#Installing a Physical POD

The following is a detailed, step-by-step recipe for installing a physical POD.

NOTE: If you are new to CORD and would like to get familiar with it, you should start by bringing up a development POD on a single physical server, called CORD-in-a-Box.

NOTE: Also see the Quick Start: Physical POD Guide for a streamlined overview of the physical POD install process.

##Terminology

This guide uses the following terminology.

  • POD: A single physical deployment of CORD.

  • Full POD: A typical configuration, and is used as example in this Guide. A full CORD POD is composed by three servers, and four fabric switches. It makes it possibile to experiment with all the core features of CORD and it is what the community uses for tests.

  • Half POD: A minimum-sized configuration. It is similar to a full POD, but with less hardware. It consists of two servers (one head node and one compute node), and one fabric switch. It does not allow experimentation with all of the core features that CORD offers (e.g., a switching fabric), but it is still good for basic experimentation and testing.

  • Development (Dev) / Management Node: This is the machine used to download, build and deploy CORD onto a POD. Sometimes it is a dedicated server, and sometime the developer's laptop. In principle, it can be any machine that satisfies the hardware and software requirements reported below.

  • Development (Dev) VM: Bootstrapping the CORD installation requires a lot of software to be installed and some non-trivial configurations to be applied. All this should happens on the dev node. To help users with the process, CORD provides an easy way to create a VM on the dev node with all the required software and configurations in place.

  • Head Node: One of the servers in a POD that runs management services for the POD. This includes XOS (the orchestrator), two instances of ONOS (the SDN controller, one to control the underlay fabric, one to control the overlay), MaaS and all the services needed to automatically install and configure the rest of the POD devices.

  • Compute Node(s): A server in a POD that run VMs or containers associated with one or more tenant services. This terminology is borrowed from OpenStack.

  • Fabric Switch: A switch in a POD that interconnects other switch and server elements inside the POD.

  • vSG: The virtual Subscriber Gateway (vSG) is the CORD counterpart for existing CPEs. It implements a bundle of subscriber-selected functions, such as Restricted Access, Parental Control, Bandwidth Metering, Access Diagnostics and Firewall. These functionalities run on commodity hardware located in the Central Office rather than on the customer’s premises. There is still a device in the home (which we still refer to as the CPE), but it has been reduced to a bare-metal switch.

Overview of a CORD POD

The following is a brief description of a generic full POD.

###Physical Configuration

A full POD includes a ToR management switch, four fabric switches, and three standard x86 servers. The following figure does not show access devices or any upstream connectivity to the metro network; those details are included later in this section.

###Logical Configuration: Data Plane Network

The following diagram is a high level logical representation of a typical CORD POD.

The figure shows 40G data plane connections (red), where end-user traffic goes from the access devices to the metro network (green). User traffic goes through different different leafs, spines and compute nodes, depending on the services needed, and where they are located. The switches form a leaf and spine fabric. The compute nodes and the head node are connected to a port of one of the leaf switches.

###Logical Configuration: Control Plane / Management Network

The following diagram shows in blue how the components of the system are connected through the management network.

As shown in this figure, the head node is the only server in the POD connected both to Internet and to the other components of the system. The compute nodes and the switches are only connected to the head node, which provides them with all the software needed.

##Sample Workflow

It is important to have a general picture of installation workflow before getting into the details. The following is a list of high-level tasks involved in bringing up a CORD POD:

  • CORD software is downloaded and built on the dev machine.
  • A POD configuration is created by the operator on the dev machine.
  • The software is pushed from the dev machine to the head node.
  • Compute nodes and fabric switches need to be manually rebooted. The CORD build procedure automatically installs the OS, other software needed and performs the related configurations.
  • The software gets automatically deployed from the head node to the compute nodes.

##Requirements

While the CORD project is for openness and does not have any interest in sponsoring specific vendors, it provides a reference implementation for both hardware and software to help users in building their PODs. What is reported below is a list of hardware that, in the community experience, has worked well.

