parallel-python

Python code samples from the High Performance Computing with R and Python workshop, Presented by Brett Vanderwerff on 11/13/18

Setup

To follow along with the guide, you will need an account on Kamiak, which can be requested at: https://wsu-its.atlassian.net/servicedesk/customer/portal/1/user/login?destination=portal%2F1

You also need to be able to log in to Kamiak. Info on logging in can be found here: https://hpc.wsu.edu/users-guide/terminal-ssh/ and

here: https://s3.wp.wsu.edu/uploads/sites/1122/2018/11/kamiak_cheat_sheet_v7.pdf

Once logged in, you will need to clone this repository into your home directory by entering the command

$git clone https://github.com/wsucougpy/parallel-python

(Note: don't add the $ for this command or any other command in this guide. The $ prefix just denotes a shell command)

After cloning the repository, you should be able to use the ls command to see the repository

$ls

Which should output something similar to:

[brett.vanderwerff@login-p1n02 ~]$ ls
parallel-python
[brett.vanderwerff@login-p1n02 ~]$

Once you have this repository, you are ready to start :)

Creating a Conda Environment

We can request an idev session using the idev command, which takes us from the login node to a compute node

$idev -t 30 

Load the anadonda3 module

$module load anaconda3

Check out the list of packages that are readily available without needing to set up an environment

$conda list -n root

Create a conda environment named “env”

$conda create -n env

(say yes when prompted)

Activate the environment “env”

$source activate env

Install pip to “env”

$conda install pip

(say yes when prompted)

Install a package from the Python Package Index to “env” (Note: this is just an example package)

$pip install bitstring

Install a package from an anaconda repository to “env” (Note: this is just an example package)

$conda install flask-sqlalchemy

(say yes when prompted)

Deactivate the “env” environment

$source deactivate

Exit the idev session

$exit

Submitting a Python Job to SLURM (serial example)

Change into the parallel-python directory using the cd command. Then use the ls command to look at the files there

$cd ~/parallel-python

$ls

serial.py is a demo script that sums all the prime numbers between 0 and 10^5. We can see the contents of this file with the cat command

$cat serial.py

serial_job.sh is a script for submitting serial.py to SLURM (Kamiak’s job scheduler). We can see the contents of this file with the cat command

$cat serial_job.sh

#!/bin/bash is the “bang” line pointing to the interpreter we want to use to run this file.

The next few lines all start with #SBATCH and are recognized as arguments by SLURM:

• partition – the partition we want to run on
• job-name – the name of our job
• output – the file we want to write the standard output to
• error – the file we want to write standard error to
• time – the amount of “wall clock” time we want for the job
• nodes – the number of nodes we want to run the program on
• ntasks-per-node – the number of times we want to run the program on our node
• cpus-per-task – the number of cores for our program

The rest of the script gives commands to:

• load the anaconda3 module
• activate the conda environment “env” that we made earlier
• use python to run the serial.py script
• deactivate the “env” environment

Submit the job to the job scheduler using the sbatch command

$sbatch serial_job.sh

(when scheduled, it takes about 30 seconds to run)

If the job is still running, we can check the progress using the squeue command (if the job already completed we won’t see anything here)

$squeue -u your.name

When the job is complete we can use ls to see that an output file has appeared. We can see the contents of this file with the cat command

$ls

$cat serial_prime*.out

Submitting a Python Job to SLURM (parallel example)

Change into the parallel-python directory using the cd command. Then use the ls command to look at the files there

$cd ~/parallel-python

$ls

parallel.py is a script that sums all the prime numbers between 0 and 10^5 using a parallel approach with the multiprocessing module from Python’s standard library. We can see the contents of this file with the cat command

$cat parallel.py

parallel_job.sh is a script for submitting parallel.py to SLURM (Kamiak’s job scheduler). We can see the contents of this file with the cat command. The only thing different between parallel_job.sh and serial_job.sh is that now we are asking for 4 cpus instead of 1

$cat parallel_job.sh

Submit the job to the job scheduler using the sbatch command

$sbatch parallel_job.sh

(when scheduled, it takes about 15 seconds to run)

Check out the results!

$ls

$cat parallel_prime*.out