Jupyter

From FarmShare

Revision as of 18:39, 15 October 2017 by Phn (Talk | contribs)
Jump to: navigation, search

Contents

Introduction

Project Jupyter evolved out of the IPython project (specifically the IPython notebook) with the goal to provide an interactive, web-browser driven, language-independent programming environment. Jupyter notebooks can be deployed on the FarmShare servers to enable an accessible, powerful, and persistent computational platform.

Features

At the end of this guide, the resulting Jupyter notebook will support:

  • Python, MATLAB, R, Julia, and SAS programming languages
  • an encrypted, token protected, and web-browser enabled programming environment
  • indefinite persistence of the Jupyter notebook server environment with simple weekly renewals (the maximum duration of Stanford Kerberos tickets)
  • file/data storage on the Stanford AFS servers (5GB user quota, Stanford-wide, automatic backups)
  • shared file/data storage to Class Disk AFS Space
  • shared file/data storage to Group AFS Space
  • easy deployment on any of the Stanford FarmShare2 systems or any Ubuntu Xenial x86_64 system with Stanford Kerberos and AFS.

Overview

The guide consists of three sections: installation, Jupyter server, and client connection.

  • Installation is only performed once per user.
  • The Jupyter server needs to be started once per FarmShare server used. It will continue running until shut down (or the server is restarted).
  • Client connection must happen every time a client computer reconnects to the Internet (e.g., a laptop wakes from sleep).

Commands to be typed in are in bold, text to substitute is in red, and the rice server and TCP port to note are in blue.

Tutorial Videos

Overview

This is a high-level overview video of the concepts involved in running Jupyter on FarmShare:

Four main steps

There are video tutorials for the main steps:

Windows specific components

There are two Windows-specific steps that are different from the above tutorial videos:

Installation

Installation graph

This only needs to be performed once per user.

SSH into FarmShare

In a terminal, SSH into a FarmShare2 computer.

$ ssh jane@rice.stanford.edu

substituting SUNet ID for jane.

A terminal application and SSH client are shipped with Mac and Linux systems. Windows does not come with an SSH client, but putty is a a free and lightweight SSH client for Windows.

Bind to Jupyter virtual environment and install default configuration

rice99:~> bash
(ignore this command if prompt already looks like: jane@rice99:~$ )

jane@rice99:~$ source /afs/ir.stanford.edu/group/bil/env/j2/bin/activate

(j2)jane@rice99:~$ generate_jupyter_wrapper

Jupyter server

Running Jupyter graph

This only needs to be performed once per FarmShare server used (and after every server reboot).

SSH into FarmShare2 if not already in a system (ok to continue on from existing SSH session used for Installation).

Launch Jupyter server

Make a note of exactly which FarmShare server is being used (e.g., rice14, rice22, etc).

(j2)jane@rice99:~$ pagsh

(j2)jane@rice99:~$ kinit -r 7d; aklog
Password for jane@stanford.edu:

(j2)jane@rice99:~$ tmux
(new blank terminal appears)

jane@rice99:~$ source /afs/ir.stanford.edu/group/bil/env/j2/bin/activate

(j2)jane@rice99:~$ keep_kerberos_afs
Run: export KRB5CCNAME=FILE:/tmp/.krb5_jane.tgt

jane@rice99:~$ export KRB5CCNAME=FILE:/tmp/.krb5_jane.tgt

(j2)jane@rice99:~$ jupyter_start

Substitute SUNet ID for jane in the export line (as instructed by the keep_kerberos_afs script). If successful, the output of this command will be

IMPORTANT: This Jupyter notebook is listening on TCP port 9876

[I 17:23:59.661 NotebookApp] Loading IPython parallel extension
[I 17:23:59.668 NotebookApp] Serving notebooks from local directory: /afs/ir.stanford.edu/users/j/a/jane
[I 17:23:59.668 NotebookApp] 0 active kernels
[I 17:23:59.668 NotebookApp] The Jupyter Notebook is running at: https://localhost:9876/?token=ba682763f27d8e2d59862badef28b0eaecb552529933176e
[I 17:23:59.668 NotebookApp] Use Control-C to stop this server and shut down all kernels (twice to skip confirmation).
[C 17:23:59.671 NotebookApp]

    Copy/paste this URL into your browser when you connect for the first time,
    to login with a token:
        https://localhost:9876/?token=ba682763f27d8e2d59862badef28b0eaecb552529933176e

The Jupyter notebook will not work until a client connection is established. Copy the https://localhost:9876/?token=... URL and save it for subsequent use. Also note the TCP port in blue that the Jupyter server is listening on.