Also note that the CORD community will be better able to help you debugging issues if your hardware and software configuration look as much as possible similar to the ones reported in the reference implementation, below.

##Bill Of Materials (BOM) / Hardware Requirements

The section provides a list of hardware required to build a full CORD POD.

###BOM Summary

QuantityCategoryBrandModelPart Num
3ComputeQuanta (QCT)QuantaGrid D51B-1UQCT-D51B-1U
4Fabric SwitchEdgeCoreAS6712-32XAS6712-32X
1Management Switch (L2 VLAN support)***
7Cabling (data plane)RobofiberQSFP-40G-03CQSFP-40G-03C
12Cabling (Mgmt)CAT6 copper cables 3M)**

###Detailed Requirements

  • 1x Development Machine. It can be either a physical machine or a virtual machine, as long as the VM supports nested virtualization. It doesn’t have to be necessarily Linux (used in the rest of the guide, below); in principle anything able to satisfy the hardware and the software requirements. Generic hardware requirements are 2 cores, 4G of memory, 60G of hdd.

  • 3x Physical Servers: one to be used as head node, two to be used as compute nodes.

    • Suggested Model: OCP-qualified QuantaGrid D51B-1U server. Each server is configured with 2x Intel E5-2630 v4 10C 2.2GHz 85W, 64GB of RAM 2133MHz DDR4, 2x hdd500GB and a 40 Gig adapter.

    • Strongly Suggested NIC:

      • Intel Ethernet Converged Network Adapters XL710 10/40 GbE PCIe 3.0, x8 Dual port.
      • ConnectX®-3 EN Single/Dual-Port 10/40/56GbE Adapters w/ PCI Express 3.0.

      NOTE: while the machines mentioned above are generic standard x86 servers, and can be potentially substituted with any other machine, it’s quite important to stick with either one of the network card suggested. CORD scripts will look for either an i40e or a mlx4_en driver, used by the two cards cards. To use other cards additional operations will need to be done. Please, see the Network Settings appendix for more information.

  • 4x Fabric Switches

    • Suggested Model: OCP-qualified Accton 6712 switch. Each switch is configured with 32x40GE ports; produced by EdgeCore and HP.
  • 7x Fiber Cables with QSFP+ (Intel compatible) or 7 DAC QSFP+ (Intel compatible) cables

    • Suggested Model: Robofiber QSFP-40G-03C QSFP+ 40G direct attach passive copper cable, 3m length - S/N: QSFP-40G-03C.
  • 1x 1G L2 copper management switch supporting VLANs or 2x 1G L2 copper management switches

##Connectivity Requirements

The dev machine and the head node have to download software from different Internet sources, so they currently need unfettered Internet access. (In the future, only the dev machine, and not the head node, will require Internet connectivity.) Sometimes firewalls, proxies, and software that prevents to access local DNSs generate issues and should be avoided.

##Cabling a POD

This section describes how the hardware components should be interconnected to form a fully functional CORD POD.

Management / Control Plane Network

The management network is divided in two broadcast domains: one connecting the POD to the Internet and giving access to the deployer (called “external” and shown in green in the figure below), and one connecting the servers and switches inside the POD (called “internal” or “management” and shown in blue). The figure also shows data plane connections in red (as described in the next paragraph).

The external and the management networks can be separated either using two different switches, or the same physical switch and by using VLANs.

NOTE: Head node IPMI connectivity is optional.

NOTE: IPMI ports do not have to be necessarily connected to the external network. The requirement is that compute node IPMI interfaces need to be reachable from the head node. This is possible also through the internal / management network.

NOTE: Vendors often allow a shared management port to provide IPMI functionalities. One of the NICs used for system management (e.g., eth0) can be shared, to be used at the same time also as IPMI port.

