How to run an interactive job on ARC: Difference between revisions

From RCSWiki
Jump to navigation Jump to search
No edit summary
Tag: Manual revert
No edit summary
Line 1: Line 1:
= Background =
Submitting and running jobs is the only way to run computations on ARC. Normally, a user creates a '''job script''' and submit it to SLURM on ARC using the <code>sbatch</code> command to run a computation. This type of jobs are non-interactive batch jobs. In some circumstances, this type of workflow may be difficult or incompatible with workflows that require interactive inputs by the user such as when you need to apply data transformation, result collection and review from large computations, or to test new methods or software. For this kind of work, the best approach is to request an '''interactive job''' from ARC's.
Submitting and running jobs is the only way to run computations on ARC. Normally, a user creates a '''job script''' and submit it to SLURM on ARC using the <code>sbatch</code> command to run a computation. This type of jobs are non-interactive batch jobs. In some circumstances, this type of workflow may be difficult or incompatible with workflows that require interactive inputs by the user such as when you need to apply data transformation, result collection and review from large computations, or to test new methods or software. For this kind of work, the best approach is to request an '''interactive job''' from ARC's.


== Why not the login node? ==
= Why not the login node? ==
The login node is designed for users to interactively use ARC. However, this usage is intended for only '''management work''', such as:
The login node is designed for users to interactively use ARC. However, this usage is intended for only '''management work''', such as:
* data management,
* data management,
Line 9: Line 10:
The '''login node''' could also be used to '''software development''', if you write scripts, or if you work on some new code for your research. The '''login node''', however, '''cannot be used''' to run anything that is CPU heavy and relates to '''research''', that is '''not management'''. See our[[General Cluster Guidelines and Policies#Login Node | guidelines and policies]].
The '''login node''' could also be used to '''software development''', if you write scripts, or if you work on some new code for your research. The '''login node''', however, '''cannot be used''' to run anything that is CPU heavy and relates to '''research''', that is '''not management'''. See our[[General Cluster Guidelines and Policies#Login Node | guidelines and policies]].


== Submitting an interactive job ==
= Submitting an interactive job ==


To '''request an interactive job''' on ARC via SLURM, you have to use the <code>salloc</code> command and  
To '''request an interactive job''' on ARC via SLURM, you have to use the <code>salloc</code> command and  
Line 33: Line 34:
For '''GPU partitions and jobs''' please see its own [[How to request an interactive GPU on ARC | How To article]].
For '''GPU partitions and jobs''' please see its own [[How to request an interactive GPU on ARC | How To article]].


== Downsides to interactive jobs ==
= Downsides to interactive jobs ==
However, one has to remember that there are some '''negatives''' of using '''interactive jobs''':
However, one has to remember that there are some '''negatives''' of using '''interactive jobs''':
* '''Resource utilization''' is typically '''quite low''' when compared with non-interactive jobs.  
* '''Resource utilization''' is typically '''quite low''' when compared with non-interactive jobs.  

Revision as of 18:51, 31 January 2024

Background

Submitting and running jobs is the only way to run computations on ARC. Normally, a user creates a job script and submit it to SLURM on ARC using the sbatch command to run a computation. This type of jobs are non-interactive batch jobs. In some circumstances, this type of workflow may be difficult or incompatible with workflows that require interactive inputs by the user such as when you need to apply data transformation, result collection and review from large computations, or to test new methods or software. For this kind of work, the best approach is to request an interactive job from ARC's.

Why not the login node? =

The login node is designed for users to interactively use ARC. However, this usage is intended for only management work, such as:

  • data management,
  • software management, and
  • job management.

The login node could also be used to software development, if you write scripts, or if you work on some new code for your research. The login node, however, cannot be used to run anything that is CPU heavy and relates to research, that is not management. See our guidelines and policies.

Submitting an interactive job =

To request an interactive job on ARC via SLURM, you have to use the salloc command and to specify resources that your interactive work will need, similar to the resource request in a normal job script. After that, SLURM will try to allocate resources for your interactive job, and then it will transfer your command line session from ARC's login node to a new command line session on the allocated compute node.

In this interactive job you can run any command on the command line without interrupting any other user's work. The resources you are using are allocated to your only and there is nobody else who can use them.

  • On ARC any interactive job should be limited to 5 hours or less. Please do not request more than 5 hours for your interactive jobs.

For example:

$ salloc -N1 -n1 -c8 --mem=16gb -t 1:00:00 

This command requests 8 CPUs and 16 GB of RAM for 1 hour on one of the compute nodes in the default partitions.

For a partition that is not a default partition, such as a bigmem partition, or the GPU partitions. You have to specify it on the command line:

$ salloc -N1 -n1 -c20 --mem=600gb -t 1:00:00 -p bigmem

This command requests 20 CPUs and 600 GB of RAM for 1 hour on a node from the bigmem partition.

For GPU partitions and jobs please see its own How To article.

Downsides to interactive jobs =

However, one has to remember that there are some negatives of using interactive jobs:

  • Resource utilization is typically quite low when compared with non-interactive jobs.
The resources are not used when you are reading the outputs, type new commands, or think about what to do next. This is inefficient
  • There is no amount of work that have to be done, so that the resources can be freed for other users.
The only way an interactive job can end is if it runs out of time, or you manually end it.
  • Typically, you can only interact with one job at a time.
In contrast, you can run multiple non-interactive jobs when running normal batch jobs.
  • If the resources you want for your interactive work are not available, you may be waiting for a long time before your interactive job starts.

Because of these reasons, interactive jobs on ARC is limited to up to 5 hours of run time.

Please remember that interactive jobs is not intended to be the primary way of using the ARC cluster.