=== The Basics ===

The neo-containers system uses cloud-container technology to abstract and generalize container creation and initialization.  There is a per-experiment control node that centralizes configuration details. At the DETER level, the experiments have the control node and a number of "pnodes" which serve as hosts for the virtualized containers. The control node configures the pnodes as well as the containerized nodes as well as itself.  The bootstrap script is run on the control node and: 1) installs configuration software on itself, 2) updates its own configuration, 3) updates the configuration of the other physical nodes (which includes virtual networking and starting virtual machines), 4) give configuration information to virtual nodes and allows them to configuration themselves. 

=== HOWTO run neo-containers ===

1. The neo-containers are started via a single script run on the {{{config}}} node. The script is on the {{{configdb}}} branch of the (existing) containers source repository. The script is {{{./bin/container_bootstrap.sh}}}. Checkout the containers repository and switch to the correct branch:
{{{
users: > cd src
users: > git clone ssh://tardis.deterlab.net:/var/local/git/benito containers
users: > cd containers
users: > git checkout configdb
}}}

2. Create a containerized experiment with an NS file and the {{{/share/containers/containerize.py}}} script. The NS file should declare a {{{config}}} node as an embedded container that uses the {{{UB14-CHEF12}}} image. Use a snippet like so:
{{{
set config [$ns node]
tb-set-node-os $config UB14-CHEF12
tb-set-hardware $config container0
tb-add-node-attribute $config containers:node_type "embedded_pnode"
}}}
The {{{UB14-CHEF12}}} image is not required, but speeds things up a bit. 
For each container in the experiment, specify  {{{image_os}}}, {{{image_type}}}, {{{mage_name}}}, and optionally an {{{image_url}}} via the {{{tb-add-node-attribute}}} syntax. Details on each attribute is given below.

* {{{image_os}}} - This is really just to distinguish Windows from non-Windows nodes. If the {{{image_os}}} starts with "{{{windows}}}", the image will be treated as a Windows node. Otherwise it'll be assumed to be some sort of Unix-y container.
* {{{image_type}}} - This setting describes the containerization tech of the node. Currently this is *always* set to "{{{vagrant}}}" as Vagrant is the only package used to spin up the containers.
* {{{image_name}}} - The name of the image. The name identifies an OS, virtualization platform, and version. This name is based on the Vagrant naming scheme. The standard, "baked in" image names are:
  * chef/centos-7.0      (virtualbox, 1.0.0)
  * chef/freebsd-10.0    (virtualbox, 1.0.0)
  * fgrehm/precise64-lxc (lxc, 1.2.0)
  * fgrehm/trusty64-lxc  (lxc, 1.2.0)
  * hashicorp/precise64  (virtualbox, 1.1.0)
  * liibvert/centos64     (libvirt, 0)
  * ubuntu/precise64     (virtualbox, 12.04.4)
  * ubuntu/trusty64      (virtualbox, 14.04)
So to create an Ubuntu 14.04 64 bit container, the name would be "{{{ubuntu/trusty64}}}". The only fully tested image is "ubuntu/trusty64". We are working on the LXC nodes. Each container that has the same name, will use a copy of the same container image. 
* {{{image_url}}} - If the image name is not one of the baked in images above, a URL must be specified from which the neo-containers system can download the container image. This URL must be resolvable from the {{{config}}} experiment node. There is an existing tested Windows 7 image at {{{http://scratch/benito/deter_win7.box}}}. The image will only be downloaded once as long as the {{{image_name}}}s are the same for the containers.