Confirm Jupyter server status

Confirm that the Jupyter server was successfully launched.

Control-B then C (releasing all keys after the Control-B before hitting C)
(opens a new tmux window and switches to it)

jane@rice99:/afs/ir.stanford.edu/users/j/a/jane$ cd

jane@rice99:~$ source /afs/ir.stanford.edu/group/bil/env/j2/bin/activate

(j2)jane@rice99:~$ keep_kerberos_afs
krenew already running, no action performed.
Run: export KRB5CCNAME=FILE:/tmp/.krb5_jane.tgt

(j2)jane@rice99:~$ export KRB5CCNAME=FILE:/tmp/.krb5_jane.tgt

(j2)jane@rice99:~$ jupyter notebook list
Currently running servers:
https://localhost:9876/?token=ba682763f27d8e2d59862badef28b0eaecb552529933176e :: /afs/ir.stanford.edu/users/j/a/jane

If successful, the output will list the new running Jupyter notebook, its TCP port, token, and home directory where the notebook was launched.

Note: every tmux window that is launched needs to have the KRB5CCNAME environment variable set, so be sure to run keep_kerberos_afs for every new tmux window launched.

Detach and logout

Detach from tmux and logout of FarmShare.

Control-B then D (releasing all keys after the Control-B before hitting D)
(detaches tmux)

(j2)jane@rice99:~$ exit

jane@rice99:~$ exit

jane@rice99:~$ exit (if exit does not work here, use logout)

Client connection

Connecting to Jupyter

This is the only step that needs to be performed every time (e.g., a client laptop wakes from sleep, connecting from a different laptop) once the installation is complete and the Jupyter server is running.

SSH into the same FarmShare system that the Jupyter server was started in above, tunneling the TCP port that the Jupyter server is listening on.

$ ssh jane@rice99.stanford.edu -L 9876:localhost:9876

substituting in the appropriate SUNet ID for jane, the appropriate FarmShare server hostname for rice99, and the appropriate TCP port for 9876.

Windows users can setup an SSH local tunnel using menu options in putty. Enter the appropriate TCP port in the Source port box and enter localhost:9876 for the Destination box and then click the Add button.

Once logged in, paste the https://localhost:9876/?token=... URL provided from the jupyter_start or jupyter notebook list commands into a web browser to connect to the Jupyter notebook. Ignore any browser security warnings.

Note: Jupyter works best with Firefox or Chrome browsers. Safari and Internet Explorer are not well supported.

Shut down Jupyter

It's best to shut down Jupyter to free resources for others when finished and will not be used for an extended period (e.g., greater than two weeks).

SSH into the same FarmShare system that the Jupyter server was started in.

$ ssh jane@rice99.stanford.edu

substituting in the appropriate SUNet ID for jane and the appropriate FarmShare server hostname for rice99.

Return to tmux window 0 where Jupyter was launched, shut down the server, and exit tmux.

rice99:~> tmux attach -d
Control-B then 0 (releasing all keys after the Control-B before hitting 0)
(returns to tmux window 0 with lots Jupyter notebook output)

[I 17:23:59.671 NotebookApp]
Control-C then Control-C (releasing all keys after the Control-C before hitting Control-C a second time)

This will shut down the Jupyter notebook. Log out of FarmShare.

Renewing virtual terminal

Renewing Jupyter graph

This only needs to be done when the Kerberos credentials expire after a week and the notebook no longer functions. To restore the notebook after this time, SSH into the same FarmShare system used to create the virtual terminal. Renewal can take place at any time after a week when use is desired.

