This repo is not currently maintained, will likely not work "as is" for its intendend purpose, and only left available for reference
The mainter has no current affiliation with Princeton University.
Setup and Example scripts (mostly) for R-based HPC at Princeton University
Contact Jonathan Olmsted (jpolmsted@gmail.com) with questions or issues.
This project comprises a set of multiple example scripts and a setup script to help prepare your HPC environment for R-based HPC. These scripts work out of the box on Tukey, Della, and Adroit. There are some additional examples of using Python and Matlab, but these are not the focus. More information on these scripts is below.
These scripts have been tested on:
- Tukey
- Della
- Adroit
If you are using one of these systems and something isn't working, please email for support.
You can create a full copy of this git repository by cloning it. To do this, run the following at the shell prompt from one of these systems:
git clone https://github.com/olmjo/tigress-scripts
On the Princeton TIGRESS systems, git
will be installed so this "just
works". The same git
command will work on your local machine if you have git
installed.
Start by navigating into the repository directory:
cd tigress-scripts
For all of the example scripts to work, the shell environment needs to be set up
correctly and particular R packages need to be installed. The helper setup
script is located in ./setup/setup.sh
. To set up your environment, simply run
the script:
cd setup
./setup.sh
Warning: If you have never used R on these machines before, you start first start R, install any package at all, and then quit R before proceeding. Without having installed any R package in this way, the subsequent setup scripts won't work.
You will be prompted by several questions like
@ Do you want to set up your account to use an updated compiler version?
Note 1: may be necessary for some R packages
Note 2: only needs to be done once per system
Note 3: may have no effect on certain machines
[y/n]
@ Do you need to DL Rmpi?
[y/n]
@ Do you need to add OpenMPI support?
Note: needs to only be added once
[y/n]
@ Do you need to install Rmpi?
[y/n]
@ Do you need to install misc. HPC R packages?
[y/n]
If you've never run this before, answer "y" to each question.
To test the openmpi setup run the following:
cd ../test/
./test_mpi.sh
You should see output like:
Process 1 on tukey out of 3
Process 2 on tukey out of 3
Process 0 on tukey out of 3
There are a series of example scripts in this project. Each example can be used
to submit a perfectly valid job on the above listed systems. They are located in
the ./examples/
subdirectory. The shell scripts with a .pbs
suffix are PBS
scripts for the Torque resource manager. The shell scripts with a .slurm
suffix are SLURM scripts. PBS scripts are not actively maintained since the
systems listed above all run SLURM. And, there is no promise that the SLURM and
PBS versions will always be "in sync".
In each example description, below, versions for either PBS, SLURM, or both are indicated.
PBS, SLURM
This is a bare bones example. It requests 1 task with 1 processor. It allows the scheduler to kill the job after 10 minutes. The script simply generates 1,000 random numbers using the Rscript interface to R.
To run under Torque:
cd ./examples/ex0/
qsub ex0.pbs
To run under SLURM:
cd ./examples/ex0/
sbatch ex0.slurm
PBS, SLURM
This script represents a reasonable starting point for simple jobs. It is more explicit about how the job should be managed than Example 0. It still requests 1 task with 1 processor. It requests only 10 minutes of time. It uses a custom name in the queue and has both the error log and the output log merged into one file which begins with log.* and has a suffix determined by the job ID. It requests emails when it begins, ends, and aborts (the email address can be specified manually, but works by default on TIGRESS systems).
Every line beginning with # is just a scheduler directive. The remainder comprises an actual shell script. The script is verbose about where it is, when it starts, and what resources were given to it by the scheduler.
The script ultimately generates 1,000 random numbers using the Rscript interface to R.
To run under Torque:
cd ./examples/ex1/
qsub ex1.pbs
To run under SLURM:
cd ./examples/ex1/
sbatch ex1.slurm
PBS, SLURM
This script includes all the reasonable defaults from Example 1. The only change is that it uses Rscript to run an external R script, which is how the job would usually be programmed.
The computational task in R is a copy of the example usage of ideal()
from the R
package pscl.
To run under Torque:
cd ./examples/ex2/
qsub ex2.pbs
To run under SLURM:
cd ./examples/ex2/
sbatch ex2.slurm
PBS, SLURM
This job script uses the sample reasonable defaults from above, but it requests 3 tasks with 1 processor each. These tasks may land on the same physical node, or not.
