Skip to content

Nextflow

Nextflow is a data-centric workflow management tool which facilitates complex and reproducible scientific computational workloads.

A workflow pattern including the following characteristics might be successfully realized by using Nextflow at NERSC:

  • It contains significant complexity in the form of different applications, data formats, repetition, conditional branching, and/or dependencies between all of these
  • The maximum scaling you project will require is low
  • Fast workflow turnaround is not a priority for you
  • You already possess a system implemented with Nextflow

Strengths of Nextflow:

  • Simple installation and basic execution
  • No persistent database needed
  • Powerful ability to express complex tasks and relationships between tasks

Disadvantages of Nextflow:

  • Easy to unintentionally create a configuration which degrades scheduler performance for all NERSC users
  • Weak integration with Slurm that is unable to adapt to NERSC scheduler policies. This can lead to long queue waiting times for tasks and excessive turnaround for workflows as a whole
  • Poor ability to manage limitations of NERSC resources when running workflows which will need to run tasks over several days or longer
  • Nextflow makes a lot of duplicate and intermediate file copies as it operates; processing large amounts of data can easily exhaust storage quotas
  • Nextflow uses parallel file systems to synchronize tasks, which is a frequent source of undesired behavior

How to use Nextflow at NERSC

Start with the Nextflow documentation Quick Start. The system Java installation provided by default on Perlmutter is sufficient to run Nextflow.

This documentation only discusses specifics of running Nextflow at NERSC; for guidance expressing your workflows with Nextflow scripts, see the official Nextflow documentation.

For the most basic proof of concept demonstration, first choose a location on $SCRATCH and install with the command curl -s https://get.nextflow.io | bash To run the pre-installed demonstration use this command: ./nextflow run hello.

Basic Example Output
elvis@login10:/nextflow> ./nextflow run hello
N E X T F L O W  ~  version 20.01.0
Launching `nextflow-io/hello` [goofy_ardinghelli] - revision: 1d43afc0ec [master]
WARN: The use of `echo` method is deprecated
executor >  local (4)
[ca/000898] process > sayHello [100%] 4 of 4 ✔
Ciao world!

Hola world!

Hello world!

Bonjour world!

elvis@login10:/nextflow> 

We can directly run this same test by creating the following file and then using the command ./nextflow run local_hello.nf This displays the same result as the first test (besides changes in unique run identifiers and time stamps).

local_hello.nf
#!/usr/bin/env nextflow                                                     
echo true                                                                   

cheers = Channel.from 'Bonjour', 'Ciao', 'Hello', 'Hola'                    

process sayHello {                                                          
    input:                                                                    
    val x from cheers                                                       
    script:                                                                   
    """                                                                     
    echo '$x world!'                                                        
    """                                                                     
}   

Warning

Nextflow has a bad habit of using hidden files excessively for things like batch submission scripts and log files. Keep this in mind when debugging a workflow.

Now that the basic operation is confirmed, we will describe how to run the Nextflow management and compute processes on more appropriate resources than a login node.

Wrapped Nextflow submission to Slurm

The easiest method is placing the ./nextflow ... command into a batch script and submitting it to Slurm with sbatch. The manager process will run on the allocated compute node, and all tasks are configured to use the local executor; it's even possible to use srun in your processes to run tasks which include MPI applications.

The major benefit of this method, besides simplicity, is only the initial submission waits in a Slurm queue; it is a good pattern for a workflow which includes a very large number of small tasks. One should not combine this method with the Nextflow 'Slurm' executor because the job running the Nextflow manager is likely to end before the requested task is finished waiting in a queue.

There are two significant caveats to running the Nextflow workflow process directly inside a Slurm job allocation:

  • The Nextflow working directory must be placed on the Perlmutter Lustre $SCRATCH filesystem. Nextflow uses a file locking feature not available on any of the other NERSC filesystems.
  • The workflow cannot run longer than the maximum wall time available to a single job in the Slurm QOS being used. This can be partially mitigated by using multiple Slurm submissions in series and passing the -resume flag to Nextflow, but only progress for completely finished tasks will be preserved from one submission to the next.

Nextflow Submits Tasks as Slurm Jobs

Nextflow configuration is used to instruct the manager process to submit its tasks to Slurm instead of running them on the local host. Place the following file in your Nextflow working directory:

nextflow.config

process {
  executor='slurm'
  queueSize = 15
  pollInterval = '5 min'
  dumpInterval = '6 min'
  queueStatInterval = '5 min'
  exitReadTimeout = '13 min'
  killBatchSize = 30
  submitRateLimit = '20 min'
  clusterOptions = '-q debug -t 00:30:00 -C cpu'
}    

Left to default configuration, a running Nextflow workflow manager process can generate a disruptive amount of communication requests to Slurm; most of these configuration file entries reduce the frequency of those requests.

Inside the individual process definitions in your scripts, you will likely wish to override the clusterOptions variable to request specific resources appropriate for that task. This can be done by adding something in the pattern of clusterOptions='-q regular -t 05:30:00 -C cpu' to the top of your task process blocks.

The choice of machine to run the Nextflow manager on is also important. NERSC recommends using the Perlmutter workflow queue for this purpose. The workflow queue is intended for users to submit lightweight tasks which coordinate computational work. Users should be prepared to infrequently check on the Nextflow manager process in their workflow queue job there and -resume a long campaign Nextflow process that has been inturrupted.