$ ssh jane@rice99.stanford.edu

substituting SUNet ID for jane and the appropriate FarmShare server for rice99.

rice99:~> tmux attach -d

(j2)jane@rice99:~$ kinit -r 7d; aklog
Password for jane@stanford.edu:

(j2)jane@rice99:~$ keep_kerberos_afs
Run: export KRB5CCNAME=FILE:/tmp/.krb5_jane.tgt

(j2)jane@rice99:~$ export KRB5CCNAME=FILE:/tmp/.krb5_jane.tgt 

Confirm that a new Kerberos ticket and AFS token have been gathered:

 (j2)jane@rice99:~$ klist
 Ticket cache: FILE:/tmp/.krb5_jane.tgt
 Default principal: jane@stanford.edu
 
 Valid starting       Expires              Service principal
 11/11/1885 12:00:00  11/12/1885 11:00:00  krbtgt/stanford.edu@stanford.edu
         renew until 11/18/1885 12:00:00
 11/11/1885 12:00:00  11/12/1885 11:00:00  afs/ir.stanford.edu@stanford.edu
         renew until 11/18/1885 12:00:00
 
 (j2)jane@rice99:~$ tokens
 
 Tokens held by the Cache Manager:
 
 User's (AFS ID 99999) tokens for afs@ir.stanford.edu [Expires Nov 12 12:00]
    --End of list--

If the ticket cache does not contain the user's SUNetID but instead looks more like FILE:/tmp/krb5cc_99999_S5kuAtfyaX then the renewal was not successful and kinit -r 7d; aklog and keep_kerberos_afs commands need to be repeated.

Once successful, detach and logout of FarmShare:

Control-B then D (releasing all keys after the Control-B before hitting D)
(detaches tmux)

This process extends the terminal's credentials for another week. The same notebook URL can be used without interruption.

Examples

Python 3

Plotly

Plotly is a popular, open-source, interactive plotting framework available for common data analysis programming languages.

import numpy as np
import plotly.offline as py
import plotly.graph_objs as go
py.init_notebook_mode(connected=True)

x = np.r_[ -np.pi : np.pi : 0.1 ]
py.iplot([go.Scatter(x=x, y=np.sin(x)), go.Scatter(x=x, y=np.cos(x))])

Jupyter python demo.png

bqplot + ipywidgets

Useful for interactive interfaces with realtime plot updating.

import numpy as np
from bqplot import pyplot as bqplt
import ipywidgets

def update_phase(n):
    line1.y = np.sin(x - n)

bqplt.clear()
x = np.r_[ -np.pi : np.pi : 0.1]
line1 = bqplt.plot(x, np.sin(x))
line2 = bqplt.plot(x, np.cos(x), 'r')
bqplt.show()

ipywidgets.interact(update_phase, n=(-2*np.pi,2*np.pi))

Jupyter python bqplot demo.gif

MATLAB

farmshare_plotly_init;

x = -pi : 0.1 : pi;
hold on;
plot(x, sin(x));
plot(x, cos(x));

Jupyter matlab demo.png

Note: If Jupyter installation was performed prior to 2017/04, run the following script in an ssh terminal to configure plotly for matlab: plotly_create_config

R

library(plotly)

x  <- seq(-pi, pi, 0.1)
y1 <- sin(x)
y2 <- cos(x)

ds <- data.frame(x, y1, y2)

p  <- plot_ly(ds, x = ~x, y = ~y1, type = 'scatter', mode='lines') %>%
      add_trace(y = ~y2)

embed_notebook(p)

Jupyter r demo.png

Julia

using Plots
plotly()

plot([sin, cos], -pi, pi)

Jupyter julia demo.png

SAS

data curves;
  do x = -constant("pi") to constant("pi") by 0.1;
    y = sin(x);
    z = cos(x);
    output;
  end;
run;

symbol1 interpol=join color=blue width=5;
symbol2 interpol=join color=red  width=5;
axis1 label=none minor=none;

proc gplot data=curves;
  plot y*x=1 z*x=2 / overlay vaxis=axis1 haxis=axis1;
run;

Jupyter sas demo.png

Useful Linux commands

  • klist - Displays Kerberos ticket status, cache file location, and expiration date
  • tokens - Lists AFS token status and expiration date
  • kinit - obtains new Kerberos ticket using SUNetID
  • aklog - obtains AFS token from a valid Kerberos ticket

Troubleshooting

The Jupyter notebooks are great when they work, but can be confusing to fix when not working. Here are some general troubleshooting steps to help get things working again.