####External Network

The external network allows POD servers to be reached from the Internet. This would likely not be supported in a production system, but is useful in development and evaluation settings, for example, making it easy to directly start/stop/reboot the head and the compute nodes. Moreover, using CORD automated scripts and tools for Jenkins pipeline requires Jenkins direct access to these interfaces. This is why IPMI/BMC interfaces of the nodes are also connected to the external network. In summary, following is the list of equipment/interfaces usually connected to the external network:

  • Internet
  • Dev machine
  • Head node - 1x 1G interface (following defined as external)
  • Head node - 1x IPMI/BMC interface (optional)
  • Compute node 1 - 1x IPMI/BMC interface (optional, but recommended)
  • Compute node 2 - 1x IPMI/BMC interface (optional, but recommended)

####Internal Network

The internal/management network is separate from the external one. It has the goal to connect the head node to the rest of the system components (compute nodes and fabric switches). For a typical POD, the internal network includes:

  • Head node - 1x 1G interface (following defined as management)
  • Compute node 1 - 1x 1G interface
  • Compute node 2 - 1x 1G interface
  • Fabric 1 - management interface
  • Fabric 2 - management interface
  • Fabric 3 - management interface
  • Fabric 4 - management interface

###User / Data Plane Network

The data plane network (represented in red in the figure) carries user traffic (in green), from the access devices to the point the POD connects to the metro network.

The fabric switches are assembled to form a leaf and spine topology. A typical full POD has two leafs and two spines. Currently, this is a pure 40G network. While spines are not connected together, each leaf is connected to both spines. In summary, the following are the devices connecting to the leaf switches:

  • Head node - 1x 40G interface
  • Compute node 1 - 1x 40G interface
  • Compute node 2 - 1x 40G interface
  • Access devices - 1 or more 40G interfaces *Metro devices - 1 or more 40G interfaces

###Best Practices

The community follows a set of best practices to better be able to remotely debug issues, for example via mailing-lists. The following is not mandatory, but is strongly suggested:

  • Leaf nodes are connected to the spines nodes starting at the highest port number on the leaf.

  • For a given leaf node, its connections to the spine nodes terminate on the same port number on each spine.

  • Leaf n connections to spine nodes terminate at port n on each spine node.

  • Leaf-spine switches are connected into the management TOR starting from the highest port number.

  • Compute node n connects to the internal (management) network switch on port n.

  • Compute node n connects to its leaf at port n.

  • The head node connects to the internal (management) network using the lowest 1G management interface.

  • The head node connects to the external network using its highest 1G management interface.

  • All servers connect to the leafs using the lowest fabric (40G NIC) interface.

Software Environment Requirements

Only the dev machine and the head node need to be prepped for installation. The other machines will be fully provisioned by CORD itself.

###Development Machine

It should run Ubuntu 16.04 LTS (suggested) or Ubuntu 14.04 LTS. Then install and configure the following software.

####Install Basic Packages

sudo apt-get -y install git python

####Install repo

curl https://storage.googleapis.com/git-repo-downloads/repo > ~/repo &&
sudo chmod a+x repo &&
sudo cp repo /usr/bin

####Configure git

Using the email address registered on Gerrit:

git config --global user.email "you@example.com"
git config --global user.name "Your Name"

####Virtualbox and Vagrant

sudo apt-get install virtualbox vagrant

NOTE: Make sure the version of Vagrant that gets installed is >=1.8 (can be checked using vagrant --version)

###Head Node

It should run Ubuntu 14.04 LTS. Then install and configure the following software.

