Array and Sequential Cluster Jobs

Recently, I found a few concepts very useful when running jobs on a Sun Grid Engine (SGE) cluster, and found others had not heard about them.

First and foremost, Googling “array jobs”, or “jobs sge cluster” will get you pretty far and so I highly suggest doing that before going anywhere else (after the –help of course). For our cluster, many topics are covered here.

Quick rundown of a batch job. Let's say you have an R script, let's call it Rscript.R. In a bash script (.sh file), let's call it Rscript.sh, you may have the commands

#!/bin/bash 
R CMD BATCH Rscript.R

Now, you need to go into your SGE computer (our's is called enigma), and run:

qsub /path/to/script/Rscript.sh

or use the -cwd option if you are running this command in the same directory as Rscript.sh. Boom! Job submitted to the nodes (which are a bunch of CPUs with aggregated memory). Let your code run and do all the fun stuff you normally would on your local machine not getting bogged down.

Array Jobs

Now you may want to run this script under certain “scenarios”. This is highly useful when I'm doing simulations. Let's say I had three parameters: x, y, z and wanted to vary these over simulations. Let's take advantage of a useful R command: expand.grid. This will give you all combinations of these 3 variables.

x <- c(0, 1, 2)
y <- c(2, 4, 6)
z <- c(1, 2)
scenarios <- expand.grid(x = x, y = y, y = z)
scenarios
   x y y
1  0 2 1
2  1 2 1
3  2 2 1
4  0 4 1
5  1 4 1
6  2 4 1
7  0 6 1
8  1 6 1
9  2 6 1
10 0 2 2
11 1 2 2
12 2 2 2
13 0 4 2
14 1 4 2
15 2 4 2
16 0 6 2
17 1 6 2
18 2 6 2

This will give you an object called scenarios that is a data.frame with named columns x, y, z (because I used the x=x syntax). Be aware stringsAsFactors is an option if you're using character vector and don't want them to be converted to factors. Also see combn if you're interested in only a subset of the combinations.

Now, we have 18 scenarios and I don't want to run a for loop, I want to parallelize them (embarrasingly). I now want to create what's called an array job. Now our Rscript.sh would be:

#!/bin/bash 
#$ -t 1-18
R CMD BATCH Rscript.R

So that 18 jobs are run when

qsub /path/to/script/Rscript.sh

is executed. Well that's all fun and good, but we need to explain how to reference these individual runs. So the -t assigns each of the jobs a “task number”. This is how we will index scenarios. So in your Rscript.R, let's add

iscen <- as.numeric(Sys.getenv("SGE_TASK_ID"))
scenarios[iscen,]

This will grab the SGE_TASK_ID (bash) environment variable (from the Sys.getenv command), and just change it to numeric. This can now be used to index the scenario that I want. Now I can run my simulation with each specific scenario in parallel. That should speed things up significantly with minimal cost of reworking code or new packages.

Sequential Jobs

I've worked on a cluster for a while and only recently have I used sequential jobs, where one job only starts after another has finished. Now, you can submit our first job:

qsub /path/to/script/Rscript.sh

and let's say it gets a task id of 1234. Now let us execute our second job:

qsub -hold_jid 1234 /path/to/script/Rscript2.sh

Now, Rscript2.sh will only run after Rscript.sh is done. This is highly useful in pipelining or an analysis scheme that depends on certain results before running.

Now, what if it's not going to be task id 1234? If you use the -N option, you can name the job, let's say “script1”, and we have:

qsub -N script1 /path/to/script/Rscript.sh
qsub -hold_jid script1 /path/to/script/Rscript.sh

Hopefully you can see how this would be useful in creating a reproducible series of jobs that would replicate an analysis/create a pipeline/do a series of steps.

Now for a quote from one of my favorite (and founder of) Jesuits: “Go forth and set the world on fire.” – St. Ignatius of Loyola

Advertisements

6 thoughts on “Array and Sequential Cluster Jobs

  1. I use array jobs all the time to do simulations, where each job does the exact same thing, but uses different (randomly generated) data. Say I want to test the performance of a model on 1000 randomly generated datasets. I’ll typically write code that will generate 10 datasets, then submit 100 versions of it as an array job, and have each one save the results to a separate file. When it’s all done I just have to write a new piece of code to bring all the results together again. Runs about 100 times faster than if I submitted everything as one loop!

    This works great whenever you’re doing a bunch of the same thing over and over again, and the jobs don’t have to talk to each other at all.

  2. I also enjoyed the sequential jobs tip – super helpful!

    My personal favorite way to submit a bunch of jobs in parallel is to write a “wrapper” bash script that creates little bash scripts for each scenario (in a genomics setting, maybe that means one script for each sample, or you could maybe envision looping over each of x, y, and z and passing them as args to Rscript…). The trick here is to use the cat/EOF syntax in bash. As an example, the file “wrapper.sh” might contain:

    #!/bin/sh
    for sample in `ls mysampledir`
    do
    cat > ${sample}_script.sh << EOF
    #!/bin/sh
    # here's where you put analysis commands
    # could pass $sample as a command line argument to R
    # could run some other software on $sample
    # could just echo the sample's name…
    echo $sample
    EOF
    qsub ${sample}_script.sh
    done

    Then you just do "sh wrapper.sh", which creates all the sample_script.sh files and submits them. I've found this way to be useful for jobs that aren't R jobs, or in situations where it's hard to make your thing a function of the task ID. For simulations that ARE in R, I like doing things the way you described in the post.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s