devstack/doc/source/guides/multinode-lab.rst

Multi-Node Lab

Here is OpenStack in a realistic test configuration with multiple physical servers.

Prerequisites Linux & Network

Minimal Install

You need to have a system with a fresh install of Linux. You can download the Minimal CD for Ubuntu releases since DevStack will download & install all the additional dependencies. The netinstall ISO is available for Fedora and CentOS/RHEL.

Install a couple of packages to bootstrap configuration:

apt-get install -y git sudo || yum install -y git sudo

Network Configuration

The first iteration of the lab uses OpenStack's FlatDHCP network controller so only a single network will be required. It should be on its own subnet without DHCP; the host IPs and floating IP pool(s) will come out of this block. This example uses the following:

• Gateway: 192.168.42.1
• Physical nodes: 192.168.42.11-192.168.42.99
• Floating IPs: 192.168.42.128-192.168.42.254

Configure each node with a static IP. For Ubuntu edit /etc/network/interfaces:

auto eth0
iface eth0 inet static
    address 192.168.42.11
    netmask 255.255.255.0
    gateway 192.168.42.1

For Fedora and CentOS/RHEL edit /etc/sysconfig/network-scripts/ifcfg-eth0:

BOOTPROTO=static
IPADDR=192.168.42.11
NETMASK=255.255.255.0
GATEWAY=192.168.42.1

Installation shake and bake

Add the DevStack User

OpenStack runs as a non-root user that has sudo access to root. There is nothing special about the name, we'll use stack here. Every node must use the same name and preferably uid. If you created a user during the OS install you can use it and give it sudo privileges below. Otherwise create the stack user:

useradd -s /bin/bash -d /opt/stack -m stack

This user will be making many changes to your system during installation and operation so it needs to have sudo privileges to root without a password:

echo "stack ALL=(ALL) NOPASSWD: ALL" >> /etc/sudoers

From here on use the stack user. Logout and login as the stack user.

Set Up Ssh

Set up the stack user on each node with an ssh key for access:

mkdir ~/.ssh; chmod 700 ~/.ssh
echo "ssh-rsa AAAAB3NzaC1yc2EAAAADAQABAAABAQCyYjfgyPazTvGpd8OaAvtU2utL8W6gWC4JdRS1J95GhNNfQd657yO6s1AH5KYQWktcE6FO/xNUC2reEXSGC7ezy+sGO1kj9Limv5vrvNHvF1+wts0Cmyx61D2nQw35/Qz8BvpdJANL7VwP/cFI/p3yhvx2lsnjFE3hN8xRB2LtLUopUSVdBwACOVUmH2G+2BWMJDjVINd2DPqRIA4Zhy09KJ3O1Joabr0XpQL0yt/I9x8BVHdAx6l9U0tMg9dj5+tAjZvMAFfye3PJcYwwsfJoFxC8w/SLtqlFX7Ehw++8RtvomvuipLdmWCy+T9hIkl+gHYE4cS3OIqXH7f49jdJf jesse@spacey.local" > ~/.ssh/authorized_keys

Download DevStack

Grab the latest version of DevStack:

git clone https://git.openstack.org/openstack-dev/devstack
cd devstack

Up to this point all of the steps apply to each node in the cluster. From here on there are some differences between the cluster controller (aka 'head node') and the compute nodes.

Configure Cluster Controller

The cluster controller runs all OpenStack services. Configure the cluster controller's DevStack in local.conf:

[[local|localrc]]
HOST_IP=192.168.42.11
FLAT_INTERFACE=eth0
FIXED_RANGE=10.4.128.0/20
FIXED_NETWORK_SIZE=4096
FLOATING_RANGE=192.168.42.128/25
MULTI_HOST=1
LOGFILE=/opt/stack/logs/stack.sh.log
ADMIN_PASSWORD=labstack
DATABASE_PASSWORD=supersecret
RABBIT_PASSWORD=supersecret
SERVICE_PASSWORD=supersecret

In the multi-node configuration the first 10 or so IPs in the private subnet are usually reserved. Add this to local.sh to have it run after every stack.sh run:

for i in `seq 2 10`; do /opt/stack/nova/bin/nova-manage fixed reserve 10.4.128.$i; done

Fire up OpenStack:

./stack.sh

A stream of activity ensues. When complete you will see a summary of stack.sh's work, including the relevant URLs, accounts and passwords to poke at your shiny new OpenStack. The most recent log file is available in stack.sh.log.

