Open stack ha design & deployment kilo

- OpenStack Control Services High Availability
OpenStack HA Design &
Deployment - Kilo
Steven004@gmail.com
Oct. 2015

Four types of HA in an OpenStack Cloud
Compute Controller
Network Controller
Database
Message Queue
Storage
....
Physical nodes
Physical network
Physical storage
Hypervisor
Host OS
….
Service Resiliency
QoS Cost
Transparency
Data Integrity
…..
Virtual Machine
Virtual Network
Virtual Storage
VM Mobility
…Physical infrastructure
OpenStack
Control services
VMs
OpenStack Compute
Applications

• Redundant in locations – Minimize downtime due to power
or network issue
• All components distributed on different labs
• Controller, Network, Compute, Storage
• Minimize downtime
• Minimize data loss risk
• Eliminate single point of failure
• Extendability
Physical Infrastructure

Physical Infrastructure - example
Zone1 Zone 2
APIs/Orchestration/Dashboard…
Compute Compute
Database for control plane (MySQL)
Message Queue
Network
Storage Storage

OpenStack Control Services
High Availability Technology

Open stack ha design & deployment kilo

Stateless / Stateful Services
State Description Services
Stateless • There is no dependency between
requests
• No need for data
replication/synchronization. Failed
request may need to be restarted
on a different node.
Nova-api, nova-
conductor, glance-api,
keystone-api, neutron-
api, nova-scheduler,
Apache web server,
Cinder Scheduler, etc.
Stateful • An action typically comprises
multiple requests
• Data needs to be replicated and
synchronized between redundant
services (to preserve state and
consistency)
MySQL, RabbitMQ,
Cinder Volume,
Ceilometer center
agent, Neutron L3,
DHCP agents, etc.

• Active/Passive
• There is a single master
• Load balance stateless services
using a VIP and a load balancer such
as HAProxy
• For Stateful services a replacement
resource can be brought online. A
separate application monitors these
services, bringing the backup online
as necessary
• After a failover the system will
encounter a “speed bump” since the
passive node has to notice the fault
in the active node and become
active
• Active/Active
• Multiple masters
• Load balance stateless services
using a VIP and a load balancer such
as HAProxy
• Stateful Services are managed in
such a way that services are
redundant, and that all instances
have an identical state
• Updates to one instance of database
would propagate to all other
instances
• After a failover the system will
function in a “degraded” state
Active/Passive or Active/Active

Do not reinvent the wheel
• Leverage time-tested Linux utilities such as Keepalived,
HAProxy and Virtual IP (using VRRP)
• Leverage Hardware Load Balancers
• Leverage replication services for RabbitMQ/MySQL such as
RabbitMQ Clustering, MySQL master-master replication,
Corosync, Pacemaker, DRBD, Galera and so on
Overall Philosophy

• Keepalived
• Based on Linux Virtual Server (IPVS) kernel module
providing layer 4 Load Balancing
• Implements a set of checkers to maintain health and Load
Balancing
• HA is implemented using VRRP Protocol, Used to load
balance API services
• VRRP (Virtual Router Redundancy Protocol)
• Eliminates SPOF in a static default routed environment
• HAProxy
• Load Balancing and Proxying for HTTP and TCP
Applications
• Works over multiple connections
• Used to load balance API services
Keepalived, VRRP, HAProxy
– for APIs (Active/Active – 2 nodes)

• Corosync
• Totem single-ring ordering and membership protocol
• UDP and InfiniBand based messaging, quorum, and cluster
membership to Pacemaker
• Pacemaker
• High availability and load balancing stack for the Linux platform.
• Interacts with applications through Resource Agents (RA)
• DRDB (Distributed Replication Block Device)
• Synchronizes Data at the block device
• Uses a journaling system (such as ext3 or ext4)
Corosync, Pacemaker and DRDB
- for APIs and MySQL (Active/Passive)

• MySQL patched for wsrep (Write Set
REPlication)
• Active/active multi-master topology
• Read and write to any cluster node
• True parallel replication, in row level
• No slave lag or integrity issues
MySQL Galera (Active/Active)
Synchronous Multi-master Cluster technology for MySQL/InnoDB

• Cinder (Block Storage) backends support
• LVM Driver
• Default linux iSCSI server
• Vendor software plugins
• Gluster, CEPH, VMware VMDK driver
• Vendor storage plugins
• EMC VNX, IBM Storwize, Solid Fire, etc.
• Local RAID support
• Swift (Object Storage) -- Done
• Replication
• Erasure coding: (not enabled)
Data Redundancy (storage HA)