Here is an example Windows and Ubuntu 14.04 container:
{{{
set r2d2 [$ns node]
tb-add-node-attribute $r2d2 containers:image_os windows
tb-add-node-attribute $r2d2 containers:image_type vagrant
tb-add-node-attribute $r2d2 containers:image_name deter/win7
tb-add-node-attribute $r2d2 containers:image_url http://scratch/benito/deter_win7.box

set c3po [$ns node]
tb-add-node-attribute $c3po containers:image_os ubuntu
tb-add-node-attribute $c3po containers:image_type vagrant
tb-add-node-attribute $c3po containers:image_name ubuntu/trusty64
}}}

3. Use the NS file to create a containerized experiment using the existing containers scripts (on users): {{{/share/containers/containerize.py [group] [experiment] [ns file]}}}
4. Modify the NS file generated by {{{containerize.py}}} to have a new image for the pnode machines. Navigate to the new experiment page and click {{{Modify Experiment}}}. Change the OS type of the pnodes to {{{PNODE_BASE}}} and the hardware type to {{{MicroCloud}}}. I.e. for each pnode in the NS file, make the lines have the form:
{{{
    tb-set-node-os ${pnode(0000)} PNODE-BASE
    tb-set-hardware ${pnode(0000)} MicroCloud
}}}
Add the bootstrap script to the {{{config}}} node, so it's execute after boot. Add the following to the {{{config}}} node stanza in the NS file:
{{{
tb-set-node-startcmd $config "~/src/containers/bin/container_bootstrap.sh"
}}}
(Of course have it match the path to your copy of the containers repository.) See the experiment Deter/swtest3 for a working example.
5. Swap in the experiment and wait a long time. There is no great way to tell when the containers are up, unfortunately. 


=== Execution Flow ===

The execution flow of is as follows. The initialization uses the existing containers system as a bootstrap. 

1. Create a containerized experiment with an NS file and the {{{/share/containers/containerize.py}}} script. The NS file should declare a {{{config}}} node as an embedded container that uses the {{{UB14-CHEF12}}} image. 

2. Modify the generated NS file.
   a. Change the OS type of the pnodes to {{{PNODE-BASE}}}. e.g. make the line in the NS file:
{{{
    tb-set-node-os ${pnode(0000)} PNODE-BASE
}}}
   b.  Add a new control node. Traditionally it's been called "config", but there is no restriction on the name. Add this to the NS file:
{{{
    set config [$ns node]
    tb-set-node-os ${config} Ubuntu1404-64-STD
    tb-set-hardware ${config} MicroCloud
    tb-set-node-failure-action ${config} "nonfatal"
}}}
   c. Remove the {{{tb-node-set-startcmd}}} line. This starts the existing DETER containers system on the nodes. We do not want that. 
3. Swap in the experiment.
4. Run the bootstrap script on your {{{config}}} node.
   a. The bootstrap script is called {{{container_bootstrap.sh}}} and it lives in the container source code on the {{{configdb}}} branch at {{{./bin/container_bootstrap.sh}}}. So clone the containers repository, switch to the branch {{{configdb}}}, and run the script.

The bootstrap script:
1. Uses the NFS mounted dir /share/chef/chef-packages to install Chef server and Chef client on the {{{config}}} node.
    a. The {{{config}}} node is both a Chef server (talks to all machines in the experiment) and a Chef client/workstation (holds the Chef git repo which contains all configuration scripts). 
2. Configures Chef Server on the {{{config}}} node. 
    a. Simply runs the chef self-configure scripts.
3. Configures Chef client/workstation on the {{{config}}} node.
    a. creates a chef-deter identity and organization (keys and names) for the Chef Server. 
    b. installs {{{git}}} on the {{{config}}} node.
    c. git clones the canonical Chef repository from the NFS mounted dir {{{/share/chef/chef-repo}}}. Thie repo contains all configuration scripts and data for the system.
    d. upload all Chef recipes, data bags, and roles (configuration scripts, data, and roles) to the Chef server on {{{config}}} (localhost). 
4. Create an experiment-specific "data bag" and upload it to the Chef Server. The data contains experiment name, project, group, name and address of the {{{config}}} machine,  and the name and address of the RESTful configuration server. This is per-experiment dynamic data and thus is generated at run time instead of statically kept in the chef repository.
5. Register the {{{config}}} node (localhost) as a Chef client to the server, download and execute all local configuration recipes that exist in the {{config_server}} chef role. (A chef role is simply a collection of recipes.) Details of these are in the next section. 

The chef repository defines three roles in the system, {{{config_server}}}, {{{pnode}}}, and {{{container}}}. The code that defines the roles can be found in the {{{./roles}}} directory in the chef repository.

