Rackspace has years of experience with running Xen at scale, starting with Xen and migrating to XenServer. We will share why we use Xen/XenServer along with some of the issues that we've experienced. We will touch on our experience with migrating from Xen to XenServer and the challenges there. We will share information about Rackspace Cloud Servers architecture, and touch briefly on OpenStack when doing so. We will explain how we use Xen to quickly deploy new Openstack services with what we call Nova on Nova. And finally, we will discuss what additional features and improvements are needed and why.

1. Scaling Xen within
Rackspace Cloud Servers
Chris Behrens
chris.behrens@rackspace.com
XenSummit 2012, San Diego, CA USA
August 28, 2012
1

2. Agenda
• Intro to Rackspace Cloud Servers
• Why Xen?
• Xen to XenServer
• OpenStack Basics
• OpenStack at Rackspace
• Scaling OpenStack Deployment
• Future
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
2

3. Intro to Cloud Servers
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
3

4. Intro to Cloud Servers
Overview
• Public Cloud product under Rackspace’s Open Cloud portfolio
• First Gen
– Code originated from Slicehost acquisition
– Initially Xen 3.x, but migrated to XenServer
• Next Gen
– Public Launch 8/1/2012
– Powered by OpenStack
• XenServer 6
• Linux, Windows, and FreeBSD VMs
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
4

5. Intro to Cloud Servers
Statistics
• 180,000+ total Rackspace customers (not just Cloud Servers)
• Tens of thousands of hosts
• Hundreds of thousands of VMs
• Millions of snapshots
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
5

6. Why Xen?
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
6

7. Why Xen?
• Thin hypervisor layer
• Open source
• Excellent driver support due to Linux dom0
• Good API
• Good performance
• Along with Linux, Windows VMs are a must
– Same virtualization technology desired for Windows
– Microsoft will support
– Good PV driver support
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
7

8. Xen to XenServer
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
8

9. Xen to XenServer
Cloud Servers use of Xen
• LVM
– Logical volumes per VM
• Base images stored as tar files
• Kernel and ramdisk outside of images
• Shell out to run ‘xm’ commands
• Snapshots w/ LVM
– Mounted and contents rsynced to another host
• No agent
– Modifications inside images done via mounting powered off VM
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
9

10. Xen to XenServer
Xen to XenServer changes
• VHDs for virtual disk images
– Tarred and gzipped
• Kernel and ramdisk inside images
• Snapshots via XenAPI
– Whole VDI chain wrapped in OVF container for upload to glance
• Agent inside VM images
– Communication via Xenstore
– File injection, password resets, networking setup
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
10

11. Xen to XenServer
Migrating
• Pick a XenServer host
• Create VDI
• Attach VDI to dom0
• Partition, mkfs, and mount the new virtual disk
• LVM Snapshot original Xen VM and mount it
• rsync snapshot contents to mounted VDI on new host
• Power off original Xen VM and mount its virtual disk
• rsync again to new host to catch changes since snapshotting
• Extract a tar file into mounted VDI containing kernel, ramdisk, agent
• Unmount and detach VDI from dom0
• Create new VM record and attach VDI
• Start the new VM
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
11

12. OpenStack
The Basics
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
12

13. OpenStack
Components
• Nova
– Compute controller with various hypervisor drivers
• Glance
– VM Image registry
• Swift
– Object storage
• Quantum/Melange
– Networking
• Cinder
– Volumes/Block storage
• Keystone
– Authentication
– Authorization
– API Endpoint discovery
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
13

14. OpenStack
Typical OpenStack Deployment
keystone
nova glance
nova-api glance-api glance-registry
RabbitMQ
MySQL
nova-scheduler
MySQL
nova-network
nova-compute swift
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
14

15. OpenStack
nova-compute w/ XenAPI
• Runs in a utility domU on each host
• Images downloaded in dom0 Physical host running XenServer"
– glance XenAPI plugin
• Images attached to nova-compute dom0" nova utility domU"
domU
– Partitioning changes glance plugin nova-compute
– file-system resizing
agent plugin
• Monitors VMs’ power_state
• Snapshots/backups xenstore plugin nova instance
• VM resize/migration domUs
– rsync to new host
• Communicates w/ agent inside VMs
– Via xenstore
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
15

16. OpenStack
XenAPI VM Calls
• VM.start / VM.start_on • VM.add_to_VCPUs_params
• VM.destroy • VM.get_VBDs
• VM.clean_reboot • VM.get_record
• VM.hard_reboot • VM.get_by_uuid
• VM.clean_shutdown • VM.get_by_name_label
• VM.hard_shutdown • VM.set_name_label
• VM.pause • VM.add_to_xenstore_data
• VM.unpause • VM.remove_from_xenstore_data
• VM.suspend • VM.assert_can_migrate
• VM.resume • VM.migrate_send
• VM.snapshot • VM.pool_migrate
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
16

