A guide to using PBS

    Introduction #

    If you have a lot of analysis to do, your institute’s high performance computing (HPC) resources might be a good solution. These are collections of many computers that are shared among a lot of people. To use them, you submit “jobs” which are queued in a scheduler called PBS.

    PBS jobs are used to run your code on HPC. The code is not run immediately, but queued to run when resources open up. The time in the queue depends on how busy HPC is and how much time your job will take.

    PBS batch jobs (also called array jobs) are useful when you want to run your analysis on many files independently. That is, the results from one file are not used for the analysis of another and they can be done in any order. This means that they each can be sent to a different computer and run in parallel at the same time. An example of this might be running a species recognizer over thousands of audio files.

    To run a batch job you need to:

    1. Create a PBS batch script
    2. Transfer your analysis scripts and PBS batch script to HPC storage
    3. Set up the environment on HPC, such as making sure python or R dependencies are available
    4. run the PBS command line command to queue the batch job

    In this guide we first show how to complete these steps with a basic, minimal batch script. Then, we extend the script to show the basics of common tasks like running python, R, and sub-batching.

    This brief guide does not cover everything about PBS. For more extensive information, try the following guides, published by different Australian Universities:

    NOTE

    HPCs tend to only run Linux, and are only accessed via the terminal. We strongly recommend you brush up on you Unix shell basic with the Software Carpentry Unix Shell course

    PBS batch job script #

    To launch a PBS batch job, you must first create a pbs batch script. This is a bash script (a kind of scripting language) with some special PBS instructions.

    In this section we will show an basic example script and explain all the bits.

    Minimal example #

    First we will just show a bare minimum empty example. This will simply output the all the rows of a csv, with each row being assigned to a different PBS job within the PBS batch.

    #!/bin/bash -l
#PBS -N MyAnalysis
#PBS -l ncpus=8
#PBS -l mem=8GB
#PBS -l walltime=00:11:00
#PBS -j oe
#PBS -m abe
#PBS -M j.bloggs@myuniversity.edu.au
#PBS -J 0-46


# cd to dir where directory was created
cd $PBS_O_WORKDIR
echo 'working directory is $PBS_O_WORKDIR'

# the environment variable $PBS_ARRAY_INDEX
# resolves to what ever the current job index
# in the array is
echo "Array index is $PBS_ARRAY_INDEX"

# this the path to the csv (relative to the directory on HPC storage where you launch the batch job)
export data_file=./list_of_audio_files.csv

# this extracts the row of the csv that corresponds to the current PBS job within the batch
line=$(sed -n "${${PBS_ARRAY_INDEX} + 1}p" "$data_file")

echo "processing row:"
echo  $line

    The above code should be put in a file that has the extension .sub, for example my_batch_job.sub

    Configuration Options #

    The first few lines, prefixed with #PBS are configuration for the script.

    • #PBS -N MyAnalysis This is the name of the job, just for your reference.
    • #PBS -l ncpus=8 The number of CPUs. The more CPUs you allocate, the longer it might take to be scheduled because it has to wait for available resources.
    • #PBS -l mem=8GB The amount of RAM needed. Again, requesting more RAM might mean a longer wait.
    • #PBS -l walltime=00:11:00 The amount of time you estimate each job will take to process at most.
      • The scheduler will try to fit your jobs in based on this, so the higher the walltime, the longer it might take to get schedule.
      • If a job takes longer than this, the scheduler will forcibly stop your job early. In this example we specify 11 minutes.
    • #PBS -j oe This merges standard error and standard output into the same output file.
    • #PBS -m abe This defines when to send you an email. Because you can’t control and don’t know when the job will start or will finish executing, or had a problem and aborted, it is useful to receive this information by email. Here we have specified that mail should be sent when the when the job is aborted by PBS (a), when the job begins execution (b) and when the job terminates (c).
    • #PBS -M j.blogs@myinstitute.edu.au This is where to send the notification emails.
    • #PBS -J 0-46 The indexes that will be executed.
      • Using the -J option will run your jobs as a batch job - i.e. it will run the job many times instead of just once.
      • In this case, it will run this script 47 times with the $PBS_ARRAY_INDEX environment variable set to the integers 0 to 46

    Environment Variables #

    There are several environment variables that are available to the script. In the above example we use the variables $PBS_O_WORKDIR and $PBS_ARRAY_INDEX

    We use the $PBS_ARRAY_INDEX to output a different row of our csv file for each job in the batch. The script is run many times (as many as we told it to) with this value changed each time.

    Here is a complete list of environment variables

    Transferring your batch script to HPC over SSH #

    After the previous section, you will have created a pbs batch script in a text editor and saved it on your computer with the extension .sub, for example my_batch_job.sub

    To submit this batch job, you first need to transfer this file (along with any other files your batch job needs, such as the csv of files to process, any scripts, etc) to your HPC storage.