=== {{{config_server}}} recipes ===

Recipes are chunks of Ruby code that is are a mixture of Ruby and the Chef configuration definition language. A recipe is downloaded from a Chef server and run when {{{chef-client}}} is run on a node. This section describes the recipes that are run on the {{{config_server}}} machine itself. These recipes configure Chef communications on the nodes and tell the {{{pnodes}}} to configure themselves. (The {{{pnode}}} role recipe section is next.)

These recipes can be found in the {{{./cookbooks/config_server/recipes}}} directory in the chef repository.

The {{{config_server}}} role consists of the following recipes, executed in order:  {{{config_db}}}, {{{config_server}}}, {{{hosts}}}, {{{bootstrap_pnodes}}}, and {{{configure_containers}}}.

* {{{config_db}}} - This recipe builds the configuration database from existing containers and DETER/emulab files found on the local machine. This database contains all the configuration information to configure the experiment. This database is RESTfully served by the {{{config_server}}} process started elsewhere. The recipe also installs required packages needed to build the database (python3 and SQLAlchemy). 

* {{{config_server}}} - This recipe installs and starts the {{{config_server}}} process. This server RESTfully serves configuration data to anyone in the experiment that requests it. After installing required packages, the recipe simply calls the standard {{{setup.py}}} script in the config_server package to install it. It then copies a local {{{/etc/init}}} style script from the chef repository to {{{/etc/init.d/config_server}}}, then invokes the it, starting the daemon. At this point configuration information requests can be made on {{{http://config:5000/...}}} to get configuration information. 

* {{{hosts}}} - This recipe (which actually lives in the {{{pnode}}} cookbook directory, adds container node names and address information to {{{/etc/hosts}}}. This is a required step for later when the {{{config_server}}} node connects to the containers.

* {{{bootstrap_pnodes}}} - This recipe connects to all pnode machines and configures them as chef clients to the new chef server. It then sets the default role for the pnode machines to the {{{pnode}}} role. Finally it connects to all the pnode machines and invokes {{{chef-client}}} locally. This kicks off the next round of configuration in the system - all pnode machines configure themselves. These recipes are described in the next section.

* {{{configure_containers}}} - This recipe simply uses the chef tool {{{knife}}} to connect to all (now running) containers and executes {{{chef-client}}}, which causes the containers to reach out to the chef server and request recipes/roles to execute. After this the containers are fully configured and integrated into DETER. 

=== {{{pnode}}} recipes ===

The {{{pnode}}} role consists of the following recipes, executed in order: {{{uml-net-group}}}, {{{mount_space}}}, {{{hosts}}}, {{{diod}}}, {{{vde}}}, and {{{vagrant}}}.

* {{{uml-net-group}}} - This recipe creates a {{{uml-net}}} group. This is just a stop-gap recipe which is needed due to a broken apt database in the PNODE-BASE image. This recipe will be removed from the role and the chef repository once the PNODE-BASE image is rebuilt. 

* {{{mount_space}}} - This recipe formats and mounts the {{{/dev/sda4}}} partition on the disk. This is subsequently used to store and run containers images, the chef repository, etc.

* {{{hosts}}} - This recipe creates entries in the pnode's {{{/etc/hosts}}} for all container nodes. 

* {{{diod}}} - This recipe installs, configures, and starts the {{{diod}}} daemon. The pnode has users home directories and other directories mounted locally. The containers also need these mounted to be fully integrated with DETER. Unfortunately it is not possible to use standard NFS mounting on directories that are themselves NFS mounted. (There are good reasons for this.) So the pnode uses {{{diod}}} to serve these directories to the container nodes. The recipe uses standard package tech to install {{{diod}}} and configures diod to export all directories that are locally mounted - which is just what we want. It then starts the daemon. 

* {{{vde}}} - This recipe installs and configures {{{VDE}}}, Virtual Distributed Ethernet on the pnode. It installs VDE2 and required packages (uml-utilities and bridge-utils). It installs a /etc/init style script to {{{/etc/init.d/vde_switch}}} to control the VDE daemon. It requests all virtual networking configuration information from the {{{config_server}}} and creates a number of VDE configuration files in {{{/etc/vde2/conf.d}}}, one for each switch needed. The recipe creates and configures all the TAP devices and bridges needed by the VDE switches. It then calls the {{{/etc/init.d/vde2}}} script to start all switches. At this point all local network plumbing is there and working. 

