Decoupling Connectivity from Routing

1
Decoupling Connectivity from Routing

• Karthik Lakshminarayanan
• UC Berkeley
• Joint work with
• Tom Anderson, Scott Shenker and Ion Stoica
• (UW and UC Berkeley)

2
Motivation

• Basic goal of routing provide connectivity
• recover from failures
• resistance to malicious nodes
• What undermines connectivity?
• flexibility of protocols thousands of knobs
• failure recovery tied to protocol convergence
  mechanisms
• Can we have routing protocols that are flexible,
  yet robust to failures and misconfigurations?

3
Basic Idea Decouple Connectivity from Routing

• Two-tiered routing
• lower layer Basic Connectivity Routing Protocol
  (BCRP)
• build sophistication on top
• Sophisticated routing fails ? fallback on BCRP
  routing

4
Basic Idea Decouple Connectivity from Routing

• Two-tiered routing
• lower layer Basic Connectivity Routing Protocol
  (BCRP)
• build sophistication on top
• Sophisticated routing fails ? fallback on BCRP
  routing

5
Key Insights behind BCRP

• Decouple failure recovery from protocol
  convergence mechanisms
• Ensure consistent network information across all
  nodes
• Append failure information in packets

6
1. Decoupling Failures from Convergence

• Traditionally protocols spread failure
  information across network
• conservative about declaring failed links to
  prevent oscillations
• Hint about failure known long before failure is
  confirmed
• e.g., single lost HELLO message
• can take local action and avoid the potentially
  failed link

7
2. Consistent Network Information

• Set of links in a managed network are stable over
  time
• e.g., intra-domain ISP networks
• Centralized node aware of the links (and their
  costs) of the network
• applies to ISP networks
• Central node disseminates the map (called
  consistent map) to all nodes in the network

8
3. Failure Information in Packets

• How does a node react when the next-hop is
  detected as failed?
• Recompute path to destination using M L, where
  M consistent map, L failed link
• Add L to the packet header
• packet contains all failed links it experiences
• downstream nodes know that L should not be used
  for routing this packet

9
Properties of BCRP (1 of 2)

• Guaranteed reachability If there is a path in
  the consistent map between S and D, then BCRP
  will route a packet from S to D
• When a failed link L is encountered
• either there is no path to D in M Lp L (where
  Lp is packet failure info)
• or graph over which path is computed
  monotonically decreases

10
Properties of BCRP (2 of 2)

• Security property As long as the correct path
  of a packet does not traverse any malicious node,
  BCRP will successfully route the packet
• Each packet is treated independently
• No routing information is disseminated
• cannot sink packets arbitrarily

11
Reducing Computation Overhead

• Eliminate per-node route computation
• Add source route to packets
• Eliminate per-packet route computation
• Precomputation

12
Precomputation

• Node N, for each of its outbound links L
• precomputes routes to all destinations using ML
• can be used when link L fails
• When is recomputation needed?
• packet might have traversed failed links before
• Recomputation is not needed as long as the
  precomputed path does not traverse any of the
  failed links contained in the packet

13
Precomputation
14
Reducing Computation Overhead

• Eliminate per-node route computation
• Add source route to packets
• Eliminate per-packet route computation
• Precomputation
• Reduce recomputation time
• Leverage incremental recomputation literature

15
Applicability

• Intra-domain routing
• Inter-domain routing
• DHT routing
• Wireless routing