Feedback Based Routing By Dapeng Zhu, Mark Gritter, and David R. Cheriton
Outline Problems with BGP Our solution Analysis Applications Related Work
Problems with BGP Vulnerability Scalability Convergence Time
Feedback Based Routing Overview of system Structural Information Propagation Algorithm for Access Routers
Analysis Attack Resistance Terrorist Black Holes Check! Bogus SYN-ACK response Nothing to see here, moving along
Scalability Scales better than BGP for three reasons Route computation/propagation removed from critical path Availability of routing system does not depend on in-time computation of shortest paths Requirements on transit routers are substantially reduced
Scalability (cont) This last point is huge! Routing system for the Internet backbone should not be dependent on the exponential growth at the edge. “Back of the envelope calculation” shows something like a mere 50mb to store an entire routing table with our scheme
Applications Defend against The Terrorists(tm) and DoS attacks Our scheme would allow for people to throttle DoS traffic without having to contact upstream providers Recognition of a pattern in the DoS traffic is the only requirement.
Applications (cont) Virtual Links with Zero Failover Time BGP currently has terrible convergence time after a link failure We propose highly available virtual links with zero failover time
Overview Separate structural and dynamic information Core Role Forward Packets Propagate structural information Edge Role Routing Decisions End to end Probing
Structural Information Propagation Edges associated with timer Renewed with announcement Removed from structure at expiration Three rule sets for packet forwarding Positive Negative Traffic Engineering
Algorithm for Access Routers Determine two disjoint routes Measure RTT TCP ICMP Periodic Renewing of Backup Routes
Related Work Differences between previous “Byzantine robustness” strategies and ours: Transit routers don't know network topology, making transit routers almost independent of network growth Since most network traffic is TCP, we use TCP SYN and SYN ACK packets as a measure of network performance We are concerned about scalability...they clearly were not.
Related Work (cont) Resilient Overlay Network Overlay network that tries to “get around” routing failures. We believe this is not sufficient. These only work when there are isolated routing failures An overlay network could not function during a The Terrorists(tm) Black Hole attack. Think about it.
Related Work (cont) RouteScience, Eye Networks, netVmg They do provide possible performance enhancement (Are these the people spamming me about Ciali$?!) Since they are edge-only, however, they do not shield against widespread network failure
Conclusion Separation of Performance Information and Structural Information Routing in the backbone is reduced purely to Structural Information Access routers maintain more than one route. Helps fight The Terrorists(tm)
The End Questions? Comments? Forfeit from the losing Offense team?