/probcomp-stack

MIT Probabilistic Computing Project software stack

Primary LanguageShell

probcomp-stack: MIT Probabilistic Computing Project software stack

We currently support two deployment mechanisms

  • An isolated EC2 instance on the group's AWS account
  • A local Docker container

EC2 model

  • Each user gets exactly one EC2 machine
  • Each user's machine is completely isolated from all the other users' machines
  • We pay for the machines
  • We have a *nix client workflow for administering the fleet
  • We can insist on each user having a fleet-wide unique username
  • We can let a given person have several usernames and thus use several independent machines, but we do not support cluster computing (in this iteration of the management software)

Deploying on and managing isolated EC2 instances

Prerequisites

  • Installed AWS command-line interface aws, and jq.

  • AWS credentials authorized to manipulate EC2 resources and use CloudFormation. For instance, the AmazonEC2FullAccess, AmazonRoute53FullAccess, and ProbcompCloudFormationFullAccess policies suffice.

  • SSH keypair the instances will trust, uploaded to AWS (e.g., bch20170503-ec2.pem, or you can make your own).

  • A DNS domain. Ours is probcomp.net (entered into the command line tool with the trailing .).

  • SSL certificate under the AWS certificate manager. Ours covers *.stack.probcomp.net.

Update the Ubuntu packages of our software

See the packaging repository.

Create a stack for a new user

Choose:

  • User name (e.g., their name)
  • Initial instance type (e.g., t2.micro [default] for testing, or c3.8xlarge for compute)
  • Initial password

Save the password in the file jupyter-passwords/<user>.passwd

Make sure the desired content (from workshop-materials, presumably) is uploaded as a compressed tar archive at https://probcomp-oreilly20170627.s3.amazonaws.com/content-package.tgz

Run

./stack.sh create <user> <instance>

If something goes wrong, the CloudFormation section of the AWS console is helpful.

Right now this script bakes in some assumptions specific to my (axch's) machine, such as the choice and location of the key pair to start the instance with. It may be necessary to make some adjustments to get it to work for you, or invoke the pieces separately.

Test it by browsing https://<user>.stack.probcomp.net

Check in the host key in the known_hosts directory, and a line about who it's for in the running-stacks.org file.

Create a fleet of many instances for some purpose

Choose

  • Base name for the fleet stacks
  • Range of indexes to set up
  • Instance type (e.g., c3.8xlarge)

Create passwords for all the instances, in files named jupyter-passwords/<base>-<i>.passwd

Make sure the desired content (from workshop-materials, presumably) is uploaded as a compressed tar archive at https://probcomp-oreilly20170627.s3.amazonaws.com/content-package.tgz

  • Perhaps this may call for (cd ../workshop-materials && aws s3 cp content-package.tgz s3://probcomp-oreilly20170627/content-package.tgz)

Run

./fleet.sh create <base-name> <low> <high> [<instance> [<ami-id>]]

Then run whichever of

./fleet.sh deb-upgrade <base-name> <low> <high>
./fleet.sh grab-content <base-name> <low> <high>
./fleet.sh marco-install <base-name> <low> <high>
./fleet.sh set-passwords <base-name> <low> <high>

are desired

Check a few by browsing https://<base>-<i>.stack.probcomp.net

Print physical cards with urls and passwords to hand out

Make sure the passwords are all present in files named jupyter-passwords/<user>.passwd.

Loop editing and rerunning the script url-cards.sh until the output is as desired.

Change the instance type for a user

  • Run ./stack.sh update <user> <new-instance>

Terminate a user's instance

Run aws cloudformation delete-stack --stack-name probcomp-stack-<user>

Stop or resatrt a user's instance

Log in to the EC2 console, find the instance tagged <user>/instance, and stop or restart it. Note that this requires an extra step to re-enable the ssh address below, per #38

SSH into the instance

  • The instance's ssh host key should be saved in known_hosts/<user>

  • login.sh <user> which runs ssh -i <private-key> -o UserKnownHostsFile=./known_hosts/<user> -o CheckHostIP=no -o StrictHostKeyChecking=yes ubuntu@ssh.<user>.stack.probcomp.net with the default key (namely, bch20170503-ec2.pem)

Transfer files to and from the instance

  • put.sh <user> <local-source> <remote-dest>
  • get.sh <user> <remote-source> <local-dest>

Change a user's Jupyter notebook password

  • SSH into the instance
  • activate the virtual environment venv
  • jupyter notebook password

Or

  • write the new password to jupyter-passwords/<user>.passwd
  • python write-jupyter-passwords.py <user>
  • ./set-jupyter-password.sh <user>

Restart a user's Jupyter notebook server, if needed

./restart-jupyter.sh

Visit the Jupyter notebook server

https://.stack.probcomp.net/

Collect usage logs

./fleet.sh fetch-logs <base-name> <low> <high>

Share usage logs with the lab

rsync -r -v logs/ probcomp-1.csail.mit.edu:/data/probcomp/ppaml/probcomp-stack-logs

Assess and explain current $ burn rate

./aws-spend.sh

Deploying locally in Docker

The stack can be bundled as a docker image, derived from components developed in separate Git repositories and distributed via Ubuntu .deb packages in https://github.com/probcomp/packaging.

Pick a directory where you want notebooks to go, say /path/to/notebooks -- perhaps under your home directory on your laptop, or perhaps the mount point of an Elastic Block Store on your Amazon EC2 instance.

Pick the name for an image, say probcomp/stack:20170503-3 for the third attempt on May 3rd 2017.

Run:

% docker build -t probcomp/stack:20170503-0 -f docker/ubuntu1604-jupyter .
% mkdir /path/to/notebooks
% docker run --rm --publish 127.0.0.1:8080:8080/tcp \
    -v /path/to/notebooks:/notebook \
    probcomp/stack:20170503-0

Then point a web browser at

http://127.0.0.1:8080/

(Beware: All other processes in the system with access to the TCP/IP stack, including those running with credentials other than yours, can reach your jupyter notebook.)

Dockerfiles available:

  • docker/ubuntu1604 Ubuntu 16.04 with system packages and probcomp stack

  • docker/ubuntu1604-jupyter Ubuntu 16.04 with system packages, probcomp stack, and pypi jupyter

The version of jupyter on pypi is newer than the version of ipython notebook in Ubuntu 16.04, but the level of QA and reliability on any particular base system such as Ubuntu 16.04 is unpredictable.

Updating Probcomp software

The current process for getting an instance with a fresh package set: