This document proposes an approach called logical xBars (LXBs) to scale SDN control to extremely large networks. LXBs introduce hierarchy and recursion by designing the network as a set of programmable forwarding units (LXBs) that can be implemented in terms of smaller LXBs. This allows SDN control to scale while preserving benefits like straightforward convergence, failure handling, and transactional network updates. LXBs can match the physical network topology and allow local policies within domains. Computation is distributed across multiple controller tiers to keep it manageable as the network scales enormously in size.

3. Extending SDN to Large-Scale
Networks
Murphy McCauley, Aurojit Panda, Scott
Shenker
Martin Casado, Teemu Koponen
VMware
UC Berkeley

4. Given a desired
policy or behavior,
how do you realize it
on the physical
network?
really large physical
network?
The Problem
Policy / Behavior
(Routing, ACLs, TE, etc.)
OpenFlow
Controller
Physical Network
(Switches, links, etc.)
Geographically distributed
10,000 or 100,000 node

5. Possible Approaches
• Beefy Central Controller
– Eventually breaks down
– Bad failure locality
– Unclear where local policy fits
• Shard and Distribute
– Requires independent shards
• Fully Distribute
– Every new problem requires new distributed algorithm
– Abandons benefits of SDN!

6. Where does that leave us?
• None of the standard approaches allow
SDN to scale to arbitrarily large networks
• Combine classic network scaling
approaches with SDN in a new way:
– Utilize aggregation and hierarchy
– Structure using a recursive abstraction

7. Old Techniques in New Setting
• Design built around Logical xBars (LXBs)
– A programmable unit of forwarding (“switch”)
– Implemented in terms of itself
• Results in design with:
– Scalability to extremely large networks
– Straightforward convergence/failure behavior
– Transactional network changes

8. xBar Basics
Ports
Switch
Link
Controller

9. xBar Basics
A
D
B
C
sw2
sw1
sw3
Table
in:A → out:D
sw1 Table
in:A → out:B
sw3 Table
in:C → out:D
Table
Table on Topology
lxb1

10. xBar Basics
lxb2lxb1 lxb3

11. Hierarchy & Recursion

12. Hierarchy & Recursion

13. Hierarchy & Recursion
State
Config

14. Hierarchy & Recursion
State
Config

15. LXB Properties
• Hierarchy cleanly stacks arbitrarily high
– Can match LXBs to physical/organizational
boundaries
– You can express local policies within “your”
LXB as long as the externally-visible policies
conform to those given by your parent

16. LXB Properties
• Computation kept in check
– 10,000 nodes in four tiers: individual LXBs
only have 10 node graphs
– Multiple levels of aggregation means that
computations can be kept towards lower
(most distributed) portion of controller
hierarchy

17. LXB Properties
• “SDN-Friendly”
– Agnostic to choice of switch control
mechanism (e.g., OpenFlow) and “bottom tier”
controller (e.g., Floodlight, NOX, etc.)

18. Status
• We’ve looked at:
– Routing (several variations), ACLs, TE
• Stuff we haven’t looked at yet:
– Multicast, anycast, … ?
• What we’d love to know:
– Characteristics of actual large networks
– Pain points for large networks – what are you hoping
SDN can solve for you?

19. murphy@berkeley.edu
Thanks!