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.
- 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)
-
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.
-
This project uses
renv
to manage R package dependencies. After opening the .Rproj file in RStudio, runrenv::restore()
to install all needed packages. -
SSH into HiperGator. Launch R and install the
clustermq
package. -
On HiperGator, edit your
~/.Rprofile
(e.g. withnano ~/.Rprofile
) to include:options( clustermq.scheduler = "slurm", clustermq.template = "~/slurm_clustermq.tmpl" )
-
Edit the
slurm_clustermq.tmpl
file if needed (but don’t touch the wildcards in double curly braces) and then copy it to HiperGator withscp slurm_clustermq.tmpl username@hpg.rc.ufl.edu:slurm_clustermq.tmpl
. More detailed instructions on setting upclustermq
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). -
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 therun.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 withtar_watch()
or by runningtar_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;
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.
Problems with HiperGator:
Problems with clustermq
:
Problems with targets
(i.e. problem still exists with tar_make()
instead of tar_make_clustermq()
:
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.