hipergator-targets-ssh

This is a minimal example of a targets workflow that can run on HiperGator HPC using the clustermq back-end for paralellization. This is intended to be run locally in RStudio. For an alternative workflow that is run on HiperGator via the command line, see this repo.

Prerequisites:

• Some familiarity with HiperGator
• Ability to SSH into HiperGator without a password. Set this up with ssh-keygen and by creating a SSH config file locally (see details here: https://help.rc.ufl.edu/doc/Authentication_With_MFA)

To Use:

1. Click the green “Use this template” button to make a copy into your own GitHub account or use the “Code” button to download a zip file of this repo.

2. This project uses renv to manage R package dependencies. After opening the .Rproj file in RStudio, run renv::restore() to install all needed packages.

3. SSH into HiperGator. Launch R and install the clustermq package.

4. On HiperGator, edit your ~/.Rprofile (e.g. with nano ~/.Rprofile) to include:

   options(
     clustermq.scheduler = "slurm",
     clustermq.template = "~/slurm_clustermq.tmpl"
   )

5. Edit the slurm_clustermq.tmpl file if needed (but don’t touch the wildcards in double curly braces) and then copy it to HiperGator with scp slurm_clustermq.tmpl [email protected]:slurm_clustermq.tmpl. More detailed instructions on setting up clustermq

   NOTE: While you’re logged into HiperGator, you should also install any R packages your project uses. Launch R and install them manually (sorry, I don’t know an easy way to use the local renv lock file to automatically install packages on HiperGator).

6. Run targets workflow:

   To run this example workflow, you can either run targets::tar_make_clustermq() in the console, or use the “Jobs” feature of RStudio to run the run.R script as a local job—this keeps the console from being tied up waiting for the jobs to run on the cluster. You can watch the progress of the pipeline with tar_watch() or by running tar_visnetwork() to see a graph like this:

graph LR
  subgraph Legend
    x7420bd9270f8d27d([""Up to date""]):::uptodate --- x5b3426b4c7fa7dbc([""Started""]):::started
    x5b3426b4c7fa7dbc([""Started""]):::started --- xbf4603d6c2c2ad6b([""Stem""]):::none
    xbf4603d6c2c2ad6b([""Stem""]):::none --- xf0bce276fe2b9d3e>""Function""]:::none
  end
  subgraph Graph
    x89d7607898dce999>"make_vects"]:::uptodate --> xc7a97a0c3f5c7239(["many_vects"]):::uptodate
    xc7a97a0c3f5c7239(["many_vects"]):::uptodate --> x7093befdb6050f4e(["means"]):::uptodate
    xc7a97a0c3f5c7239(["many_vects"]):::uptodate --> x5746aa97a820c5ac(["sds"]):::uptodate
    x5746aa97a820c5ac(["sds"]):::uptodate --> xc75015002057228e(["sd_means"]):::uptodate
    x7093befdb6050f4e(["means"]):::uptodate --> xe00cc1033f4aed9a(["means_mean"]):::uptodate
    x6e52cb0f1668cc22(["readme"]):::started --> x6e52cb0f1668cc22(["readme"]):::started
  end
  classDef uptodate stroke:#000000,color:#ffffff,fill:#354823;
  classDef started stroke:#000000,color:#000000,fill:#DC863B;
  classDef none stroke:#000000,color:#000000,fill:#94a4ac;
  linkStyle 0 stroke-width:0px;
  linkStyle 1 stroke-width:0px;
  linkStyle 2 stroke-width:0px;
  linkStyle 8 stroke-width:0px;

Loading

Notes

If you only want certain targets to run on Hipergator, you can control this with the deploy argument to tar_options_set() and tar_target(). Targets with deploy = "main" will run locally and targets with deploy = "worker" will run on HiperGator. Set the default behavior with tar_options_set() inside of _targets.R and then adjust individual targets as needed. See _targets.R for an example and see the targets manual for more detail. Note that you will still have to wait for remote targets to finish running for the pipeline to finish, so if any targets take a very long time to run, this alternative approach might be better.

Currently (as of May 2022) there is a bug with R version 4.1+ on HiperGator that affects this workflow. To get around this, I’ve edited the ssh_clustermq.tmpl and slurm_clustermq.tmpl files to load R version 4.0.

Troubleshooting:

Problems with HiperGator:

• https://help.rc.ufl.edu/doc/UFRC_Help_and_Documentation

Problems with clustermq:

• https://mschubert.github.io/clustermq/articles/userguide.html#ssh

Problems with targets (i.e. problem still exists with tar_make() instead of tar_make_clustermq():

• https://docs.ropensci.org/targets/index.html#help

How it works:

Using tar_make_clustermq() sends necessary data to HiperGator via SSH and spawns worker jobs using the slurm_clustermq.tmpl file as a template for the SLURM submission scripts for each worker.

The parallelization happens at the level of targets. In this example, a list of numeric vectors is stored as many_vects. Then, independently, means and standard deviations are calculated for each vector in the list. These two targets (the means and the sd’s) should be able to run on separate workers in parallel if things are set up correctly. Parallelizing code within a target (e.g. a function that does parallel computation) will require more setup.