• No need to HA support for L2 networking, which is located in
compute node
• Problems
• Routing on Linux server (max. bandwith approximately 3-4 Gbits)
• Limited distribution between more network nodes
• East-West and North-South communication through network node
• High Availability
• Pacemaker&Corosync
• Keepalived VRRP
• DVR + VRRP – should be in Juno release
Networking – Vanilla Neutron L3 agent
Reference:
• Neutron/DVR
• L3 High Availability
• Configuring DVR in OpenStack Juno

HA methods in different vendors
Vendor Cluster/Replication Technique Characteristics
RackSpac
e
Keepalived, HAProxy, VRRP, DRBD, native
clustering
Automatic - Chef
for 2 controller nodes installation
Red Hat Pacemaker, Corosync, Galera Manual installation/Foreman
Cisco Keepalived, HAProxy, Galera
Manual installation, at least 3
controller
tcp cloud
Pacemaker, Corosync, HAProxy,Galera,
Contrail
Automatic Salt-Stack deployment
Mirantis Pacemaker, Corosync, HAProxy,Galera Automatic - Puppet
HP
Microsoft Windows based installation with
Hyper-V
MS SQL server and other Windows
based methods
Ubuntu Juju-Charms, Corosync, Percona XtraDB, Juju+MAAS

Comparison
Database Replication
method
Strengths Weakness/Limit
ations
Keepalived/HAPro
xy/VRRP
Works on MySQL
master-master
replication
Simple to
implement and
understand.
Works for any
storage system.
Master-master
replication does
not work beyond
2 nodes.
Pacemaker/Coros
ync/DRBD
Mirroring on Block
Devices
Well tested More complex to
setup. Split Brain
possibility
Galera Based on write-
set Replication
(wsrep)
No Slave lag Needs at least 3
nodes. Relatively
new.
Others MySQL Cluster,
RHCS with
DAS/SAN storage
Well tested More complex
setup.

• HAProxy for load
balancing
• MySQL Galera –
active/active
• RabbitMQ cluster
Sample OpenStack HA architecture -1

• HAProxy for load balancing
• MySQL Galera –
active/active
• RabbitMQ cluster
• DVR + VRRP for network
Sample OpenStack HA architecture - 2
HAProxy
VIP
HAProxy
Keepalived
Controller
keystone
glance
cinder
horizon
rabbitmq
nova
Controller
keystone
glance
cinder
horizon
rabbitmq
nova
MySQL MySQL
galera
Storage Storage
Block Block
Object Object
Network / Compute Network / Compute
DVR + VRRP

• OpenStack High Availability Guide
• Ubuntu OpenStack HA wiki
• RackSpace OpenStack Control Plane High Availability
• TCP Cloud OpenStack High Availability
• Configuring DVR in OpenStack Juno
• OpenStack High Availability – Controller Stack by Brian
Seltzer
Reference

Overall Picture
VIP
1
PortalHost1
PortalHost2
HAProxy1
HAProxy2
Keepalived
JumpBox4
JumpBox3
ExternalNetwork Internal Network
VIP2
Controller
keystone
glance
cinder
horizon
rabbitmq
nova
MySQL
Controller
keystone
glance
cinder
horizon
rabbitmq
nova
MySQL
Compute
Compute
Compute
Compute
Compute
Storage
JumpBox1
JumpBox2
Network
L3
Network
L3
Keepalived
VIP
rabbitmq
MySQL
To external Network

• OpenStack Release: Kilo
• Host computers
• Cisco UCS for Controller, Compute, Network nodes
• SuperMicro Computer for Storage nodes
• Host OS: Ubuntu 14.04 Server
• Network Switches: Cisco Nexus – N7K, N5K, N2K
• IP assignment:
• All hosts are using Lab internal IP address to save IP
addresses resource
• For Management/tunnel/storage/… cloud networks
• Use Jumpbox to access the all the cloud host computers from
outside, 4 Jumpbox are set up for redundancy
• HAProxies for internal load balancing and dashboard portal
for outside.
Equipment and Software

• Two portal hosts for redundancy
and load balance
• Same configurations on both
• One node hosts 3 VMs for 2 jumpbox
and 1 haproxy
• Jumpbox to Cloud management
• All IP addresses in Cloud are private,
reachable via Jumpbox from outside
• Applications: VNC, Java, Wireshark, …
• Repository mirroring for Linux(Ubuntu
14.04) and OpenStack (Kilo)
• Mirror required since internal network can
not access Internet directly
• Locate on Jumpbox
• Dashboard portal (on HAProxies)
• VIPs for load-balance
• VIP1 for external network access
• VIP2 for load balance of all Cloud APIs,
Database, MessageQ, …
• Important: Two VIPs should be in one
VRRP group.
Set up Portal Hosts
VIP1
PortalHost1
PortalHost2
HAProxy1
HAProxy2
Keepalived
JumpBox4
JumpBox3
External Network
(Cisco)
VIP2
JumpBox1
JumpBox2
Internal Network
(Cloud)
Step 1

Portal Hosts Index Example
VIP1
PortalHost1
IPMI: 10.10.10.6
Host: 10.10.10.9
PortalHost2
IPMI: 10.10.10.7
Host: 10.10.10.8
HAProxy1
HAProxy2
Keepalived
JumpBox4
JumpBox3
External Network
(Cisco)
VIP2
JumpBox1
JumpBox2
Internal Network
(Cloud)
10.10.10.11
10.10.10.12
10.10.10.10
gw 10.10.10.1
192.168.222.240
gw 192.168.222.1
192.168.222.251
192.168.222.252
192.168.222.242
192.168.222.241
192.168.222.243
192.168.222.244
Windows 10.10.10.14
Ubuntu 10.10.10.13
Ubuntu 10.10.10.15
Windows 10.10.10.16
Assume the 10.x.x.x is the company network IPs, 192.168.*.* is for lab internal use

• HAProxy configuration for VIP1 (external network)
• Keepalived configuration
• VIP1 and VIP2 should be in one VRRP group
• Once there is one interface fail, the whole function will be taken over
by another host
• HAProxy configuration for VIP2 (internal network)
• http://docs.openstack.org/high-availability-guide/content/ha-aa-
haproxy.html
HAProxy and Keepalived set up
Step
1.1

• 4 jumpbox set up
• 2 Windows and 2 Linux
• Software installed:
• VNC
• WireShark
• Vmclient
• Putty …
• Repository Mirror for Ubuntu 14.04 and OpenStack Kilo set
up in 2 Linux jumpbox
• The internal network will get package from the Jumpbox directly
Jumpbox and Repository Mirroring
Step
1.2

• NIS servers set up for the cloud infrastructure, for
• Host configuration
• Authentication
• …
• Two NIS servers set up on the HAProxy hosts
• Master and slave
NIS set up on HAProxy host (option)

• 3 UCS hosts for Controller, Database, and MessageQ
• Located in two racks
• Better to have Network / Compute all located in UCS-B
hosts
• Be sure the Mac Pools are set differently in different FIs,
otherwise, there will be Mac Address conflict
• Complete all cabling and network configuration on UCSes
and upper switches
• Verify all network connectivity of IPMI ports
• Write down all the configuration in a detailed document
Cloud host
Step 2
Similar setting when using other compute hosts

• 2 Portal Hosts
• 2 HAProxy, 4 Jumpbox
• 2 Portal hosts, for each:
• IPMI: VLAN aaa (external network)
• Eth0: (7 external IP) – VLAN eee
• Eth1: (7 Internal IPs) – VLAN mmm -- VLAN access port.
• All Cloud hosts
• IPMI vlan/network (lab internal)
• accessible via jumpbox from external network
• Management vlan/network (lab internal)
• Accessible via jumpbox from external network
• Tunnel vlan/network (lab internal) – not accessible from external
• Storage vlan/network (lab internal) – not external accessible
• Other internal network (e.g. internal VLAN network)
VLANs / IP design

• Network configuration for each node
• Each host in one network can connect each other
• Each host can reach HAProxy and JumpBox via management interface
• Each host can reach HAProxy VIP2 via management interface
• Hosts set up: controller-vip is used for APIs
• Install the Ubuntu Cloud archive keyring and repository
• Use the mirror address, instead of the standard one
• Update packages for each system, via mirror on jumpbox
• Verification
1. NTP: ntpq –c peers
2. Connectivity: can reach HAProxy and Jumpbox
3. Repository setup: /etc/apt/sources.list..d/ …
4. Upgraded the packages
Host system preparation Check on
each host

• NTP Source:
• Select a stable NTP source from external network as standard time server
• The VMs on portal hosts should configure to follow the standard time
server listed above
• Jumpboxes and HAProxy
• All internal hosts in cloud should follow the HAProxy host
• Using the VIP2
NTP set up
Step 3