####Install Basic Packages

sudo apt-get -y install curl jq

####Install Oracle Java8

sudo apt-get install software-properties-common -y &&
sudo add-apt-repository ppa:webupd8team/java -y &&
sudo apt-get update &&
echo "oracle-java8-installer shared/accepted-oracle-license-v1-1 select true" | sudo debconf-set-selections &&
sudo apt-get install oracle-java8-installer oracle-java8-set-default -y

####Create a User with "sudoer" Permissions (no password)

sudo adduser cord &&
sudo adduser cord sudo &&
sudo echo 'cord ALL=(ALL) NOPASSWD:ALL' >> /etc/sudoers.d/90-cloud-init-users

####Copy Your Dev Node ssh Public-Key

On the head node:

ssh-keygen -t rsa &&
mkdir /home/cord/.ssh/authorized_keys &&
chmod 700 /home/cord/.ssh &&
chmod 600 /home/cord/.ssh/authorized_keys

From the dev node:

cat ~/.ssh/id_rsa.pub | ssh cord@{head_node_ip} 'cat >> ~/.ssh/authorized_keys'

###Compute Nodes

The CORD build process installs the compute nodes. You only need to configure their BIOS settings so they can PXE boot from the head node through the internal ( management) network. In doing this, make sure:

  • The network card connected to the internal / management network is configured with DHCP (no static IPs).

  • The IPMI (sometime called BMC) interface is configured with a statically assigned IP, reachable from the head node. It’s strongly suggested to have them deterministically assigned, so you will be able to control your node as you like.

  • Their boot sequence has (a) the network card connected to the internal / management network as the first boot device; and (b) the primary hard drive as second boot device.

NOTE: Some users prefer to connect as well the IPMI interfaces of the compute nodes to the external network, so they can have control on them also from outside the POD. This way the head node will be able to control them anyway.

###Fabric Switches: ONIE

The ONIE installer should be already installed on the switch and set to boot in installation mode. This is usually the default for new switches sold without an Operating System. It might not be the case instead if switches have already an Operating System installed. In this case rebooting the switch in ONIE installation mode depends by different factors, such the version of the OS installed and the specific model of the switch.

###Download Software onto the Dev Machine

From the home directory, use repo to clone the CORD repository:

mkdir cord && cd cord &&
repo init -u https://gerrit.opencord.org/manifest -b master &&
repo sync

NOTE: master is used as example. You can substitute it with your favorite branch, for example cord-2.0 or cord-3.0. You can also use a "flavor" specific manifests such as “mcord” or “ecord”. The flavor you use here is not correlated to the profile you will choose to run later but it is suggested that you use the corresponding manifest for the deployment you want. AN example is to use the “ecord” profile and then deploy the ecord.yml service_profile.

When this is complete, a listing (ls) inside this directory should yield output similar to:

ls -F
build/         incubator/     onos-apps/     orchestration/ test/

###Build the Dev VM

Instead of installing the prerequisiste software by hand on the dev machine, the build environment leverages Vagrant to spawn a VM with the tools required to build and deploy CORD. To create the development machine the following Vagrant command can be used:

cd ~/cord/build
vagrant up corddev

This will create an Ubuntu 14.04 LTS virtual machine and will install some required packages, such as Docker, Docker Compose, and Oracle Java 8.

WARNING: Make sure the VM can obtain sufficient resources. It may takes several minutes for the first command vagrant up corddev to complete, as it will include creating the VM, as well as downloading and installing various software packages. Once the Vagrant VM is created and provisioned, you will see output ending with:

==> corddev: PLAY RECAP *********************************************************************
==> corddev: localhost                  : ok=29   changed=25   unreachable=0    failed=0

The important thing is that the unreachable and failed counts are both zero.

NOTE: From the moment the VM gets created, it shares a folder with the OS below (the one of the server or of your personal computer). This means that what was the installation root directory (~/cord), will be also available in the VM under /cord.

###Log into the Dev VM

From the build directory, run the following command to connect to the development VM created

vagrant ssh corddev

Once inside the VM, you can find the deployment artifacts in the /cord directory.

In the VM, change to the /cord/build directory before continuing.

cd /cord/build

###Fetch Docker Images

The fetching phase of the build process pulls Docker images from the public repository down to the VM, and clones the git submodules that are part of the project. This phase can be initiated with the following command:

./gradlew fetch

NOTE: The first time you run ./gradlew it will download the gradle binary from the Internet and installs it locally. This is a one time operation, but may be time consuming, depending on the speed of your Internet connection.

