R¶

R runs from the command line and in jobs through the bio/R module. Packages install into your own library, and R has to be kept to the CPUs a job requested. RStudio, on the RStudio server or on a compute node, is under RStudio.

Load R¶

$ module load bio/R
$ R --version

bio/R loads R 4.4.0. bio/R/4.4.1 selects another version, and module -t avail 2>&1 | grep 'bio/R/' lists them. tools/R and bioinformatics/R are aliases of bio/R. For a newer R than the modules provide, install it in a conda environment (see Python and conda) with conda create -n r-env -c conda-forge r-base.

Install packages¶

Start R on a login node or in an interactive session and install as usual:
```
install.packages("vegan")
```
R reports that the system library is not writable and offers to create a personal library under your home directory. Accept. Every package you install from then on goes there, and every job on every node sees it.
For a package that compiles C or Fortran code, install from a login node or an interactive session. Those nodes have the system development libraries such packages need and the compute nodes do not, so an install started inside a batch job fails at the compile step.

Use the requested CPUs¶

Give R functions that run in parallel the number of workers from NSLOTS, so the job uses the slots it requested however it was submitted:

numcores <- as.integer(Sys.getenv("NSLOTS"))
cl <- makeCluster(numcores, type = "FORK")

Start R workers with the FORK type

Use type = "FORK" on Hydra. Without it each worker is a separate R process with its own copy of the data, so the job uses far more resources than expected and may be killed.

R's linear-algebra libraries start one thread per CPU on the node unless told otherwise, which oversubscribes the node. Set OMP_NUM_THREADS in the job file before starting R:

Job	Set
serial	`export OMP_NUM_THREADS=1`
`-pe mthread N`, threads only	`export OMP_NUM_THREADS=$NSLOTS`
`-pe mthread N`, all slots used by `makeCluster()`	`export OMP_NUM_THREADS=1`
`makeCluster()` with K workers plus threads	`export OMP_NUM_THREADS=$((NSLOTS / K))`

model.job

#$ -S /bin/sh
#$ -N model -cwd -j y -o model.log
#$ -pe mthread 8
#$ -l mres=16G,h_data=2G,h_vmem=2G
#
echo + `date` job $JOB_NAME started in $QUEUE with jobID=$JOB_ID on $HOSTNAME
module load bio/R
export OMP_NUM_THREADS=1
Rscript model.R
echo = `date` job $JOB_NAME done

R in a conda environment¶

Several jobs activating the same R conda environment at once can hang. The environment's activate.d/activate-r-base.sh runs R CMD javareconf, and concurrent runs overwrite each other's files. Either edit that script so the line reads

R CMD javareconf > /dev/null 2>&1 || true

or put the environment's bin directory on PATH in the job file instead of activating it, which works for most packages but skips the other variables activation sets.

For help, email [email protected] with the full error message and whether you were using the RStudio server, RStudio on a node, or R from the command line.

R¶

Load R¶

Install packages¶

Use the requested CPUs¶

R in a conda environment¶

Further reading¶