simple

Simple example

This example mocks up what your experimental setup might look like: we're going to conduct a gridsearch over some parameters.

Follow the exercise below. This illustrates the experiment framework and should take you no more than 20 minutes to complete.

Exercise

We're going to:

1. make a file experiment.txt which contains all the commands to run our gridsearch. Each line is the command to run one setting of the gridsearch.
2. make a file slurm_arrayjob.sh which will be passed to the sbatch command, and be run on the selected node

0. Setup and start an interactive session

Firstly, if you haven't already, clone this repository, following the base README.md installation instructions.

Next, make sure you have made run_experiment available by running the setup instructions in the slurm experiments README.

IMPORTANT NOTE: Running processes locally i.e. here on the headnode (the node you arrive at when you ssh to a given cluster) is VERY BAD! The headnode is responsible for coordinating all the jobs for slurm; if we clog it with lots of additional jobs, it will slow everything down for everyone. Thankfully, it's easy to simply jump on one of the nodes in the cluster and work from there - everything works the same as if you were on the headnode:

cluster_name=cdtcluster  # e.g. mlp or ilcc-cluster
ssh ${USER}@${cluster_name}.inf.ed.ac.uk

# Run this slurm command to be allocated to a node in the default partition
# We have set a maximum time and a few other constraints:
srun --time=02:00:00 --mem=4000 --cpus-per-task=1 --pty bash

# You can also use our shortcut command which essentially does the above!
interactive

When you are finished, execute the exit command to return to the headnode.

Tip - if the srun command is taking a long time to allocate you an interactive session, run sinfo and establish if you should use a specific partition for this. For example:

$ sinfo
> PARTITION         AVAIL  TIMELIMIT  NODES  STATE NODELIST
> Teach-Interactive    up    2:00:00      3   idle landonia[01,03,25]
> ...
$ srun --partition=PGR-Standard ...

1. Make a python script to run an experiment train.py

Well actually...

We've made one for you! This will read data, configure the model, and return results. Have a look at the file you're going to run (either on the server with e.g. cat train.py or in your browser train.py).

Try running the file with different options to see how it works:

# edit this if you cloned the repo elsewhere
repo_dir=/home/$USER/cluster-scripts
cd $repo_dir/experiments/examples/simple

# Print the man page
python3 train.py -h

# Set the input path and output path, with arguments -i and -o, and run!
python3 train.py -i data/input -o data/output

# You have just run an experiment! Check out what it did...
ls data/output
cat data/output/*

# Try running another experiment!
python3 train.py -i data/input -o data/output --lr 0.001
ls data/output
cat data/output/*

# Which learning rate was more accurate?!

You should be able to answer the following:

1. What are the arguments you can supply to train.py
2. Which arguments are optional, and what happens if you do not set them?

2. Create experiment.txt - the commands to run

So, we want to make a file which lists all the experiments you want to run, like you were doing manually in step 1, but you want to do 1000s. We have made a script which will generate this file for you. Generating your experiments.txt file is a good idea for you to do in general because it's:

1. much quicker than manually writing, copy pasting, and editing
2. much safer (it's reproducible!)
3. easy to make another file with slightly different settings

Again have a look at the file you're going to run (either on the server with e.g. cat gen_experiment.py or in your browser gen_experiment.py).

In your terminal, change directory as below, run the script to make the experiment file, and a look at the file:

cd $repo_dir/experiments/examples/simple
python3 gen_experiment.py
head experiment.txt
wc experiment.txt

You should be able to answer the following:

1. How many jobs are we going to run in our gridsearch
2. If each job was going to take around an hour, and we are able to run 10 jobs in parallel at any given time, how long will the full experiment take to run

3. Create the slurm_arrayjob.sh - the bash script for the slurm arrayjob

So we could simply run experiments.txt (e.g. bash experiments.txt), and it will run each experiment in sequence, one after the other... but that would be very slow!

We're going to use slurm to schedule these experiments to run in parallel. This way, the whole lot could take as little time as running just one.

The slurm command for sending jobs to the cluster's nodes is sbatch. At the most basic level, it just takes a bash script from you and executes that same script on every node it is sent to. We are going to use it in 'array mode'.