WARNING: It is unfortunately fairly common to see this command fail due to network timeouts. If theis happens, be patient and run again the command.

Once the fetch command has successfully run, the step is complete. After the command completes you should be able to see the Docker images that were downloaded using the docker images command on the development machine:

docker images
REPOSITORY                  TAG                 IMAGE ID            CREATED             SIZE
opencord/onos               <none>              e1ade494f06e        3 days ago          936.5 MB
python                      2.7-alpine          c80455665c57        2 weeks ago         71.46 MB
xosproject/xos-base         <none>              2b791db4def0        4 weeks ago         756.4 MB
redis                       <none>              74b99a81add5        11 weeks ago        182.8 MB
xosproject/xos-postgres     <none>              95312a611414        11 weeks ago        393.8 MB
xosproject/cord-app-build   <none>              003a1c20e34a        5 months ago        1.108 GB
consul                      <none>              62f109a3299c        6 months ago        41.05 MB
swarm                       <none>              47dc182ea74b        8 months ago        19.32 MB
nginx                       <none>              3c69047c6034        8 months ago        182.7 MB
xosproject/vsg              <none>              dd026689aff3        9 months ago        336 MB

###Build Docker Images

Bare metal provisioning leverages utilities built and packaged as Docker container images. The images can be built by using the following command.

./gradlew buildImages

Once the buildImages command successfully runs the task is complete. The CORD artifacts have been built and the Docker images can be viewed by using the docker images command on the dev VM:

docker images --format 'table {{.Repository}}\t{{.Tag}}\t{{.Size}}\t{{.ID}}'
REPOSITORY                  TAG                 SIZE                IMAGE ID
opencord/mavenrepo          latest              338.2 MB            2e29009df740
cord-maas-switchq           latest              337.7 MB            73b084b48796
cord-provisioner            latest              822.4 MB            bd26a7001dd8
cord-dhcp-harvester         latest              346.8 MB            d3cfa30cf38c
config-generator            latest              278.4 MB            e58059b1afb2
cord-maas-bootstrap         latest              359.4 MB            c70c437c6039
cord-maas-automation        latest              371.8 MB            9757ac34e7f6
cord-ip-allocator           latest              276.5 MB            0f399f8389aa
opencord/onos               <none>              936.5 MB            e1ade494f06e
python                      2.7-alpine          71.46 MB            c80455665c57
golang                      alpine              240.5 MB            00371bbb49d5
golang                      1.6-alpine          283 MB              1ea38172de32
nginx                       latest              181.6 MB            01f818af747d
xosproject/xos-base         <none>              756.4 MB            2b791db4def0
ubuntu                      14.04               187.9 MB            3f755ca42730
redis                       <none>              182.8 MB            74b99a81add5
xosproject/xos-postgres     <none>              393.8 MB            95312a611414
xosproject/cord-app-build   <none>              1.108 GB            003a1c20e34a
consul                      <none>              41.05 MB            62f109a3299c
swarm                       <none>              19.32 MB            47dc182ea74b
nginx                       <none>              182.7 MB            3c69047c6034
xosproject/vsg              <none>              336 MB              dd026689aff3

NOTE: not all the docker machines listed are created by the CORD project but are instead used as a base to create other images.

Prepare POD Configuration File

Each CORD POD deployment requires a POD configuration file that describes how the system should be configured, including what IP addresses should be used for the external and the internal networks, what users the system should run during the automated installation, and much more.

POD configuration files are YAML files with extension .yml, contained in the /cord/build/config directory in the dev VM. You can either create a new file with your favorite editor or copy-and-edit an existing file. The sample.yml configuration file is there for this purpose. All parameters have a descriptions. Optional lines have been commented out, but can be used in case as needed.

More information about how the network configuration for the POD can be customized can be found in an Appendix: POD Network Settings.

