Skip to content

Latest commit

 

History

History
154 lines (110 loc) · 4.17 KB

README.ARC3.md

File metadata and controls

154 lines (110 loc) · 4.17 KB

Notes for using ARC3

Installing conda on ARC3

ARC3 doesn't have conda installed so we first need to install conda. NOTE: replace earlcd with your own username in the instructions below.

The commands below will create a directory for yourself in /nobackup on ARC3 and install conda there

mkdir -p /nobackup/$USER/
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
chmod +x Miniconda3-latest-Linux-x86_64.sh
./Miniconda3-latest-Linux-x86_64.sh -b -p /nobackup/$USER/miniconda3
source ~/.bashrc

You should now have conda available and can install the dependencies.

Setting up SSH aliases and keys

Defining aliases and settings up ssh keys will save you a lot of time being able to directly ssh into ARC3 and copy to/from ARC3 in one command.

  1. In the following replace earlcd with your own username and paste into .ssh/config (create the file if it doesn't exist):
Host leeds-login
        HostName feeble.leeds.ac.uk
        User earlcd

Host leeds-arc3
        HostName arc3.leeds.ac.uk
        ProxyCommand ssh leeds-login -Y -W %h:%p
        User earlcd
  1. Create an ssh-key
# ssh-keygen
  1. Use ssh-copy-id to set up passwordless login to leeds-login and leeds-arc3. After each of these commands you will be prompted for your password.
ssh-copy-id leeds-login
ssh-copy-id leeds-arc3

You can now directly ssh into ARC3 with

ssh leeds-arc3

And copy files with for example

scp localfile.txt leeds-arc3:~/{path-in-homedir}/

Training on ARC3

First log in to ARC3

[earlcd@cloud9 ~]$ ssh arc3.leeds.ac.uk

You can either a) run the training interactively by first requesting an interactive job from the scheduler with a GPU attached or b) submit a batch job to the scheduler to do the training non-interactively

a) Interactive job

While you're making sure that you have everything installed correctly and the training data in the correct place it can be easiest to do the training interactively.

Request a node with a GPU attached

[[email protected] ~]$ qrsh -l coproc_k80=1,h_rt=1:0:0 -pty y /bin/bash -i

Once the node has spun up and the prompt is again available you will notice the hostname has now changed (to something containing "gpu", here db12gpu1). Now activate your conda environment (installed as above) and start the training

[[email protected] ~]$ conda activate convml_tt
[[email protected] ~]$ python -m convml_tt.trainer <path-to-your-dataset>

Or you can open a jupyter notebook and forward traffic from the GPU so that you can run the training using a jupyter notebook:

[[email protected] ~]$ jupyter notebook --no-browser --port 8888 --ip=0.0.0.0

To do this you'll need to start a second ssh connection to ARC3 from your work station, this time forwarding the local port 8888 to port 8888 on the GPU node (here db12gpu1) you have been allocated on ARC3:

[earlcd@cloud9 ~]$ ssh arc3 -L 8888:db12gpu1.arc3.leeds.ac.uk:8888

Open up a local browser on your workstation and browse to http://localhost:8888

b) Batch job training

To submit a job to the SGE scheduler on ARC3 you will need a script like the following:

#!/bin/bash

#Run with current environment (-V) and in the current directory (-cwd)
#$ -V -cwd

# request one hour
#$ -l h_rt=01:00:00

# with on k80 GPU
#$ -l coproc_k80=1

# setup conda, change this to the path where you have conda installed
. "/nobackup/earlcd/anaconda3/etc/profile.d/conda.sh"
conda activate convml_tt

# extract data to TMPDIR
echo "Extract data to $TMPDIR"
DATAFILE="/nobackup/earlcd/convml_tt/data/goes16__Nx256_s200000.0_N500study_N2000train.tar.gz"
tar zxf $DATAFILE -C $TMPDIR
ls $TMPDIR
tree -d $TMPDIR

# run training
DATA_PATH="$TMPDIR/Nx256_s200000.0_N500study_N2000train/"
python -m convml_tt.trainer $DATA_PATH --gpus 1 --num-dataloader-workers 24 --max-epochs 10 --log-to-wandb --batch-size 128

The above script ensures you have a node with a GPU, have conda available and the environment activated and also copies the dataset to the local storage on the node where you are training (which is a lot faster than using nobackup to read training data from).