  • Confirm the rice machine being used
    • Jupyter will only work if the same rice machine used to launch Jupyter is also used to establish the client connection
    • Jupyter notebooks launched on a given riceXX machine are only accessible via that riceXX machine
  • Confirm tmux session has valid Kerberos tickets and AFS tokens
  • Confirm that Jupyter is running
      • In an open j2 tmux window, check that Jupyter is running via jupyter notebook list
  • Confirm that the SSH tunnel is in place
    • The SSH tunnel needs to be from the personal computer (e.g., laptop) to rice
    • Refresh the browser window with the Jupyter URL (https://localhost:XXXX)
    • Do not rely on an existing browser window running Jupyter to help with debugging, its behavior can be misleading

Custom Jupyter notebook

A custom virtual environment can be deployed and used instead of the default one used by these instructions if necessary. This could be useful for using other Jupyter kernels, additional python modules, or a different version of python.

Create virtual environment:

virtualenv -p python3 /farmshare/user_data/jane/venv
source /farmshare/user_data/jane/venv/bin/activate

This is installed in the /farmshare/user_data/ ZFS pool to preserve AFS user quota space, as virtual environments can get large (the default environment deployed here exceeds 1.5GB). These ZFS pools are accessible across all farmshare machines.

Note, this will create a virtual environment with python3. To use python2, specify -p python2 instead of -p python3.

From there, install the desired packages using pip (and update pip and all virtualenv packages to the latest versions):

pip install pip-review
pip-review -a

Most people performing numerical analyses, interacting with the web, or manipulating data with Jupyter will want to install the following common packages:

 pip install numpy scipy jupyter sympy pandas blaze matplotlib plotly seaborn statsmodels SQLAlchemy Pillow Requests lxml beautifulsoup4

The full list of packages installed in the virtual environment (generated by a pip freeze) used in these instructions can be found at /afs/ir.stanford.edu/group/bil/env/scripts/requirements.txt

Once installed, jupyter can be launched from the command line as usual. The jupyter_start script in /afs/ir.stanford.edu/group/bil/env/scripts/ will also work in most cases since it relies on the environment path to launch jupyter.

Technical Details

  • Jupyter setup is best performed via the Linux console. This guide is mostly step-by-step, but general familiarity with Linux is helpful.
  • SSH provides an encrypted remote shell into another system and is the primary way that Jupyter will be installed and accessed.
  • The first two commands of the installation (bash and source) switch the shell to bash and then update the environment to use a pre-built Jupyter installation. The prompt changes to have the (jupyter) prefix when the environment change is successful. SUNet IDs ceated after 2014 have their default shell set to bash and can omit the call to bash.
  • If of interest, the pip packages installed for the Jupyter environment are stored in a requirements.txt file at /afs/ir.stanford.edu/group/bil/env/jupyter/requirements.txt. Additional packages may be installed upon request.
  • The jupyter_config_wrapper command calls a script that will create encryption keys for the Jupyter notebook, configure the jupyter config file, and sets the default tmux shell to bash. Note, this script modifies existing Jupyter notebook config files (if they exist), but will not overwrite any parameters that have changed from their default values.
  • These instructions runs the notebook within a persistent virtual terminal so it will continue to run even after the user has logged out of FarmShare.
  • The jupyter_start script accepts an optional port argument, specifying which port to connect to (e.g., jupyter_start 9876).
  • Once the Jupyter notebook server is running, it is ready to accept client connections. For security, it only accepts connections from localhost (i.e., connections originating from that specific FarmShare system itself). Connections from other systems (e.g., a laptop) are created through SSH tunnels.
  • Since the encryption keys were self-signed, the browser will warn about an insecure connection, but this warning can be disregarded.
  • If this notebook URL is misplaced, tmux attach -d brings up the virtual terminal and the notebook status can be queried using the verification step on tmux window 1 using jupyter notebook list.
Personal tools
Toolbox
LANGUAGES