• The MySQL/Maria Galera are deployed in 3 hosts: 2 controllers and another
• Make sure InnoDB is configured
• Configure HAProxy to listen on galera cluster api, and load balance (Port:
3306) .
• Verification
• Create table on one node, can be access/manipulate from another
• Mysql work well through VIP2, and verify tolerance of single node failure
• Access from Jumpbox, work fine.
• References:
• http://docs.openstack.org/high-availability-guide/content/ha-aa-db-mysql-
galera.html
• Product webpage： http://www.codership.com/content/using-galera-cluster/。
• Download: http://www.codership.com/downloads/download-mysqlgalera。
• Document: http://www.codership.com/wiki。
• More information about wsrep, see https://launchpad.net/wsrep
MySQL/MariaDB Galera Setup
Step 4

• Deploy on 3 nodes, including two controllers
• Configure them as a cluster, all nodes are disk nodes
• Configure HAProxy for load balance (port: 5672) , to use multiple
rabbit_hosts instead.
• Verification
• rabbitmqadmin tool?
• rabbitmqctl status
• References:
• http://docs.openstack.org/high-availability-guide/content/ha-aa-
rabbitmq.html
• http://88250.b3log.org/rabbitmq-clustering-ha
• OpenStack High Availability: RabbitMQ
RabbitMQ Cluster Setup
Step 5

• Services contain:
• All OpenStack API services
• All OpenStack Schedulers
• Memcashed service (multiple instances can be configured, consider later)
• API services
• User VIP2 when configuring Keystone endpoints
• All configuration files should refer to VIP2
• Schedulers: use RabbitMQ as the message system, hosts configured:
• http://docs.openstack.org/admin-guide-cloud/content/section_telemetry-cetral-
compute-agent-ha.html
• Telemetry central agent set up can be load balanced:
• See also: http://docs.openstack.org/high-availability-guide/content/ha-aa-controllers.html
Control services set up

• Installation
• Add Database into MySQL, and grant privileges (once)
• Install Keystone components on each node (one each node)
• Configure the keystone.conf
• Configure the backend to sql database
• Disable caching if needed (need to try)
• Configure the HAProxy for the API
• Configure the keystone token backend to sql (the default is
memcached)
• Services/Endpoints/Users/Roles/Projects & Verification
• Create admin, demo, service projects and corresponding user/role, …
• Create on one node and verify it on another node
• Work with VIP2
• See also:
• http://docs.openstack.org/high-availability-guide/content/s-keystone.html
• http://docs.openstack.org/kilo/config-reference/content/section_keystone.conf.html
Identity Service - Keystone
Step 6

• Shared storage is required for Glance
HA
• In the pilot cloud, the controller local file
system is used as image storage, for HA,
it will not work
• Use the swift as Glance backend.
• Swift itself needs to be HA
• At least two storage nodes
• At least two swift proxy nodes
• Installed on controllers with glance
• Use keystone for authentication, instead
of Swauth
Image Service - Glance
Step 7
VIP2
…
Controller
keystone
glance
MySQL
Swift Proxy…
Controller
keystone
glance
MySQL
Swift Proxy
HAProxy1 HAProxy2
Swift

• Installation
• Install two swift proxy
• Proxies can be located on the controller nodes; Configure VIP2 for them for load-
balance
• Install two storage nodes in B-series nodes, two disks for each, total 4
• Configure 3 replicators for HA
• No account error fix – upgrade swift-client to 2.3.2.
• There is a bug, fixed in 2.3.2 (not in Kilo release)
• Verification
• File can be put into the storage via one of proxies, and get from another
• Object write/get via VIP2
• Failure cases
• See also: http://docs.openstack.org/kilo/install-
guide/install/apt/content/ch_swift.html
• https://bugs.launchpad.net/python-swiftclient/+bug/1372465
Object Storage Installation
Step
7.1

• Install Glance on each controller
• Use the file system as the backend first, verify it works with local
file system
• Configure the HAProxy for Glance API and Glance Registry service
• There would be warning about Unknown version in Glance API log
• Need to change HAProxy setting about httpchk to fix it
• option httpchk Get /versions
• See also:
• http://docs.openstack.org/juno/config-reference/content/section_glance-
api.conf.html
• https://bugzilla.redhat.com/show_bug.cgi?id=1245572
• http://docs.openstack.org/high-availability-guide/content/s-glance-
api.html
Image Service - Glance
Step
7.2

• Prerequisites:
• Swift object store had been installed and verified
• Glance installed on controllers and verified with local file system as the backend
• Integration:
• Configure the swift store as the glance backend
• Configure the keystone token backend to sql (important)
• Or configure multiple memcached hosts in the configuration file
• Verification
• Upload image and list images successfully in each controller node
• See also:
• http://behindtheracks.com/2014/05/openstack-high-availability-glance-
and-swift/
• http://thornelabs.net/2014/08/03/use-openstack-swift-as-a-backend-
store-for-glance.html
Integration of Glance and Swift
Step
7.3

