1. www.mpls2012.com MPLS And The Data Center Adrian Farrel Old Dog Consulting / Juniper Networks adrian@olddog.co.uk afarrel@juniper.net

2. Agenda What do I mean by “Data Center”? Design goals and requirements Handling mobility within the data center Connectivity between data center sites Can MPLS add value? 2

3. Everyone’s Data Center is Different There are some common fundamental concepts Racks of servers VMs hosted on blades VMs connected On server In rack In DC In other DCs Connectivity to the external services 3 IP/MPLS Network Storage DC FW LB VM-based Appliances NAT VMs on Server Blades VM VSw Top of Rack Switch Services L2 L3

4. Design Goals Provide separate logical tenant networks in Data Center over common IP physical infrastructure Design Goal: 100K tenants, 10M Virtual Machines (VMs) Need a data plane encapsulation Examples exist Virtual Extensible Local Area Networks (VXLAN) Network Virtualization using Generic Routing Encapsulation (NVGRE) Discovery is needed Data plane learning seems popular ARP doesn’t scale and needs to be suppressed Maybe the control plane can help A control plane is also required Static configuration is a solution (Hypervisor with SDN?) A control plane can make life a lot easier 4

5. Multi-Tenancy : Requirements Multi-tenancy has become a core requirement of data centers Including for Virtualized Machines (VMs) and VM multi-tenancy It prooves a real stretch Three key requirements needed to support multi-tenancy are Traffic isolation Address independence Fully flexible VM placement and migration IETF’s NVO3 WG considers approaches to multi-tenancy that reside at the network layer rather than using traditional isolation (e.g., VLANs) An overlay model to interconnect VMs distributed across a data center We already have network layer overlay solutions More about this later 5

6. Mobility Virtual Machines need to be moved between blades How often? Dynamic load balancing Planned service Failure recovery How much? Blades, servers, racks How seamless? Application re-start Packet loss Hitless Challenges are recovery/preservation of connectivity VMs need to preserve identity L2 or L3? Need rapid location discovery/advertisement 6

7. Inter Data Center Connectivity Many reasons for connectivity Applications in different DCs need to talk VMs may be gathered into VPNs (virtual VPNs?) One application’s data might be stored in anther DC Stored data has to be synched between DCs Connectivity between DC sites is like VPN connectivity Except it may be “tunnelling” virtual VPN connectivity And, of course, connectivity to the outside world 7

8. What do we Mean by MPLS? Odd time and place to be asking this question MPLS offers a versatile encapsulation technique Small headers Nested encapsulation Simple forwarding Special meaning labels MPLS provides a range of control plane protocols These have different applicabilities Some are more complex than others Supports static configuration 8

9. The E-VPN Designed for scalability and ease of deployment Provider Edge (PE) can be in ToR switch and/or Hypervisor Operator defined networks – mesh, hub & spoke, extranets, etc Control plane learning using BGP VM Mobility – all PEs know VM’s E-VPN location VPN and Virtual LAN auto-discovery ARP flood suppression Control-plane scaling using Route Reflectors, RT Constrain, ESI, MAC aggregation Control & data plane traffic for VPNs only sent to PE with active VPN members Scalable fast convergence using Block MAC address withdrawal Support for MAC prefixes (e.g., default MAC route to external DC) Broadcast & Multicast traffic over multicast trees or ingress replication Active/active multi-homing CE sees LAG, PEs see Ethernet Segment (set of attachments to same CE) 4B tenant VPNs, 4B virtual LANs per tenant VPN 9

10. MPLS E-VPN Routes MAC Advertisement Route Distributes MAC & IP address to PE & MPLS label binding Per EVI Ethernet AD Route Distributes Ethernet Segment to PE & MPLS label binding Used in active/active multi-homing Both carry a 24 bit MPLS label field Use of MPLS label is very similar to VNID but supports local significance Distribute VNID in MPLS label field Either global or local significance Local significance allows it to represent EVI, Port, MAC address, or MAC address range Data plane encapsulation specified using Tunnel Encapsulation attribute (RFC 5512) Distributed with both of the above routes 10

11. E-VPN is Encapsulation Agnostic E-VPN Instance can support multiple data plane encapsulations (MPLS, VXLAN, NVGRE, etc.) MPLS encapsulation is just one option Encapsulations advertised in BGP, ingress uses encapsulation supported by egress This use of BGP is not complicated Broadcast & multicast use encapsulation-specific shared trees Allows interoperability with existing E-VPN & L3VPN deployments This makes inter-DC really easy 11

12. Is MPLS The Answer? What was the question? Do we need another control plane protocol? Why can’t we use what we already have? Frankly, BGP is not that hard and does what we need Can we integrate the DC with the outside world? Gateways, tunnelling and encapsulation are always possible Protocol gateways are a bit of a mess E-VPN and L3VPN connectivity just works Do we need another L2 encapsulation? There are plenty available, just pick your favorite This is an MPLS conference 12

13. Questions? afarrel@juniper.net adrian@olddog.co.uk 13