Transcript Routing Scalability

Fabric: A Retrospective on
Evolving SDN
Presented by: Tarek Elgamal
Outline
What is ideal Network design ?
What are the SDN Limitations ?
What is Multiprotocol label switching (MPLS) ?
How the MPLS design address the SDN limitations ?
How can we extend SDN to support MPLS ideas ?
Ideal Network Design
Ideal Hardware:
• Simple: Hardware should be inexpensive to build and operate
• Vendor-neutral: Users easily switch between hardware vendors
without complex upgrades
• Future-proof: Users should not need to upgrade hardware
unnecessarily
Ideal Software
• Flexibility: supports configuration for a wide variety of
requirements (e.g., isolation, security, traffic engineering, routing
algorithms)
“Today’s Internet does not satisfy any!”
MPLS
SDN
Ideal Network Design
Host-Network Interface:
SDN
How the hosts inform the network of
their requirements
Operator-Network Interface
Operator-Network
Interface
Host-Network
Interface
How operator (ISP) controls the
network infrastructure
Packet-switch Interface
What fields the switch needs from the
packet to forward it.
Ideal network design should
separate between the three
MPLS
Packet-Switch
Interface
Multi-protocol Label Switching (MPLS)
•
•
•
Data packets are assigned labels that correspond with paths through
the network instead of end points
Routers follow the predefined labelled path
Eggress router
Routers propagate dynamic changes in the network
Removes label
delivers packet
and paths are updated accordingly
Label
IP Packet
Label
IP Packet
IP Packet
IP Packet
Ingress router
receives and labels
packet
Switching router:
Only does switching
based on the label
Advantages of MPLS
•
•
Distinction between network edge and network core
• Core hardware become much simpler
MPLS labels decoupled from host protocol (e.g. IPv4)
• Distinguishes between Host — Network
and Packet— Switch Interfaces
Label
IP Packet
Ingress router
receives and labels
packet
IP Packet
Label
IP Packet
Eggress router
Removes label
delivers packet
IP Packet
Network Interfaces
Host-Network Interface:
How the hosts inform the network of
their requirements
Operator-Network Interface
How operator controls the network
infrastructure
Packet-switch Interface
What fields the switch needs from the
packet to forward it.
Ideal network design should
separate between the three
SDN
Operator-Network: OpenFlow
Host-Network: IPv4
Same
Packet-Switch:
IPv4
Extended SDN
Operator-Network
Interface
Host-Network
Operator-Network:
OpenFlow
Interface
Host-Network:
IPv4
Packet-Switch: MPLS ideas
MPLS
Host-Network: IPv4
Packet-Switch: MPLS (packet
routing through label only)
Packet-Switch
Interface
SDN vs MPLS
Ideal Hardware:
• Simple: Hardware should be inexpensive to build and operate
• SDN forward on IPv4 headers (100s of bits)
• MPLS forwards on the label only (10s of bits)
• Vendor-neutral: Users easily switch between hardware vendors without
complex upgrades
• SDN uses today’s routers which has rich set of features
• MPLS routers only need to forward packets on a predefined path
(minimal set of features would suffice)
• Future-proof: Users should not need to upgrade hardware unnecessarily
• SDN has to deal with host’s protocol (e.g. IPv4)
• MPLS decouples edge protocol from core protocol
(IPv4  IPv6 will not affect core of the network)
Extended SDN
Network components:
• Host, Edge, Fabric (switch for basic
packet transport only)
• Two logical controllers (edge and
fabric controllers)
Network Interfaces:
• Host – Network : Ingress edge switch
• Operator– Network : Edge controller
• Packet– Switch: Fabric elements and
controller
Extended SDN - Design Principles
Separation of forwarding
• Minimal set of forwarding
primitives
Separation of control
• Two separate roles:
• Fabric transports packet
• Switch provides security,
Mobility, and other services
The main principle is to allow edge and
core technologies to evolve separately
Fabric Component
• Interconnects the edge as fast and cheaply as possible.
• Supports any number of edge designs (i.e., different addressing
schemes and policy models)
• Supports point-to-point and point-to-multipoint communication
• Have packet-dropping mechanism under congestion
• Should not support filtering ,isolation, stateful flow tracking
Fabric Path Setup
Two suggested approaches
• Forwarding on pre-calculated paths
• Reduces provisioning overhead, if both the edge and the fabric are
part of the same administrative domain
• Explicit provisioning of the path by the provider (ingress and egress
switches exchange control messages to establish path with specific
requirements, then packet transfer starts)
• Used with customer-provider relationship
Mapping between Edge/Fabric Addresses
Two suggested approaches
• Address translation: edge addresses are replaced with fabric internal
addresses, translated back to edge addresses at the destination
• Downside: Couples the edge and core addressing schemes
Fabric
Header
IP Header
IP Header


• Encapsulation: encapsulated with another header that carries the
Fabric internal addresses
• More general and decouples addressing schemes
IP Header

Fabric
IP Header
Header
Discussion
Can SDN with OpenFlow be used to implement MPLS instead of
changing the whole architecture ?
What are some suggested changes to OpenFlow to adopt this
architecture ?
Do you think simplifying the hardware infrastructure achieves
vendor-neutrality ?