Configure Compute Nodes

The compute nodes only run the OpenStack worker services. For additional machines, create a local.conf with:

[[local|localrc]]
HOST_IP=192.168.42.12 # change this per compute node
FLAT_INTERFACE=eth0
FIXED_RANGE=10.4.128.0/20
FIXED_NETWORK_SIZE=4096
FLOATING_RANGE=192.168.42.128/25
MULTI_HOST=1
LOGFILE=/opt/stack/logs/stack.sh.log
ADMIN_PASSWORD=labstack
DATABASE_PASSWORD=supersecret
RABBIT_PASSWORD=supersecret
SERVICE_PASSWORD=supersecret
DATABASE_TYPE=mysql
SERVICE_HOST=192.168.42.11
MYSQL_HOST=$SERVICE_HOST
RABBIT_HOST=$SERVICE_HOST
GLANCE_HOSTPORT=$SERVICE_HOST:9292
ENABLED_SERVICES=n-cpu,q-agt,n-api-meta,c-vol,placement-client
NOVA_VNC_ENABLED=True
NOVNCPROXY_URL="http://$SERVICE_HOST:6080/vnc_auto.html"
VNCSERVER_LISTEN=$HOST_IP
VNCSERVER_PROXYCLIENT_ADDRESS=$VNCSERVER_LISTEN

Note: the n-api-meta service is a version of the api server that only serves the metadata service. It's needed because the computes created won't have a routing path to the metadata service on the controller.

Fire up OpenStack:

./stack.sh

A stream of activity ensues. When complete you will see a summary of stack.sh's work, including the relevant URLs, accounts and passwords to poke at your shiny new OpenStack. The most recent log file is available in stack.sh.log.

Starting in the Ocata release, Nova requires a Cells v2 deployment. Compute node services must be mapped to a cell before they can be used.

After each compute node is stacked, verify it shows up in the nova service-list --binary nova-compute output. The compute service is registered in the cell database asynchronously so this may require polling.

Once the compute node services shows up, run the ./tools/discover_hosts.sh script from the control node to map compute hosts to the single cell.

The compute service running on the primary control node will be discovered automatically when the control node is stacked so this really only needs to be performed for subnodes.

Cleaning Up After DevStack

Shutting down OpenStack is now as simple as running the included unstack.sh script:

./unstack.sh

A more aggressive cleanup can be performed using clean.sh. It removes certain troublesome packages and attempts to leave the system in a state where changing the database or queue manager can be reliably performed.

./clean.sh

Sometimes running instances are not cleaned up. DevStack attempts to do this when it runs but there are times it needs to still be done by hand:

sudo rm -rf /etc/libvirt/qemu/inst*
sudo virsh list | grep inst | awk '{print $1}' | xargs -n1 virsh destroy

Options pimp your stack

Additional Users

DevStack creates two OpenStack users (admin and demo) and two projects (also admin and demo). admin is exactly what it sounds like, a privileged administrative account that is a member of both the admin and demo projects. demo is a normal user account that is only a member of the demo project. Creating additional OpenStack users can be done through the dashboard, sometimes it is easier to do them in bulk from a script, especially since they get blown away every time stack.sh runs. The following steps are ripe for scripting:

# Get admin creds
. openrc admin admin

# List existing projects
openstack project list

# List existing users
openstack user list

# Add a user and project
NAME=bob
PASSWORD=BigSecret
PROJECT=$NAME
openstack project create $PROJECT
openstack user create $NAME --password=$PASSWORD --project $PROJECT
openstack role add Member --user $NAME --project $PROJECT
# The Member role is created by stack.sh
# openstack role assignment list

Swift

Swift, OpenStack Object Storage, requires a significant amount of resources and is disabled by default in DevStack. The support in DevStack is geared toward a minimal installation but can be used for testing. To implement a true multi-node test of swift, additional steps will be required. Enabling it is as simple as enabling the swift service in local.conf:

enable_service s-proxy s-object s-container s-account

Swift, OpenStack Object Storage, will put its data files in SWIFT_DATA_DIR (default /opt/stack/data/swift). The size of the data 'partition' created (really a loop-mounted file) is set by SWIFT_LOOPBACK_DISK_SIZE. The Swift config files are located in SWIFT_CONF_DIR (default /etc/swift). All of these settings can be overridden in (wait for it...) local.conf.

