Example CS cluster scripts for incoming MSc students.

• Access to a cluster account through https://hpc.cs.ucl.ac.uk/account-form/
• Optionally, storage through: https://hpc.cs.ucl.ac.uk/storage-form/
• If you don't want to jumphost, access to a VPN.

Jumping through knuckles:

ssh -J [email protected] [email protected]

Or though VPN:

ssh [email protected]

All commands below are run on a login node, in this case it's beaker. You can't (an you wouldn't want to) run computation on these nodes, but you can schedule jobs.

You have limited space in your home directory, both in size and quantity of files. Full fat anaconda will be too expensive (loads of potentially useless packages), install miniconda instead:

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh;
chmod +x Miniconda3-latest-Linux-x86_64.sh;
./Miniconda3-latest-Linux-x86_64.sh;

Let's run a simple script that will run a python file in your code directory. First make a directory in your home called runLog.

The python script, let's call it mscPythonTestScript.py, is just a simple hello-world:

print("Hello from beyond!")

See the included qsub file, mscTuturialTestScript.qsub.sh. It has comments for what every line means.

Schedule the script via:

qsub mscTuturialTestScript.qsub.sh

See an output file in ~/runLog!

Storing data in your home directory is a nono. Accessing data in your home directory from a run node is a bigger nono. You'll get mad emails asking why you did that. Instead, copy any data you need to the node where the processing happens. Here's a handy script that copies data from /home/YOUCLUSTERUSERNAME/datasets/testData.zip to the scratch space of the machine where your job will run:

#!/bin/bash

UNIQUEID=$(uuidgen)
UNIQUEID=${UNIQUEID:0:13}

mkdir /scratch0/$USER/
BASEDIR="/scratch0/$USER/"

COPYDIR="${BASEDIR}${UNIQUEID}/"
mkdir $COPYDIR

rsync -ar --info=progress2 /home/$USER/datasets/testData.zip $COPYDIR

cd $COPYDIR

unzip -q testData.zip
rm testData.zip

Your python script should read from this directory. Here's an example of what that looks like in mscPythonReadTestScript.py

import sys

print("Hello from beyond!")


print("Here's your data:")
file = open(sys.argv[1] + '/testData.txt', 'r')
Lines = file.readlines()
print(Lines[0]) 

Your qsub file should now call the script to copy your data and launch your python script with the correct directory where your data is copied. See mscTuturialCopyDataTestScript.qsub.sh for what this looks like.

Handy tool for pretty printing status:

qstat -ext -xml | tr '\n' ' ' | sed 's#<job_list[^>]*>#\n#g' | sed 's#<[^>]*>##g' | grep " " | column -t

Interactive Session with one GPU:

qrsh -l tmem=2G,gpu=true,h_rt=00:05:00,tscratch=70G -pe gpu 1

https://hpc.cs.ucl.ac.uk/quickstart/

https://hpc.cs.ucl.ac.uk/full-guide/

https://hpc.cs.ucl.ac.uk/data-storage/