598  # Project ID

#SBATCH -p devcore  # Asking for cores (for test jobs and as opposed to multiple nodes) 

#SBATCH -n 1  # Number of cores

#SBATCH -t 00:10:00  # Ten minutes

#SBATCH -J Template_script  # Name of the job

# go to some directory

cd /proj/sens2023598/
pwd -P

# load software modules

module load bioinfo-tools
module list

# do something

echo Hello world!  
  • Run it:

    $ sbatch

Node types

  • Bianca has three node types: thin, fat and gpu.
    • thin being the typical cluster node with 128 GB memory
    • fat nodes having 256 GB or 512 GB of memory.
      • You may specify a node with more RAM, by adding the words -C fat to your job submission line and thus making sure that you will get at least 256 GB of RAM on each node in your job.
      • If you absolutely must have more than 256 GB of RAM then you can request to get 512 GB of RAM specifically by adding the words -C mem512GB to your job submission line.
      • Please note that requesting 512 GB can not be combined with requesting GPUs.
    • You may also add -C gpu to your submission line to request a GPU node with two NVIDIA A100 40 GB.
      • Please note that all GPU nodes have 256 GB of RAM, and are thus "fat" as well. All compute nodes in Bianca has 16 CPU cores in total.
  • Please note that there are only 5 nodes with 256 GB of RAM, 2 nodes with 512 GB of RAM and 4 nodes with 2xA100 GPUs. The wait times for these node types are expected to be somewhat longer.

Some Limits

  • There is a job wall time limit of ten days (240 hours).
  • We restrict each user to at most 5000 running and waiting jobs in total.
  • Each project has a 30 days running allocation of CPU hours. We do not forbid running jobs after the allocation is over-drafted, but instead allow to submit jobs with a very low queue priority, so that you may be able to run your jobs anyway, if a sufficient number of nodes happens to be free on the system.
Summary about the Bianca Hardware
  • Intel Xeon E5-2630 v3 Huawei XH620 V3 nodes with 128, 256 or 512 GB memory
  • GPU nodes with two NVIDIA A100 40GB GPUs each.

Cores per node: 16, or on some 128

Details about the compute nodes

  • Thin nodes
    • 194 compute nodes with 16 cores and a 4TB mechanical drive or 1TB SSD as SCRATCH.
  • Fat nodes
    • 74 compute nodes, 256 GB memory
    • 14 compute nodes, 512 GB memory
    • 10 compute nodes, 256 GB memory each and equipped with 2xNVIDIA A100 (40GB) GPUs
  • Total number of CPU cores is about 5000
  • Login nodes have 2vCPU each and 16GB memory
  • Network
    • Dual 10 Gigabit Ethernet for all nodes


  • Local disk (scratch): 4 TB
  • Home storage: 32 GB at Castor
  • Project Storage: Castor

Other Slurm tools

  • squeue — quick info about jobs in queue
  • jobinfo — detailed info about jobs
  • finishedjobinfo — summary of finished jobs
  • jobstats — efficiency of booked resources
    • use eog to watch the png output files
  • bianca_combined_jobinfo

What kind of work are you doing?

  • Compute bound
    • you use mainly CPU power (more cores can help)
  • Memory bound
    • if the bottlenecks are allocating memory, copying/duplicating

Job efficiency (no type-along)

  • Check the efficiency!
  • Generate jobstats plots for your jobs

    • Firstly, find some job IDs from this month
      • $ finishedjobinfo -m <username>
    • Write down the IDs from some interesting jobs.
    • Generate the images:
      • $ jobstats -p ID1 ID2 ID3
    • Watch the images:
      • $ eog <figure-files.png>
  • The figures

    • blue line: the jobs CPU usage, 200% means 2 cores
    • horizontal dotted black line: the jobs max memory usage
    • full black line: RAM used at 5 minute intervals

Example demo

Examine the jobs run by user douglas. The relevant job numbers are the jobs with the highest jobid= numbers that have the names names and These should appear at the end of the output.

  • You can be in your ~ dir!
  • Some background info may be found in the extra material.

    finishedjobinfo -u douglas

  • We find these are job numbers 18 for and 19 for Generate jobstats plots for each job.

    jobstats -p 18 19

  • This generates two PNG image files, one for each job. These are named cluster-project-user-jobid.png. Examine them both using an image viewer.

    eog bianca-sens2023598-douglas-18.png bianca-sens2023598-douglas-19.png



The judgement

This job has booked many more cores and memory (RAM) than it needs.


The judgement

This job needs more memory (RAM).

Discovering job resource usage with jobstats

Extra exercise (if time allows)

Submit a Slurm job
  • Make a batch job to run the demo "Hands on: Processing a BAM file to a VCF using GATK, and annotating the variants with snpEff". Ask for 2 cores for 1h.
    • You can copy the file in /proj/sens2023598/workshop/slurm to your home folder and make the necessary changes.
  • edit a file using you preferred editor, named, for example, with the content
  • alternatively copy the /proj/sens2023598/workshop/slurm/ file and modify it cd ~ cp /proj/sens2023598/workshop/slurm/ .
  • edit and add the SBATCH commands
#SBATCH -A sens2023598
#SBATCH -J workflow
#SBATCH -t 01:00:00
#SBATCH -p core
#SBATCH -n 2

cd ~
mkdir -p myworkflow
cd myworkflow

module load bioinfo-tools

# load samtools
module load samtools/1.17

# copy and example BAM file
cp -a /proj/sens2023598/workshop/data/ERR1252289.subset.bam .

# index the BAM file
samtools index ERR1252289.subset.bam

# load the GATK module
module load GATK/

# make symbolic links to the hg38 genomes
ln -s /sw/data/iGenomes/Homo_sapiens/UCSC/hg38/Sequence/WholeGenomeFasta/genome.* .

# create a VCF containing inferred variants
gatk HaplotypeCaller --reference genome.fa --input ERR1252289.subset.bam --intervals chr1:100300000-100800000 --output ERR1252289.subset.vcf

# use snpEFF to annotate variants
module load snpEff/5.1
java -jar $SNPEFF_ROOT/snpEff.jar eff hg38 ERR1252289.subset.vcf > ERR1252289.subset.snpEff.vcf

# compress the annotated VCF and index it
bgzip ERR1252289.subset.snpEff.vcf
tabix -p vcf ERR1252289.subset.snpEff.vcf.gz
  • make the job script executable

    $ chmod a+x

  • submit the job

    $ sbatch


  • You are always in the login node unless you:
    • start an interactive session to do development or hands-on work
    • start a batch job to run jobs not needing any manual input
  • Slurm is a job scheduler
    • add flags to describe your job.
  • There is a job wall time limit of ten days (240 hours).