Dags Server

Load modules

In order to use Jupyter notebook and python3.7, I need to load module python/conda/3.7 first.

To automatically load specific software module or modules every time you log into the server, edit your .bashrc file:

1

> module load python/conda/3.7

Install packages

Keep a habit of using virtual environments, especially in a remote server where I am not the root. One biggest different between local machines and remote machines is that you won’t have the permission to install packages as you want, and you are only allowed to install packages on your local directory. If you want to install a python package for general purposes, use pip install --user packagename. But usually, we work on projects and should set different environments for different projects, so creating a conda virtual environment and install packages inside would be a better idea.

1
2
3
4
5

> module load python/conda/3.7
> conda create -n envname python=3.7
> conda activate envname
> conda install packages
> conda deactivate

Note: in order to use conda activate, we need to run conda init bash to initialize conda commands.

Now you can install packages you would like to use, without worring about not having root permissions.

To see the information of the environments, run:

1

> conda info --envs

Jupyter notebook

References:

• Running a Jupyter notebook from a remote server
• How do I remove an SSH forwarded port

If you have created a conda virtual environment as before and have activated it, then jupyter notebook is defaultly installed. There are several steps.

1. In the remote host, run:

1

> jupyter notebook --no-browser --port=8889

1. In the local computer, run:

1

> ssh -N -f -L localhost:8888:localhost:8889 x500@dags

Note: I have configured dags in my ssh configuration, so I just use the shortcut of the host name.

1. Open local browser and type localhost:8888.

Use functions to simplify

Use functions or aliases to make the above steps easier. The reference used functions and for simplicity, I aliases.

1. In the remote host, add the function below to .bashrc.

1
2
3

function jpt(){
	jupyter notebook --no-browser --port=$1
}

1. In the local host, add the function below to .zshrc. (I am using zsh in my local computer.)

1
2
3

function jptt(){
	ssh -N -f -L localhost:"$2":localhost:"$1" x500@dags
}

Note: for zsh, do remember to add single quote or double quote.

1. Then every time you want to run jupyter notebook, just do:

remotely,

1

> jpt 8889

locally,

1

> jptt 8889 8888

Open browser and type localhost:8888. Done!

Use aliases to simply

For simplicity, I set the default remote port to be 8889, and the default local port to be 8888. Feel free to change it.

1. Edit .bash_aliases in the remote server.

1

alias jpdags="jupyter notebook --no-browser --port=8889"

1. Edit .zshrc in the local machine.

1
2
3
4
5

# establish ssh tunnel and open the browser
alias jpdags="ssh -N -f -L localhost:8880:localhost:8889 x500@dags; open 'http://localhost:8880/tree?'"

# kill related ssh processes
alias kjpdags="ps aux | grep '8889.*dags'"

Now, the browser window will automatically open.

Note: remember to add single quote around the url link and the regular expression in grep.

1. Close jupyter in the remote server using ‘Ctrl + c’.

2. Kill relevant ssh processes by running kjpdags and kill <PID> with PID shown in the grep results.

Install Packages

Keep in mind that all of these packages should be installed in a virtual environment.

PyTorch

First, check the available cuda modules and the GPU information. In dags server, the newest version of cuda is 10.0.130, and the cuda version for the GPUs is 10.2. Which version should I use then? 10.2 means the GPU can deal with at most 10.2 version of cuda, so it is fine to have lower versions. In the official pytorch website, the available cuda versions for 1.5.0 are 9.2, 10.1, and 10.2. I tried all of them using conda, and none of them works. I then tried pip under the version 10.1, and it worked!

1
2
3
4

> module avail
> nvcc -V

> nvidia-smi

1

> pip install torch==1.5.0+cu101 torchvision==0.6.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html

Helpful Tools

Remote Development using VS Code

References

• Remote Development using SSH

Install the Remote Development extension pack in VS Code, and this would enable us to use VS Code to edit remote files.

• Open VS Code, press F1 to open the command plate, select Remote-SSH: Connect to Host…, and choose the remote server. I have set ssh configurations, so I can directly see the available remote servers. Done!
• VS Code would remember your action, so next time you open VS Code, you can directly check recent activities to quickly connect ssh.

TMUX

It is convenient to use tmux to run multiple processes without logging into the same machine for several times. Also, the multiscreen functionality is helpful.

• $tmux: create a new session.
• $tmux list-sessions: list sessions.
• Ctrl + b: controlling keyword
  • +x: kill a pane
  • +": split a pane horizontally
  • +%: split a pane vertically

Other Configurations

aliases

• I uncommented unalias rm in the .bash_aliases file, since the rm questions are disturbing.