The R script uses an MPI backend to parallelize an R foreach loop across multiple nodes. A total of 3 * 1 = 3 processors will be used for this job (but 1 task is kept for the "master" process). When running the R script, we pass the value "10" as an unnamed argument. The R script then uses this value to determine how many iterations of the foreach loop to run.
Each iteration of the foreach loop simply pauses for 1 second and then returns some contextual information in a dataframe. This information includes where that MPI process is running and what it's "id" is.
To run under Torque:
cd ./examples/ex3/
qsub ex3.pbs
To run under SLURM:
cd ./examples/ex3/
sbatch ex3.slurm
PBS, SLURM
This script now requests an array of jobs based on the template. For jobs in this array (indexed from 1 to 3), the shell script will run given the requested resources. Because the log file depends on the job ID, each of the three jobs will generated different log.* output. Because R can read environmental variables we are able to use the index on an object of interest to us in R (e.g., a vector of names of files in a directory to be processed).
With this setup, each sub-job is requesting the same resources.
To run under Torque:
cd ./examples/ex4/
qsub ex4.pbs
To run under SLURM:
cd ./examples/ex4/
sbatch ex4.slurm
PBS
This script uses the default setup (see ex1), requests one task with 5
processors and runs a Matlab script. The Matlab script executes a loop
sequentially and then in parallel where each of MC
iterations takes MC/DUR
seconds by construction. The parallel loop (i.e., the one using the parfor
construct) should be about PROCS
times faster. This approach does not
generalize to parallel execution across nodes.
cd ./examples/ex5/
qsub ex5.pbs
PBS, SLURM
This example is less a demonstration of features available (e.g., there is no use of job arrays or command line arguments) and, instead, shows a computational job that provides a good template for many other embarrassingly parallel computational tasks. Here, the goal is use the non-parametric bootstrap to approximate the sampling distribution of correlation coefficients based on samples of size 25. The correlation of interest is between average undergraduate GPA and average LSAT scores among students at 82 different law schools.
The output generated from the R script is just the deciles from this distribution (without acceleration or bias-correction).
To run under Torque:
cd ./examples/ex6/
qsub ex6.pbs
To run under SLURM:
cd ./examples/ex6/
sbatch ex6.slurm
PBS, SLURM
This example mirrors Example 6. However, it demonstrates use of a single task, where that task uses multiple processes.
To run under Torque:
cd ./examples/ex7/
qsub ex7.pbs
To run under SLURM:
cd ./examples/ex7/
sbatch ex7.slurm
PBS
This PBS script uses the default setup (see Example 1), requests 5 processors on
a single node, and runs a Python script. The Python script executes a loop
sequentially and then does the equivalent in parallel. Eeach of MC
iterations
takes MC/DUR
seconds by construction. The map
-based parallel evaluation
should be about PROCS
times faster. This approach does not generalize to
multiple nodes.
cd ./examples/ex8/
qsub ex8.pbs
PBS
This PBS script uses the default setup (see Example 1), requests 5 processors on
a single node, and runs a Python script. The Python script executes a loop
sequentially and then does the equivalent in parallel. Eeach of MC
iterations
takes MC/DUR
seconds by construction. The map
-based parallel evaluation
should be about PROCS
times faster. This approach does not generalize to
multiple nodes.
cd ./examples/ex8/
qsub ex8.pbs
PBS, SLURM
PBS
- common scheduler directives: Example 1
- sequential execution of scripts
- in R: Examples 2, 10
- in Matlab: Example 5
- in Python: Example 8
- misc. shell commands in job scripts: Example 1
- job arrays: Example 4
- passing command line arguments
- in R: Example 4
- reading shell environmental variables:
- in R: Examples 3, 6, 7
- in Matlab: Example 5
- in Python: Example 8
- dynamic parallelization (i.e., not hard coding processors): Examples 3, 5, 6, 8
- single-node, multiple-core parallelization
- in R: Example 7
- in Matlab: Example 5
- in Python: Example 8
- multiple-node parallelization
- in R
- with openmpi: Examples 3, 6
- with job arrays: Example 4
- in R