UMIACS Servers

Notes on using UMIACS servers

Modules

Use modules to load programs you need to run.

Notes

• You can load modules in your .bashrc file

# List loaded modules
module list

# Load a module
module load [my_module]

# List all available modules
module avail

Some useful modules in my .bashrc file

module load tmux
module load cuda/10.0.130
module load cudnn/v7.5.0
module load Python3/3.7.6
module load git

Python

Do not install anaconda in home. You will run out of space.
Load the Python 3 module adding the following to your .bashrc file

module load Python3/3.7.6
export PATH="${PATH}:$(python3 -c 'import site; print(site.USER_BASE)')/bin"

Then run the following to get pip installed

curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py
python get-pip.py --user

Notes

• You will need to install things with pip --user
• You may need to add your local site-packages to your PYTHONPATH environment variable
  • Add this to .bashrc:
  • export PYTHONPATH="${PYTHONPATH}:/nfshomes/$(whoami)/.local/lib/python3.7/site-packages/"
• You can also install using pip install --target=/my-libs-folder/

Conda

If you must install conda, install it somewhere with a lot of space like scratch.

Install PyTorch

pip install --user torch===1.3.1 torchvision===0.4.2 -f https://download.pytorch.org/whl/torch_stable.html

Installing Packages to a Directory

pip install geographiclib -t /scratch1/davidli/python/

MBRC Cluster

See UMIACS MBRC

SLURM Job Management

See https://docs.rc.fas.harvard.edu/kb/convenient-slurm-commands/

1 GPU

srun --pty --gres=gpu:1 --mem=16G --qos=high --time=47:59:00 -w mbrc00 bash

2 GPUS mbrc00

srun --pty --gres=gpu:2 --mem=16G --qos=default --time=23:59:00 -w mbrc00 bash

CPU-only on scavenger QOS

srun --pty --account=scavenger --partition=scavenger \
     --time=3:59:00 \
     --mem=1G -c1 -w mbrc00 bash

Notes

• You can add -w mbrc01 to pick mbrc01
• -c 4 for 4 cores

See Jobs

See my own jobs

squeue -u <user> -o "%8i %10P %8j %10u %10L %5b"

Formatting

• %L is remaining time
• %b is the number of GPUs

See all jobs

squeue

SFTP

Note: If you know of an easier way, please tell me.

On your PC
Start an sshd for forwarding. You can do this in a docker container for privacy purposes.

On the cluster:
Generate an sshd host key:

ssh-keygen -t ed25519 -a 100 -f /nfshomes/dli7319/ssh/ssh_host_ed25519_key

Create the following sshd_config file

#	$OpenBSD: sshd_config,v 1.103 2018/04/09 20:41:22 tj Exp $
Port 5981
HostKey /nfshomes/dli7319/ssh/ssh_host_ed25519_key
AuthorizedKeysFile	.ssh/authorized_keys
Subsystem	sftp	/usr/libexec/openssh/sftp-server

Start the sshd daemon and proxy the port to your local sshd. You can make a script like this:

#!/bin/bash

LOCAL_PORT=5981
REMOTE_PORT=22350
REMOTE_SSH_PORT=22450
REMOTE_ADDR=$(echo "$SSH_CONNECTION" | awk '{print $1}')

/usr/sbin/sshd -D -f sshd_config & \
ssh -R $REMOTE_PORT:localhost:$LOCAL_PORT root@$REMOTE_ADDR -p $REMOTE_SSH_PORT 

On your PC:
Proxy the sshd from the local docker to your localhost.
Connect to the the sshd on the cluster

Class Accounts

See UMIACS Wiki: ClassAccounts

Class accounts have the least priority. If GPUs are available, you can access 1 GPU up to 48 hours.
However, your home disk only has 18GB and installing PyTorch takes up ~3GB.
You cannot fit a conda environment in here so just use the python module.

The ssh endpoint is

class.umiacs.umd.edu

Start a job with:

srun --pty --account=class --partition=class --gres=gpu:1 --mem=16G --qos=default --time=47:59:00 -c4 bash

My .bashrc

#PS1='\w$ '
PS1='\[\e]0;\u@\h: \w\a\]${debian_chroot:+($debian_chroot)}\[\033[01;32m\]\u@\h\[\033[00m\]:\[\033[01;34m\]\w\[\033[00m\]\$'

# Modules
module load tmux
module load cuda/10.0.130
module load cudnn/v7.5.0
module load Python3/3.7.6
alias python=python3

export PATH="${PATH}:${HOME}/bin/"
export PATH="${PATH}:${HOME}/.local/bin/"

.bashrc

My .bashrc

#PS1='\w$ '
PS1='\[\e]0;\u@\h: \w\a\]${debian_chroot:+($debian_chroot)}\[\033[01;32m\]\u@\h\[\033[00m\]:\[\033[01;34m\]\w\[\033[00m\]\$'

if test -f "/opt/rh/rh-php72/enable"; then
    source /opt/rh/rh-php72/enable
fi

export NVM_DIR="$HOME/.nvm"
[ -s "$NVM_DIR/nvm.sh" ] && \. "$NVM_DIR/nvm.sh"  # This loads nvm
[ -s "$NVM_DIR/bash_completion" ] && \. "$NVM_DIR/bash_completion"  # This loads nvm bash_completion

command_exists() {
  type "$1" &> /dev/null ;
}


# Modules
if command_exists module ; then
  module load tmux
  module load cuda/10.2.89
  module load cudnn/v8.0.4
  module load Python3/3.7.6
  module load git/2.25.1
  module load gitlfs
  module load gcc/8.1.0
  module load openmpi/4.0.1
  module load ffmpeg
  module load rclone
fi
if command_exists python3 ; then
  alias python=python3
fi

if command_exists python3 ; then
  export PATH="${PATH}:$(python3 -c 'import site; print(site.USER_BASE)')/bin"
fi
export PYTHONPATH="${PYTHONPATH}:/nfshomes/dli7319/.local/lib/python3.7/site-packages/"

export PATH="${HOME}/bin/:${PATH}"

Software

git

The MBRC cluster has an git available in the modules.
Then you can download git-lfs compiled and drop it in ~/bin/.
Make sure ${HOME}/bin is in your path and run git lfs install

Notes

• Make sure you have a recent version of git
  • E.g. module load git/2.25.1

Copying Files

There are 3 ways that I copy files to the scratch drives:

• For small files, you can copy to your home directory under /nfshomes/ via SFTP to mbrcsub00. I rarely do this because the home directory is only a few gigs.
• For large files, I typically use rclone to copy to my terpmail Google Drive and then copy back to the scratch drives with a cpu-only job. Do not do this with thousands of small files; it will take forever since Google Drive has a limit on files per second. Also note that Google Drive has a daily limit of 750GB in transfers.
• For mounting, I have a convoluted system where I start SSHD in a job and port forward the SSH port to my local PC. See above for more details.