17. OpenStack
Rackspace Specifics
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
17

18. OpenStack Rackspace Specifics
Source Code
• Rackspace tracks trunk
– Deployed code generally less than 2 weeks behind trunk
• Some custom patches on top of trunk
– Features specific to Rackspace
– Custom scheduling to meet Rackspace needs
– Scaling with OpenStack Compute Cells
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
18

19. OpenStack Rackspace Specifics
OpenStack Compute Cells
• Scaling beyond small nova deployments
– Effectively ties multiple nova deployments together
• Developed by Rackspace
• Targeted for Grizzly OpenStack release
• Hierarchical tree
– Multiple parents supported
• API cell(s) at top of the tree
• Compute cell(s) underneath
• DB and RabbitMQ per cell
• Scalability
– Splits connections to DB and RabbitMQ
– Less work to schedule new VM placement
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
19

20. OpenStack Rackspace Specifics
OpenStack Nova Multi-cell Diagram
API Cell
RabbitMQ
MySQL
nova-api
nova-cells
Compute Cell 1" Compute Cell 2"
nova-cells nova-cells
RabbitMQ
RabbitMQ
MySQL
MySQL
nova-scheduler nova-scheduler
nova-network nova-network
nova-compute nova-compute
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
20

21. Deploying OpenStack
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
21

22. Deploying OpenStack
‘Inception’
• Private internal cloud
– Based on Openstack, called iNova
– OpenStack services for public cloud run virtualized under iNova
• Easy to spin up additional services for load balancing
– Enables quick reaction to load spikes
• Easier deployments of new compute cells
• Automated reaction to downed hosts
• Enables new strategies for deployments
– Replace vs Upgrade
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
22

23. Deploying OpenStack
iNova Diagram
iNova"
OpenStack Control
Public Cloud
Infrastructure for
XenServer
Public Cloud"
hosts
XenServer hosts"
domUs
XenServer host w/ Public Cloud
OpenStack Control
Infrastructure for iNova" XenServer
hosts
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
23

24. Deploying OpenStack
Continuous Deployment
• Trunk and custom branches merged multiple times daily
– Unit Tested and Packaged
• Configuration managed with puppet
– Follows same QE and Continuous Deployment rules as code
• QE Environment
– OpenStack control infrastructure deployed on iNova
– nova-compute upgraded on 200 hosts
– Smoke tests
• Staging Deploy
– Uses the same packages and puppet manifests
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
24

25. Future
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
25

26. Future
Improvements desired from Xen
• Better VM resource isolation
– I/O scheduling weights per virtual disk
• Built in live migration without shared storage
• More information about VM resource usage
– Disk usage
– Help us improve OpenStack code
– Help customers react to bad situations
• XenAPI call to write to live xenstore
• Improve XenServer upgrade experience
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
26

27. Future
Rackspace’s Other Interests
• Better VDI introduction for new VMs
– Look at VDI.introduce vs mv’ing VDIs and doing an sr-scan
• Incremental backups
– Delta VHDs instead of all VHDs
• Updated plugins using python >= 2.6
– Cleans up workarounds for no built-in ‘uuid’ module
– Allows use of newer python features like context managers
• Better continuous deployment and testing
– Virtualize all of OpenStack
• Requires spinning XenServer on XenServer
• Wipe out compute nodes for QE environment
• Enable virtual labs for OpenStack developers
– Puppet used to configure XenServer
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
27

28. Future
iNova with Dev/QE Environments
OpenStack QE
iNova" Environment"
OpenStack
XenServer
Control
‘hosts’
Infrastructure
XenServer hosts"
OpenStack Dev
Environment"
domUs
OpenStack
XenServer
Control
‘hosts’
Infrastructure
XenServer host w/
OpenStack Control
Infrastructure for iNova"
RACKSPACE® HOSTING | WWW.RACKSPACE.COM
28

29. "
Chris Behrens"
chris.behrens@rackspace.com"
RACKSPACE® HOSTING | 5000 WALZEM ROAD | SAN ANTONIO, TX 78218
US SALES: 1-800-961-2888 | US SUPPORT: 1-800-961-4454 | WWW.RACKSPACE.COM
RACKSPACE® HOSTING | © RACKSPACE US, INC. | RACKSPACE® AND FANATICAL SUPPORT® ARE SERVICE MARKS OF RACKSPACE US, INC. REGISTERED IN THE UNITED STATES AND OTHER COUNTRIES. | WWW.RACKSPACE.COM
29