* {{{vagrant}}} - This recipe installs and configures {{{vagrant}}}. Vagrant is a front-end for configuring and spawning virtual machines. it supports many different virtual machine image (qemu, LXE, VirtualBox, etc), but the current neo-containers system only uses VirtualBox images. (It has a few default images already configured and installed on the PNODE-BASE DETER image.) The recipe installs {{{vagrant}}} and {{{virtualbox}}} (although they are already installed on the PNODE-BASE image, so this is really a NOOP command. After that, it's big job is to create the {{{Vagrantfile}}} file in {{{/space/vagrant}}} that describes the container images and basic configuration. The {{{Vagrantfile}}} created does a few basic things: configure virtual NICs, setup basic networking so the containers route correctly to the DETER control net, and invoke a basic NOOP chef configuration that *only* register the container node with the chef server. It does not invoke {{{chef-client}}} on the containers to configure the nodes, that happens next. The recipe also sets up a few things on the pnode for smoother vagrant operations: a {{{vagrant}}} user and home directory at {{{/vagrant}}} are created, an {{{/etc/profile.d/vagrant.sh}}} file is created with vagrant-specific environment variables, Virtual Box is configured to not store disk image copies on NFS mounted dirs, but in {{{/space/vagrant}}}. Finally {{{/etc/init.d/vagrant}}} is invoked to spawn the containers. At this point, the containers are running, but not configured/integrated into DETER (no user accounts, mounts, etc). 

=== {{{container}}} recipes ===

The {{{container}}} role consists of the following recipes: {{{apt}}}, {{{hosts}}}, {{{groups}}}, {{{accounts}}}, {{{mounts}}}. All these recipes work to integrate the container into the running DETER experiment.

* {{{apt}}} - Runs {{{apt clean}}} and {{{apt update}}} on the nodes. It would be nice to get rid of this recipe as it will slam the DETER apt repository server. Need to investigate why and if this is still needed. 

* {{{hosts}}} - Adds container names and addresses to {{{/etc/hosts}}} so all containers can resolve each others' names. 

* {{{groups}}} - Requests group information from the {{{config_server}}} and creates the DETER groups on the container.

* {{{accounts}}} - Requests user account information from the {{{config_server}}} and recreates the accounts locally on the container. This includes creating mount points for the user home directories. 

* {{{mounts}}} - This recipe installs the client side {{{diod}}} package and configures it to talk to the local pnode on which the container runs. It requests mount information from the {{{config_server}}} and mounts what it finds via {{{diod}}}. 


=== Outstanding Issues ===

This section lists outstanding issues for the container system. And not "outstanding" as in "awesome", but outstanding as in this still has to get done.

* {{{diod}}} - The diod daemon (re)mounts NFS mounted filesystems on the containers. The pnode acts is a server, the containers are clients.
  * {{{diod}}} fails to fully mount filesystems sometimes during the application of the mounts recipe. It fails in {{{mount -a}}}. It fails to allocate memory. This only seems to happen when running in chef. Running {{{sudo mount -a}}} on the command line of the container works just fine. Which is a little annoying, but so it goes.
* Windows - is fragile and doesnt' really work. Working on this now.
* LXC - Does not come up. Looked into this briefly. Tried booting outside of vagrant and still had issues. See here for a getting started doc: https://linuxcontainers.org/lxc/getting-started/
* Using {{{~glawler}}} - some scripts are still run (via recipes) out of {{{~glawler}}}. They should be fed a path to a {{{containers}}} repo or just assume {{{/share/containers}}}. 
* Vagrant
  * Does do some things in parallel, but not bringing up virtualbox machines. This makes things very slow.
  * chef provisioning via vagrant does not work well. It registers the node with the chef server, but does not properly setup the node for future runs. Does not create proper {{{client.rb}}} config files.
* everything else.