##Publish Docker Images to the Head Node

Publishing consists of pushing the build docker images to the Docker repository on the target head node. This step can take a while as it has to transfer all the image from the development machine to the target head node. This step is started with the following command:

./gradlew -PdeployConfig=config/podX.yml publish

Once the publish command successfully runs this task is complete. When this step is complete, a Docker registry has been created on the head node and the images built on the dev node have been published to the head node registry.

WARNING: This command sometimes fails for various reasons. Simply rerunning the command often solves the problem.

Verify that the containers are running, using the docker ps command on the head node.

docker ps --format 'table {{.ID}}\t{{.Image}}\t{{.Command}}\t{{.CreatedAt}}'
CONTAINER ID        IMAGE               COMMAND                  CREATED AT
c8dd48fc9d18        registry:2.4.0      "/bin/registry serve "   2016-12-02 11:49:12 -0800 PST
e983d2e43760        registry:2.4.0      "/bin/registry serve "   2016-12-02 11:49:12 -0800 PST

Alternatively, the docker registry can be queried from any node that has access to the head node. You should be able to observe a list of docker images. Output may vary from deployment to deployment. The following is an example from an R-CORD deployment:

curl -sS http://head-node-ip-address:5000/v2/_catalog | jq .
{
  "repositories": [
    "config-generator",
    "consul",
    "cord-dhcp-harvester",
    "cord-ip-allocator",
    "cord-maas-automation",
    "cord-maas-switchq",
    "cord-provisioner",
    "gliderlabs/consul-server",
    "gliderlabs/registrator",
    "mavenrepo",
    "nginx",
    "node",
    "onosproject/onos",
    "redis",
    "swarm",
    "xosproject/chameleon",
    "xosproject/exampleservice-synchronizer",
    "xosproject/fabric-synchronizer",
    "xosproject/gui-extension-rcord",
    "xosproject/gui-extension-vtr",
    "xosproject/onos-synchronizer",
    "xosproject/openstack-synchronizer",
    "xosproject/vrouter-synchronizer",
    "xosproject/vsg",
    "xosproject/vsg-synchronizer",
    "xosproject/vtn-synchronizer",
    "xosproject/vtr-synchronizer",
    "xosproject/xos",
    "xosproject/xos-client",
    "xosproject/xos-corebuilder",
    "xosproject/xos-gui",
    "xosproject/xos-postgres",
    "xosproject/xos-synchronizer-base",
    "xosproject/xos-ui",
    "xosproject/xos-ws"
  ]
}

NOTE: This example uses the curl and jq to retrieve data and pretty print JSON. If your system doesn't have these commands installed, they can be installed using sudo apt-get install -y curl jq.

##Head Node Deployment

Head node deployment works as follows:

  • Makes the head node a MaaS server from which the other POD elements (fabric switches and compute nodes) can PXE boot (both to load their OS and to be configured).
  • Installs and configures the containers needed to configure other nodes of the network.
  • Installs and configures OpenStack.
  • Provisions XOS, which provides service provisioning and orchestration for the CORD POD.

This step is started with the following command:

./gradlew -PdeployConfig=config/podX.yml deploy

NOTE: Be patient: this step can take a couple hours to complete.

WARNING: This command sometimes fails for various reasons. Simply re-running the command often solves the problem. If the command fails it’s better to start from a clean head node. Most of the time, re-starting from the publish step (which creates new containers on the head node) helps.

If the process runs smoothly, the output should be similar to:

PLAY RECAP *********************************************************************
localhost                  : ok=5    changed=2    unreachable=0    failed=0   

Monday 19 June 2017  22:59:22 +0000 (0:00:00.233)       0:00:03.370 *********** 
=============================================================================== 
setup ------------------------------------------------------------------- 
1.35s
setup ------------------------------------------------------------------- 1.18s
automation-integration : Template do-enlist-compute-node script to /etc/maas/ansible/do-enlist-compute-node --- 0.46s
automation-integration : Have MAAS do-ansible script run do-enlist-compute-node script --- 0.23s
Include variables ------------------------------------------------------- 0.12s
:PIdeployPlatform
:deploy

BUILD SUCCESSFUL

Total time: 57 mins 25.458 secs

This step is complete when the command successfully runs.

