1. Show how to prototype a MC study interactively on Okapi.
2. Show how to unit test by submitting a slurm job.
   2b) Students conduct unit tests.
   
3. Provide information to scale to embarrasingly large parallel simulations.

You’ll be able to…

• Appreciate the “why” you may need to compute on a server.
• Understand the basics of slurm batch scheduling.
  
• Become familiar with available software and workflows for computing on a server.
  
• Log onto Okapi the Server and submit a basic slurm job to obtain MC results.

• Brief motivation / introduction to Monte Carlo Simulation
• R gives us unique access to great simulation tools (unique compared to other languages).
• Why simulate? Welcome to the 21st century! Two reasons:
  1. Often, simulations can be easier than hand calculations
  2. Often, simulations can be made more realistic than hand calculations.
• See M. Davidian’s slides for more.

You can follow Dr. Eric Olson’s guidance on the Okapi the Server homepage to open a ssh portal to enable a virtual desktop from off-campus.

On Mac/linux open a terminal and execute:

ssh -L 33389:localhost:3389 -l <NETID> okapi.math.unr.edu

replacing <NETID> with your user name. Then open up a MS Remote Desktop instance. Let’s do that now.

From Rizzo (2019) (Example 7.1):

Suppose that \(X_1, X_2\) are iid from a standard normal distribution. To estimate the mean difference \(E|X_1 - X_2|\) we will obtain a Monte Carlo estimate of \(\theta = E|X_1 - X_2|\) based on \(m=1000\) replicates of random samples \(x^{(j)} = (x_1^{(j)}, x_2^{(j)}), j=1,\ldots,m\) of size 2 from \(N(0,1)\). Then compute the replicates \(\hat{\theta}^{(j)} = \frac{1}{m}\sum |x_1^{(j)} - x_2^{(j)}|\)

set.seed(44)
m <- 1000
g <- numeric(m)
for (i in 1:m) {
    x <- rnorm(2)
    g[i] <- abs(x[1] - x[2])
}
est <- mean(g)
est

## [1] 1.119111

The exact theoretic answer by integration is \(E|X_1-X_2|=2/\sqrt{\pi} \dot{=} 1.128379\).

args = commandArgs(trailingOnly = TRUE) ## Expect command line args at the end. 
tmpM = as.numeric(args[1])
tmpSeed = as.numeric(args[2])
estimateTheta <- function( m = 1000, seed = NULL ) {
    if (!is.null( seed )) { set.seed( seed ) }
    est <- mean ( replicate( n = m, expr = {
        x <- rnorm(2)
        abs(x[1] - x[2])
    } ) )
    return( est )
}
(est <- estimateTheta( m = tmpM, seed = tmpSeed )) # print and store
jobID <- paste( tmpM, tmpSeed, sep="_" )
saveRDS( data.frame( m = tmpM, seed = tmpSeed, est = est),
        paste0( "thetaHat_", jobID, ".rds") ) 

For more on this topic see this nice blog post:

https://thecoatlessprofessor.com/programming/r/working-with-r-on-a-cluster/.

 Rscript ./estimateTheta.R 100 44 > estimateTheta.Rout

• Lets create a script in R Studio and save for submitting to the scheduler.

• Create a script that will submit small-scale simulation through the slurm scheduler.
• Later, we’ll execute via the slurm command sbatch estimateTheta.slm.

#!/bin/bash
#SBATCH -n 1
#SBATCH --mem=8GB
Rscript ./estimateTheta.R 100 4444 > estimateTheta.Rout

Before executing the script via the slurm command sbatch estimateTheta.slm, check the queue

sinfo
squeue -p fast

sbatch estimateTheta.slm

#!/bin/bash
#SBATCH -n 1
#SBATCH --mem=8GB
Rscript ./estimateTheta.R 1000000 44 > estimateTheta.Rout

sbatch estimateTheta.slm

squeue -p fast
squeue -u aschissler

less estimateTheta.Rout

scancel <JOBID>
scancel -u aschissler ## cancel all your jobs

1. Log onto okapi.
2. Create estimateTheta.R and estimateTheta.slm
3. Check the queue.
4. Submit a small-scale unit test.
5. Submit a larger-scale, more precise simulation.
6. Check the queue as you go.
7. Check the output.

#!/bin/bash
today=`date +%Y-%m-%d.%H:%M:%S`
id=0
echo "$today"

for l in 15000 10000 1000 500 250 100 50 25 12
do
    for group in 0 1
    do
    id=$((id+1))
    ## id=$(echo $l-$k)
    echo "working on pear"$id
        echo "#!/bin/bash   
#SBATCH -n 1
#SBATCH --job-name="$id"
#SBATCH --mem=4GB

cd ~/Research/Mul_NB/brca_nb_pearson_vital

time Rscript ./estimate_R_tcga_brca_nb_pearson_group.R "$l" "$group" " > "launch-"$id".slm"
        chmod a+rx "launch-"$id".slm"
        sbatch < "launch-"$id".slm"
        rm "launch-"$id".slm"
    done
done

• rslurm Getting Started vignette to specify a job, submit a job, and inspect the results.
• rslurm automates slurm scheduling for embarrasingly large parallel jobs in the Rprogramming language
• Quick look at Getting Started vignette.

• slurmR is a light-weight wrapper to automate slurm jobs for USC Biostats. https://github.com/USCbiostats/slurmR

• conda
• Python-based virtual environments
• Install packages/tools without sudo.
• Some tutorials by Alex Knudson.

Summary of workflow

1. Prototype your script.
2. Unit test your script: first quick progressing to more intense.
3. Deploy many jobs (emabarrasingly parallel) for comprehensive simulation settings.

Software used

• Mac OSX / Linux terminal
  
• MS Remote Desktop.
• slurm. See here for some convenient commands.
• R, rmarkdown, Rstudio,

• Thank you to Dr. Eric Olson, Chair of Computing Committee.
• Questions?
• Dr. Ahn would you like to describe the HW assignment?

Refererences