We need to make a bash script which:

1. configures any slurm options we want to use
2. moves data from the Distributed File System (DFS - available on all nodes) to the scratch space of the node slurm has allocated the job to (e.g. /disk/scratch - which is only available on this specific node)
3. runs one of the lines in experiments.txt

You're in luck, because we have made you just such a script! Have a look at slurm_arrayjob.sh and answer the following questions (answers at the bottom of the README, no cheating!):

1. where is the data located on the distributed filesystem?
2. where does the python script read the data from?
3. how did the data get there?
4. does the data get moved every time?
5. why are these different places?
6. where is the output data?
7. what happens to the output data if you do a second run of the experiment?
8. how does the script run different lines of experiment.txt?

4. Run your experiment!

Finally, it's time to see it in action. To set off all the jobs, run:

sbatch --array=1-35%10 slurm_arrayjob.sh experiment.txt

To observe your jobs running, you can execute:

myjobs  # shorthand for squeue -u ${USER}
# or if you want to see it in real time
watch myjobs  # hit ctrl-c to exit

You'll find all the logs printed in the current working directory:

ls
cat slurm-*

A more general way of doing the above (which doesn't rely on you knowing the number of lines in experiment.txt):

run_experiment -b slurm_arrayjob.sh -e experiment.txt

With any luck, you should now have all your results!

ls data/output
tail -n +1 data/output/*  # a little trick to print filenames and file contents

CONGRATULATIONS! You just used the cluster to run all your experiments!

You can delete your logs easily using:

rm slurm-*.out

Answers

1.

1. What are the arguments you can supply to train.py
   • There are 4 arguments:
     • -i INPUT - path to read the input data from
     • -o OUTPUT - path output data should be written to
     • --lr - learning rate to use
     • --weight_decay - weight_decay to use
2. Which arguments are optional, and what happens if you do not set them? --lr and --weight_decay are optional. If you don't specify them the default values are used. The default values are 1e-4 and 0.01 respectively (you can find this specified in the script)

2.

1. How many jobs are we going to run in our gridsearch
   • 35 - 5 learning rates x 7 weight decays
2. If each job was going to take around an hour, and we are able to run 10 jobs in parallel at any given time, how long will the full experiment take to run
   • It's a fair estimate to say 3.5 hours, but actually it will take 4 hours: when the 3rd batch of 10 concurrent jobs have finished, you've still got 5 left over, which will still take an hour

3.

1. where is the data located on the distributed filesystem?
   • Well...right here! Specifically, if you cloned this git repo to ~/cluster-scripts, as recommended, this location has the same path on all the nodes, i.e. /home/${USER}/cluster-scripts/experiments/examples/simple/data/input.zip points to the distributed filesystem whichever node you are on
2. where does the python script read the data from?
   • The scratch space of node which the job gets allocated to e.g. the path /disk/scratch/${USER}/... which is only accessible from this node
3. how did the data get there?
   • In slurm_arrayjob.sh, rsync moves the zip file over, and unzip unzips the zip file
4. does the data get moved every time?
   • No. The option -u in rsync will only update the file if the version on the distributed filesystem in newer, or the file doesn't exist on the scratch disk.
   • Bonus. If the data is already unzipped, the -n flag on the zip command always opts not to overwrite existing files.
5. why are these different places?
   • The DFS is very slow to read and write to. Read your experiment data from the node's scratch disk, write your results there, and then copy the results over at the end of the job
6. where is the output data?
   • Well, the output for individual runs is contained on the scratch disk of the node, but the amalgamated outputs are sent back to the DFS. The last part of slurm_arrayjob.sh moves the data back to the DFS.
7. what happens to the output data if you do a second run of the experiment?
   • The scratch disk of each node may now have two files inside (it's possible the same experiment will get sent to the same node again, and the output file will simply be overwritten). However, when rsync looks at these two files and compares them with the files on the DFS, it will decide to only send the one newer file (as the older file on scratch will either be unchanged on the DFS, or a newer file will already exist on the DFS)
8. how does the script run different lines of experiment.txt?
   • The script takes one argument as input - the path to the experiment file. It uses the slurm environment variable ${SLURM_ARRAY_TASK_ID} - if you run the command sbatch --array=3-14 ..., the jobs will receive numbers 3 to 14 inclusive. That number is stored in ${SLURM_ARRAY_TASK_ID} and accessible within each job. The script simply picks line number ${SLURM_ARRAY_TASK_ID} from the file experiments.txt and runs it.