###MaaS

As previously mentioned, once the deployment is complete the head node becomes a MaaS region and rack controller, basically acting as a PXE server and serving images through the management network to compute nodes and fabric switches connected to it.

The Web UI for MaaS can be viewed by browsing to the head node, using a URL of the from http://head-node-ip-address/MAAS.

To login to the web page, use cord as the username. If you have set a password in the deployment configuration password use that, otherwise the password used can be found in your build directory under <base>/build/maas/passwords/maas_user.txt. After the deploy command installs MAAS, MAAS itself initiates the download of an Ubuntu 14.04 boot image that will be used to boot the other POD devices. This download can take some time and the process cannot continue until the download is complete. The status of the download can be verified through the UI by visiting the URL http://head-node-ip-address/MAAS/images/, or via the command line from head node via the following command:

APIKEY=$(sudo maas-region-admin apikey --user=cord) && \
maas login cord http://localhost/MAAS/api/1.0 "$APIKEY” && \
maas cord boot-resources read | jq 'map(select(.type != "Synced"))'

If the output of of the above commands is not an empty list ([]) then the images have not yet been completely downloaded. Depending on your network speed, this could take several minutes. Please wait and then attempt the last command again, until the returned list is empty.

When the list is empty you can proceed.

###Compute Node and Fabric Switch Deployment

The section describes how to provision and configure software on POD compute nodes and fabric switches.

####General Workflow

Once it has been verified that the Ubuntu boot image has been downloaded, the compute nodes and the fabric switches may be PXE booted.

Compute nodes and switches should be simply rebooted. The head node (through MaaS) will act as DHCP and PXE server. It will install the OSs and will make sure they are correctly configured.

At the end of the process, the compute and switch elemlents should be visible through the CORD CLI utilities and MAAS.

WARNING: make sure your computes nodes and fabric switches are configured as prescribed in the Software Environment Requirements section.

####Important Commands: cord harvest and cord prov

Two important commands are available to debug and check the status of the provisioning. They can be used from the head node CLI.

  • cord harvest: Tracks the nodes harvesting process. Nodes and switches should appear here, as soon as they get an IP and are recognized by MaaS. To see if your devices have been recognized, use the following command:
cord harvest list
  • cord prov: Tracks the provisioning process, meaning the configuration process that happen soon after the OS has been installed on your devices. To see the provisioning status of your devices, use the following command:
cord prov list

The following status values are defined for the provisioning status:

  • Pending: The request has been accepted by the provisioner but not yet started
  • Processing: The request is being processed and the node is being provisioned
  • Complete: The provisioning has been completed successfully
  • Error: The provisioning has failed and the message will be populated with the exit message from provisioning.

Logs of the post deployment provisioning can be found in /etc/maas/ansible/logs on the head node.

For a given node, the provisioning re-starts automatically if the related entry gets manually removed. This can be done with the following command:

cord prov delete node_name

Please refer to Re-provision Compute Nodes and Switches for more details.

####Static IP Assignment

If you want to assign a specific IP to either a compute node or a fabric switch, it should be done before booting the device. This is achieved through a configuration file: /etc/dhcp/dhcpd.reservations.

To help you, a sample file is available: /etc/dhcp/dhcpd.reservations.sample. For each host you want to statically assign an IP, use this syntax:

host <name-of-your choice> {
	hardware ethernet <host-mac-address>;
	fixed-address  <desired-ip>;
	}

####Compute Nodes

The compute node provisioning process installs the servers as OpenStack compute nodes.

The compute node will boot, register with MaaS, and then restart (eventually multiple times).

