Neutron Deployment at Scale Igor Bolotin, Cloud Architecture Vinay Bannai, SDN Architecture
Feb 26, 2016
Neutron Deployment at Scale
Igor Bolotin, Cloud ArchitectureVinay Bannai, SDN Architecture
eBay Inc. enables commerce by delivering flexible and scalable solutions that foster merchant growth.
About ebay inc
With 145 million active buyers globally, eBay is one of the world's largest online marketplaces, where practically anyone can buy and sell practically anything.
With 148 million registered accounts in 193 markets and 26 currencies around the world, PayPal enables global commerce, processing almost 8 million payments every day.
eBay Enterprise is a leading provider of commerce technologies, omnichannel operations and marketing solutions. It serves 1000 retailers and brands.
• Business case
• Cloud at eBay Inc
• Deployment Patterns
• Problem Areas
• How we addressed them
• Future Direction
• Summary
• Q & A
Outline of the Presentation
• Agility− Reduce time to market
− Enable innovation
• Efficiency− Elastic scale
− Reduce overall cost
• Multi-Tenancy
• Availability
• Security & Compliance
• Software Enabled Data Centers
What our businesses need?
5
Cloud at eBay Inc
eBay IncCloud
Region/DC Region/DC Region/DC Region/DC
AZ AZ
Nova Cells
OpenstackControllers
Identity & ImageManagement
AZ AZ
Global Orchestration
• Private Cloud for all eBay Inc properties
• Global Orchestration with traffic and load balancing
• Identity Management− Region level (eventually global)
• Image Management− Region level
• Nova/Cinder/Neutron− Availability Zones
− Active/Active servers
• Trove
• Zabbix for monitoring
• All services run behind a load balancer VIP
Deployment Patterns
• Shared Cloud between tenants
• Different types of tenants− eBay Production, PP production, StubHub, GSI Enterprise etc
− Dev/QA
− Sandbox environment, internal tenants (IT, VPCs)
• Production Traffic− All bridged and no overlays
− No DHCP
• Dev/QA and some of the internal tenants− Overlays
− DHCP
Deployment Patterns (contd.)
8
Gateway Nodes
Physical Racks
• Hypervisor Scale Out
• Overlay Networks
• Bridged Networks
• Neutron Services (DHCP, Metadata, API server)
• SDN Controllers
• Network Gateway Nodes
• Upgrade
Areas With Scale Issues
10
Hypervisor Scale Out
Nova API
Nova Cells
Nova Cells
Nova Sched
Nova Cells
Nova Sched
Neutron APINeutron
APINeutron API
Nova APINova
API
Nova CellsNova
Cells
Nova Cells
Nova Cells
DHCP AgentDHCP
Agent
SDN ContrlSDN
Contrl
• Several hundreds of hypervisor in a cell
• Multiple cells in a AZ
• Several thousands of hypervisors in a AZ
• Nova cells mitigate hypervisor scale
• Neutron scaling− Majority of the hypervisors support Bridged VM’s
− Hybrid mode with both overlay and bridged VM’s
Hypervisor Scale Out
Network Virtualization Layer
L2
VMVM VM VM VM
L2
L2
L3
VMVM VM VM VM
Tenant onOverlayNetwork
Tenant onBridged Network
Bridged and Overlay
• Overlay technology − VXLAN
− STT
• Handling BUM traffic− ARP
− Unknown unicast
− Multicast
• Logical switches and routers
• Distributed L3 routers− Direct tunnels from hypervisor to hypervisor
• Scale out deployment of Gateway nodes
Overlay Networks
• Keystone tokens
• Single threaded Quantum/Neutron server
• DHCP Servers
• Healing Instance Info Caching Interval
Neutron Services
15
Keystone Token Generation and Authentication
Nova Server
Neutron Server
CinderServer
Client
Client
Client
KeystoneServer
ImageServer
• UUID based Token– Needs to be authenticated by
the keystone server for every call
• PKI based Token– Authenticated by the servers
using Keystone certs• Token caching
– Prevents unnecessary token creation
• Applies to uuid based tokens− 98% of tokens generated by inter-API services
− 92% are quantum/neutron related
− Average of 25 to 30 tokens/sec created by quantum/neutron alone
− RPC call overhead, bloated token table
• Fix− Use token caching (1 hour)
− Use PKI for service tenant
− Reduces network chatter and improves performance
• Openstack bugs− Bug id : 1191159
− Bug id : 1250580
Token Caching
• Prior to Havana
• One api thread handling both REST calls and the RPC calls
• Broke up the api to two threads− One handles REST API calls
− The other handles RPC calls
• Havana fixes− DHCP renewals not handled by neutron servers, instead dhcp_release
− Multi-worker support
Neutron Multi Worker
• All nova computes regularly poll neutron server
• To get network info of the instances running on the compute node
• Default is 10 seconds
• Hundreds of hypervisors and tens of thousands of VM’s will add up− Even though only one instance is checked for each interval
• We adjusted the interval to 600 seconds
Heal Instance Info Cache Interval
• The most common source of problems
• We employ multiple strategies− DHCP active/standby
− Planning to support DHCP active/active
− No DHCP/Config Drive option
• Production Environment− No DHCP
− Config drive management
− Requires “cloudinit” aware images
DHCP Scaling
SDN Controllers
SDNController
SDNController
SDNController
Neutron API
Neutron API
Neutron API
Nova API
OSCtrl
OSCtrl
OS Ctrl
• Only with overlay networks
• Scale out architecture
• Problems with high CPU utilization
• Number of flows in the gateway node
• East – West traffic also hitting VIPs− Load Balancer running as a appliance on a hypervisor
− Using SNAT
• OVS Enhancements− Use megaflows in openvswitch
− Multi-core version of ovs-vswitchd
Network Gateway Nodes
• Prior to OVS 1.11
• Megaflow introduces wildcarding in kernel module
• Fewer misses and punts to user space
• Reduced number of flows in kernel
• Requires OVS 1.11 or greater
• Cons− Using security groups nullifies the effects of megaflows
OVS ImprovementsMegaflows
23
OVS ImprovementsMulti-Core
vswitchd
Kernel module
User Space
Kernel Space
cpu cores
vswitchd
Kernel module
OVS < 2.0 OVS >= 2.0
• VPC model
• Neutron Tagging Blueprint− Network Assignment for Bridged VM’s
− Network Selection for VPC tenants
− Network Scheduling
− Additional meta data information
• Blueprint− https://blueprints.launchpad.net/neutron/+spec/network-tagging
Future Work
• There are two primary ways to plumb a VM into a network− Pass the net-id to the nova boot
− Create a port and pass it to nova boot
• Nova schedules the instance without much knowledge about the underlying network
• BP proposes to address this issue as one of the use cases
Network Assignment in Bridged VM’s
Rack 1N1
Rack 2N2
Rack 3N3
Rack 4N4
FZ1 FZ2 FZ3 FZ4
Network Tagging
26
VM
• Know your requirements
• Understand your size and scale
• Pick the SDN controller based on your needs
• Design with multiple failure domains
• Overlay, Bridged or Hybrid
• Monitor your cloud for performance degradation
Summary