Volumes

DevStack will automatically use an existing LVM volume group named stack-volumes to store cloud-created volumes. If stack-volumes doesn't exist, DevStack will set up a 10Gb loop-mounted file to contain it. This obviously limits the number and size of volumes that can be created inside OpenStack. The size can be overridden by setting VOLUME_BACKING_FILE_SIZE in local.conf.

stack-volumes can be pre-created on any physical volume supported by Linux's LVM. The name of the volume group can be changed by setting VOLUME_GROUP_NAME in localrc. stack.sh deletes all logical volumes in VOLUME_GROUP_NAME that begin with VOLUME_NAME_PREFIX as part of cleaning up from previous runs. It is recommended to not use the root volume group as VOLUME_GROUP_NAME.

The details of creating the volume group depends on the server hardware involved but looks something like this:

pvcreate /dev/sdc
vgcreate stack-volumes /dev/sdc

Syslog

DevStack is capable of using rsyslog to aggregate logging across the cluster. It is off by default; to turn it on set SYSLOG=True in local.conf. SYSLOG_HOST defaults to HOST_IP; on the compute nodes it must be set to the IP of the cluster controller to send syslog output there. In the example above, add this to the compute node local.conf:

SYSLOG_HOST=192.168.42.11

Using Alternate Repositories/Branches

The git repositories for all of the OpenStack services are defined in stackrc. Since this file is a part of the DevStack package changes to it will probably be overwritten as updates are applied. Every setting in stackrc can be redefined in local.conf.

To change the repository or branch that a particular OpenStack service is created from, simply change the value of *_REPO or *_BRANCH corresponding to that service.

After making changes to the repository or branch, if RECLONE is not set in localrc it may be necessary to remove the corresponding directory from /opt/stack to force git to re-clone the repository.

For example, to pull nova, OpenStack Compute, from a proposed release candidate in the primary nova repository:

NOVA_BRANCH=rc-proposed

To pull glance, OpenStack Image service, from an experimental fork:

GLANCE_BRANCH=try-something-big
GLANCE_REPO=https://github.com/mcuser/glance.git

Notes stuff you might need to know

Reset the Bridge

How to reset the bridge configuration:

sudo brctl delif br100 eth0.926
sudo ip link set dev br100 down
sudo brctl delbr br100

Set MySQL Password

If you forgot to set the root password you can do this:

mysqladmin -u root -pnova password 'supersecret'

Live Migration

In order for live migration to work with the default live migration URI:

[libvirt]
live_migration_uri = qemu+ssh://stack@%s/system

SSH keys need to be exchanged between each compute node:

1. The SOURCE root user's public RSA key (likely in /root/.ssh/id_rsa.pub) needs to be in the DESTINATION stack user's authorized_keys file (~stack/.ssh/authorized_keys). This can be accomplished by manually copying the contents from the file on the SOURCE to the DESTINATION. If you have a password configured for the stack user, then you can use the following command to accomplish the same thing:

   ssh-copy-id -i /root/.ssh/id_rsa.pub stack@DESTINATION

2. The DESTINATION host's public ECDSA key (/etc/ssh/ssh_host_ecdsa_key.pub) needs to be in the SOURCE root user's known_hosts file (/root/.ssh/known_hosts). This can be accomplished by running the following on the SOURCE machine (hostname must be used):

   ssh-keyscan -H DEST_HOSTNAME | sudo tee -a /root/.ssh/known_hosts

3. Verify that login via ssh works without a password:

   ssh -i /root/.ssh/id_rsa.pub stack@DESTINATION

In essence, this means that every compute node's root user's public RSA key must exist in every other compute node's stack user's authorized_keys file and every compute node's public ECDSA key needs to be in every other compute node's root user's known_hosts file. Please note that if the root or stack user does not have a SSH key, one can be generated using:

ssh-keygen -t rsa

The above steps are necessary because libvirtd runs as root when the live_migration_uri uses the "qemu:///system" family of URIs. For more information, see the libvirt documentation.