• Install Nova related packages
• In two controller nodes and in one compute node
• Compute nodes need to be set up as “Virturelization Host”
• Otherwise, the installation later will fail due to dependency issue.
• Configure HAProxy for Nova services
• List the Nova services
• Verify if RabbitMQ works in the HA environment
• There should be redundant Nova APIs, Schedulers, Conductors,
… listed
• Further verification needs network nodes set up
Install Compute Services
Step 8

• DVR + 3 network nodes (distributed SNAT / DHCP redundency)
• Multiple options, but no one is perfect
• Pacemaker&Corosync
• Keepalived VRRP
• DVR + VRRP – should be in Juno/Kilo releases
• References:
• http://docs.openstack.org/networking-guide/scenario_dvr_ovs.html
• https://blogs.oracle.com/ronen/entry/diving_into_openstack_network_ar
chitecture
• http://assafmuller.com/2014/08/16/layer-3-high-availability/
Network Redundancy
Step 9

Service Layout for DVR mode
• Network Node
• Services required the same as the
central mode
• Compute Node
• Compute node is doing networking too
• L3 services added in Compute node
• Take most networking functions too

• Support GRE/VXLAN/VLAN/FLAT
network
• In our system, GRE is used for
tunneling between instances and SNAT
• VLAN network is not required if we do
not use it.
• Network Node is mainly for network
central services like DHCP,
Metadata, and SNAT
• Just north/south traffic with fixed IP
need network node forwarding
• Compute nodes handle DNAT
• East/West traffic and North/South
traffic with a floating IP will not go
through network node
DVR General Architecture

• This is just for load balance of networking, not HA really
• Install all services listed in the service layout picture
• On Network node and Compute node respectively
• Configuration
• router_distrbuted = True
• ……
• Create routers, networks and instances
• Verification
• North/south for instances with a fixed IP address (SNAT, via Network
node)
• North/south for instances with a floating IP address (DNAT, via Compute
node only)
• East/west for instances using different networks on the same router (via
Compute node only)
Install one Network and two Compute node
Step
9.1

• Add one more network Node
• DHCPD redundancy
• Networking L3 Agent redundancy
• Networking Metadata Agent
• Kilo does not support DVR & L3HA mechanism combination
• This is not implemented in our practice, but it should be
feasible to implement
• The key is to keep all configuration (static/dynamic) sync-up
• Two ways to go:
• PaceMaker + CoroSync …
• VRRP + Keepalived (need to reboot network node when one is down)
L3 network redundancy - TBD
Step
9.2

• Cinder Services installation
• Install Cinder services in each controller
• Configure HAProxy for the API
• Storage Nodes set up
• SuperMicro equipment as storage
• Linux soft-raid for disk redundancy
• GlusterFS for node redundancy
• Verification
• Create and access volume through client on
Jumpbox – try both controllers
• Do failover cases on disk level
• Do failover cases on node level
Volume redundancy
Step 10
VIP2
…
Controller
keystone
MySQL
Cinder
…
Controller
keystone
MySQL
Cinder
HAProxy1 HAProxy2
It’s the same to use any other storage node, e.g. normal computer,
we use SuperMicro since it provides >100T storage on one node
RaidRaid
SuperMicro Storage Node
Gluster FS

• Horizon Services installation
• Install Horizon services in each controller
• Configure Horizon services
• Use the external url name for console setting (instead of controller)
• Configure the memcached
• /etc/openstack-dashboard/local_settings.py – CACHES LOCATION changes to VIP2
• /etc/memcached.conf change 127.0.0.1 to the controller IP.
• Configure HAProxy for the API
• Configure the VIP1 to the internal controllers proxy
• Make sure the Dashboard is accessible from external network
• Verification
• From Jumpbox to access the dashboard
• From external network to access the dashboard
Dashboard redundancy
Step 11

• HEAT Services installation
• Install HEAT services in each controller
• Configure HEAT services
• Configure HAProxy for the HEAT API
• Verification
Orchestration redundancy
Step 12

• Ceilometer Services installation
• Install Ceilometer services in each controller
• Configure Ceilometer services
• Configure HAProxy for the Ceilometer API
• Verification
Telemetry redundancy - TBD
Step 13

Open stack ha design & deployment kilo

Recommended

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Open stack ha design & deployment kilo (20)

Recently uploaded (19)

Open stack ha design & deployment kilo