1. Martin Casado, Nate Foster, and Arjun Guha CACM, October 2014
Abstractions for SDN
Martin Casado, Nate Foster, and Arjun Guha
CACM, October 2014

2. Introduction
SDN was invented to addressing some long-standing challenges in networking
SDN starts from two ideas
generalize network hardware to provide a standard collection of packet-processing functions instead of a fixed set of narrow features
decouple software that controls network from devices that implement it
The design of SDN
makes it possible to evolve network without having to change underlying hardware
enables expressing network algorithms in terms of appropriate abstractions for particular applications

3. Introduction
One or more controller machines execute general-purpose programs that respond to events (changes in network topology, connections initiated by end hosts, shifts in traffic load, or messages from other controllers) by computing a collection of packet-forwarding rules
Controllers then push these rules to switches, which implement required functionality efficiently using specialized hardware

4. New Applications Shortest-path routing Policy-based access control
Adaptive traffic monitoring
Traffic engineering
Network virtualization
without new (distributed) protocols on proprietary hardware
Need new abstractions to facilitate new applications
(like OS providing rich abstractions for managing resources, need new abstractions for network applications)

5. Network-Wide Structures
SDN controllers can compute network-wide structures that give global visibility into network state → simplify networking applications
Network Information Base (NIB) in controller
evaluating Dijkstra's shortest-path algorithm over the structure representing the topology
spanning tree: difficult with distributed algorithms vs. Prim’s algorithm

6. Distributed Updates Configuration updates to be eventually consistent
if network configuration is recalculated due to link failure, a packet may traverse a switch once in original state and a second time in updated state → forwarding loops or dropping packets
eventually consistent updates do not always suffice in SDN
an SDN controller might manage both filtering rules and forwarding rules, and these rules may be critical for ensuring invariants such as access control or isolation between traffic of tenants sharing network → invariants can be violated during periods of transition

7. Modular Composition
In OS, processes allow multiple users to share hardware resources; processes interact via well-specified interfaces
SDN controller as network OS, but lacks abstraction analogous to process
Network programming needs modularization
SDN building blocks = {forwarding, broadcast, monitoring, access control, …}

8. Virtualization (1)
Decouples software that controls network from the underlying forwarding hardware
Does not decouple forwarding logic from underlying physical network topology
SDN controllers now provide primitives for writing applications in terms of virtual network elements
Decoupling programs from topology also creates opportunities for making SDN applications more scalable and fault tolerant

9. Virtualization (2) Access control Multi-tenant datacenter
encoding MAC or IP addresses into configuration
topology changes (due to mobility) undermines security
when access control lists are configured in terms of a virtual switch connected to each host, then policy remains stable even if topology changes
Multi-tenant datacenter
allow multiple tenants to impose different policies on devices in a shared physical network
overlapping addresses and services lead to complicated forwarding tables → hard to guarantee that traffic generated by one tenant will be isolated from other tenants
using virtual switches, each tenant can be provided with a virtual network they can configure however they like without interfering with other tenants

10. Virtualization (3) Scale-out router
in large networks, it can be necessary to make a collection of physical switches behave like a single logical switch
for example, a large set of low-cost commodity switches could be assembled into a single carrier-grade router.
besides simplifying forwarding logic for individual applications, this approach can also be used to obtain scalability—because such a router only exists at the logical level, it can be dynamically augmented with additional physical switches as needed
Virtualization makes applications more portable and scalable by decoupling forwarding logic from specific physical topologies

11. Virtualization (4) Virtualization abstraction Virtualization mechanism
e.g., VMware’s NSX, Frenetic languages,
Virtualization mechanism
hypervisor

12. Formal Verification (1)
Manual low-level network configuration → unreliable and/or insecure networks
SDN standardizes interface to network hardware → tools to build and operate reliable networks
Network invariants: properties can be checked automatically using tools that formally model the state of network and controller
What properties?
Many properties are topology-specific
they can be stated and verified given a model of the structure of network

13. Formal Verification (2)
Topology-specific properties
Connectivity: eventually packets are delivered except due to congestion or failures
Loop freedom: no loop in network
Waypointing: packets emitted by untrustworthy hosts traverse middlebox that scans for malicious traffic before being forwarded to intended destinations
Bandwidth: provide minimum bandwidth specified in SLA
Topology-agnostic properties or properties for large classes of topologies [correctness criteria]
Access control
Host learning
Spanning tree

14. Formal Verification (3)
Both types of properties are difficult to establish in traditional networks, as they require reasoning about complex state distributed across many heterogeneous devices
Uniform interfaces provided by SDN simply verification
To verify configurations
model both topology and switch configurations
e.g., Header Space Analysis, FlowChecker, Antester, VeriFlow, NetKAT
To verify controllers (control programs)
e.g., NICE, NetCore, VeriCon, Flowlog

15. Formal Verification (4)
Need for tools that can provide rigorous guarantees about behavior, performance, reliability, and security of networked systems
By standardizing interfaces for controlling networks, SDN makes it feasible to build tools for verifying configurations and controllers against precise formal models
Future directions
developing custom logics and decision procedures for expressing and checking properties
enriching models with additional features such as latency and bandwidth
better integrating property checking and debugging tools into SDN controller platforms