Compute nodes are given a random hostname, in the “Canonical way”, of an adjective and a noun (e.g., popular-feast.cord.lab). The name will be different for every deployment.

After this is complete, an entry for each node will be visible:

  • From the MaaS UI, at http://head-node-ip-address/MAAS/#/nodes

  • From the OpenStack CLI on the head node, using the command

source ~/admin-openrc.sh &&
nova hypervisor-list
  • From CORD head node CLI, using the cord harvest command

In MaaS, the new node will be initially in a New state. As the machines boot, they should automatically transition from New through the states Commissioned, Acquired and Deployed.

Once the node is in the Deployed state, it will be provisioned for use in a CORD POD by the automated execution of an Ansible playbook.

The post deployment provisioning of the compute nodes can be queried using the cord prov command.

After a correct provisioning you should see something similar to:

cord prov list
ID                                         NAME                   MAC                IP          STATUS      MESSAGE
node-c22534a2-bd0f-11e6-a36d-2c600ce3c239  steel-ghost.cord.lab   2c:60:0c:cb:00:3c  10.6.0.107  Complete
node-c238ea9c-bd0f-11e6-8206-2c600ce3c239  feline-shirt.cord.lab  2c:60:0c:e3:c4:2e  10.6.0.108  Complete

Once the post deployment provisioning on the compute node is complete, this task is complete.

####Fabric Switches

Similar to the compute nodes, the fabric switches will boot, register with MaaS, and then restart (eventually multiple times).

If a name hasn’t been assigned to the switches (see the static IP assignment section above), usually switches have a name in the form UKN-XXXXXX.

When the fabric switches get an IP and go through the harvesting process, they should be visible in MaaS, under the devices tab (http://head-node-ip-address/MAAS/#/devices).

As with the compute nodes, following the harvest process, the provisioning will happen. After a correct provisioning you should see something similar to:

cord prov list
ID                                         NAME                   MAC                IP          STATUS      MESSAGE
cc:37:ab:7c:b7:4c                          UKN-ABCD                cc:37:ab:7c:b7:4c  10.6.0.23   Complete
cc:37:ab:7c:ba:58                          UKN-EFGH                cc:37:ab:7c:ba:58  10.6.0.20   Complete
cc:37:ab:7c:bd:e6                          UKN-ILMN                cc:37:ab:7c:bd:e6  10.6.0.52   Complete
cc:37:ab:7c:bf:6c                           UKN-OPQR                cc:37:ab:7c:bf:6c  10.6.0.22   Complete

NOTE: cord prov list output for compute nodes is not shown here for simplicity.

Once the post deployment provisioning on the fabric switches is complete, the task is complete.

##Access to CORD Services

Your POD is now installed. You can now try to access the basic services as described below.

###ONOS (Underlay)

A dedicated ONOS instance is installed on the head node to control the underlay infrastructure (the fabric). You can access it with password “rocks”

  • From the head node CLI: ssh -p 8101 onos@onos-fabric

  • Using the ONOS UI, at: http://<head-node-ip>/fabric

###ONOS (Overlay)

A dedicated ONOS instance is installed on the head node to control the overlay infrastructure (tenant networks). You can access it with password “rocks”

  • From the head node CLI: ssh -p 8102 onos@onos-cord

  • Using the ONOS UI, at: http://<head-node-ip>/vtn

###OpenStack

###XOS UI

XOS is the cloud orchestrator that controls the entire POD. It allows you to define new service and service dependencies.. You can access XOS:

  • Using the XOS GUI at http://<head-node-ip>/xos

  • Using the XOS admin UI at http://<head-node-ip>/admin/

Getting Help

If it seems that something has gone wrong with your setup, there are a number of ways that you can get help -- in the documentation on the OpenCORD wiki, on the OpenCORDSlack channel (get an invitation here), or on the CORD-discuss mailing list. See the How to Contribute to CORD wiki page for more information.