    There are several ways to do this. Many universities allow users to access storage via the network. This might look like a network drive, or network share for you. For Mac and Linux users you can access this share using the SMB protocol. However since each university is different you’ll need to access your institution’s documentation for advice.

    As far as we know, every HPC supports access via SSH. This means you can use programs like WinSCP or rclone . Again, you’ll need to refer to your institution’s documentation for details.

    Submitting a PBS Batch Job #

    In this section we are going to do the following steps

    1. In a terminal, SSH in to your institution’s HPC
    2. Submit the job
    3. Check the status
    4. Check the output

    1. SSH to your institute’s HPC #

    Note: This process may differ between institutions, so consult your the relevant documentation.

    In a terminal use the command ssh username@hostname for example ssh joebloggs@hpc.myuniversity.edu.au

    The username and hostname will be provided by your institute. There may be a firewall that prevents you from doing this from off campus without connecting to your institute’s VPN.

    Once you have connected with ssh, use the cd command to change directory to the location where you copied the my_batch_job.sub file in the previous section.

    2. Submit the job #

    Use the command qsub my_batch_job.sub to submit your batch job. This will queue it to be started when resources open up. The batch job will be given an ID, which is outputted to the terminal (e.g. 2178039[].pbs will be output for an id of e.g 2178039[]). Take a note of this so you can use it to check the job’s status.

    3. Check the status #

    To check the status use the command qstat -xJf myjobid (where myjobid is the value returned by the previous qsub command). This spits out a lot of information, for example Output_Path and Error_Path show you where the files will be saved with the output and error output.

    The main things you want to check is whether the job has started, how many jobs within the batch job have been completed and remain, which is the line that begins with array_state_count. You can use grep to view only this line of the output with qstat -xJf myjobid | grep array_state_count

    4. Check the output #

    In our example batch script, we are simply outputting the line of the csv that corresponds to the array index. If the batch job ran correctly we will have an output file for each array index submitted. Remember, we specify the array indexes in the #PBS -J 0-987 configuration option. The location of these files is given in the in output of the qstat -xJf command.

    The output should look like this (this is job number 31)

    working directory is /home/myusername/pbs_test_1
Array index is 31
processing row:

PBS Job 2180532[31].pbs
CPU time  : 00:00:00
Wall time : 00:00:04
Mem usage : 0b

    If your job writes any logging (like echo commands) you can view the output by running cat on the output files.

    Sub Batching #

    So far we have been processing one audio file (row of our csv) per pbs batch job array index. Your institution may have imposed a limit on how many array indexes you can submit with the -J option. You should check with your institution’s HPC documentation to find what this limit is.

    If you have more files to process than this limit, then you will need to process multiple files per array index. The following example batch script shows one method to accomplish this.

    In this example imagine we have a csv with 136 rows and a limit of 50 (in reality the limit will be more like 2000, but let’s keep things small for this demo)

    We need to calculate the set_size, i.e. the number of rows per job, with ceiling(num_rows / limit) which in this case equals 3. We then use the set_size to calculate the number of indexes to submit with ceiling(num_rows / set_size) which in this case equals 46.

    The pbs batch script below uses these values to calculate the which rows to process for each array index. It then has a loop that iterates over these rows one at a time. Note that this may increase the walltime you need to allocate to the job.

    #!/bin/bash -l
#PBS -N MyAnalysis
#PBS -l ncpus=1
#PBS -l mem=2GB
#PBS -l walltime=00:05:00
#PBS -j oe
#PBS -m abe
#PBS -M j.bloggs@myuniversity.edu.au
#PBS -J 0-46


cd $PBS_O_WORKDIR
echo "working directory is $PBS_O_WORKDIR"
echo "Array index is $PBS_ARRAY_INDEX"

# this the path to the csv (relative to the directory on HPC storage where you launch the batch job)
export data_file=./list_of_audio_files.csv

# load the anaconda3 module so we can use python
module load anaconda3

# activate the virtual environment we created earlier
source activate myenv

# number of rows per job
set_size=3
# total number of rows
number_records=136
let "start_row = PBS_ARRAY_INDEX * $set_size"
let "end_row = $start_row + $set_size - 1"
# if the set size does not divide exactly into the num_rows, for the last array index we need to
# make sure we don't go past the last row.
let "end_row = ($end_row > $number_records - 1) ? $number_records - 1 : $end_row"

# loop from the start to the end row, calling the python script on each execution
# alternatively, you could code your python script to do the loop,
# passing in the start and end rows as arguments

for (( idx = start_row;  idx <= end_row; idx++ )); do
  echo "starting $idx, batch $PBS_ARRAY_INDEX"
  python3 check_files.py -f $data_file -r $PBS_ARRAY_INDEX
done