Interdomain Routing and Games Hagay Levin, Michael Schapira and Aviv Zohar The Hebrew University 1 On the Agenda • Motivation: Are Internet protocols incentive compatible? • Interdomain routing & path vector protocols • Convergence issues • BGP as a game • Hardness of approximation of social welfare • Incentive compatibility • Conclusions 2 Are Current Network Protocols Incentive Compatible? • Protocols for the network have been dictated by some designer • Okay for cooperative settings • But what if nodes try to optimize regardless of harm to others? • Example: TCP congestion control – Requires sender to transmit less when the network is congested – This is not optimal for the sender (always better off sending more) 3 Secure Network Protocols • A lot of effort is going into re-designing network protocols to be secure. • Routing protocols are currently known to be very susceptible to attacks. – Even inadvertent configuration errors of routers have caused global catastrophes. • Designers are also concerned about incentive issues in this context. • Our work highlights some connections between incentives and security of BGP. 4 Interdomain Routing • Messages in the Internet are passed from one router to the other until reaching the destination. • Goal of routing protocols: decide how to route packets between nodes on the net. • The network is partitioned into Autonomous Systems (ASes) each owned by an economic entity. – Within ASes routing is cooperative – Between ASes inherently non-cooperative • Routing preferences are complex and uncoordinated. Always choose shortest paths. Load-balance my outgoing traffic. UUNET AT&T My link to UUNET is for backup purposes only. Comcast Qwest Avoid routes through AT&T if at all possible. 5 Path Vector Protocols • The only protocol currently used to establish routes between ASes (interdomain routing): The Border Gateway Protocol (BGP). • Performed independently for every destination autonomous system in the network. • The computation by each node is an infinite sequence of actions: receive routes from neighbors choose “best” neighbor send updates to neighbors 6 Example of BGP Execution 5 4 41d 41d 23d 1d 1 1d 23d 2 23d 23d 3d 3d 3 d d d d receive routes from neighbors choose “best” neighbor send updates to neighbors 7 Our Main Results Informally • Theorem: In “reasonable economic settings”, BGP is almost incentivecompatible (And can be tweaked to be incentive compatible). • Theorem: In these same settings it is also almost collusion proof. – To make it fully collusion proof we need a somewhat stronger assumption. 8 BGP – Not Guaranteed to Converge 1 12d 1d … 2d 2 23d 2d ... 12d 1d d 31d 3 31d 3d … • Other examples may fail to converge for certain timings and succeed for others. 9 Finding Stable States • Previously known: It’s NP-Hard to determine if a stable state even exists. [Griffin, Wilfong] We add: • Theorem: Determining the existence of a stable state requires exponential communication. • In practice, BGP does converge in the Internet! Why? 10 The Gao-Rexford Framework: An economic explanation for network convergence. Neighboring pairs of ASes have one of: • a customer-provider relationship • a peering relationship peer providers peer customers Restrict the possible graphs and preferences: • No customer-provider cycles (cannot be your own customer) • Prefer to route through customers over peers, and peers over providers. • Only provide transit services to customers. Guarantees convergence of BGP. 11 Dispute Wheels • A Dispute Wheel [Griffin et. al.] – A sequence of nodes ui and routes Ri, Qi. – ui prefers RiQi+1 over Qi. • If the network has no dispute wheels, BGP will always converge. • Also guarantees convergence with node & link failures. Gao-Rexford Shortest Path No Dispute Wheel Robust Convergence 12 Modeling Path Vector Protocols as a Game • The interaction is very complex. – Multi-round – Asynchronous – Partial-information • Network structure, schedule, other player’s types are all unknown. • No monetary transfers! – More realistic – Unlike most works on incentive-compatibility in interdomain routing. 13 Routing as a Game • The source-nodes are the strategic agents • Agent i has a value vi(R) for any route R • The game has an infinite number of rounds • Timing decided by an entity called the scheduler – Decides which nodes are activated in each round. – Delays update messages along selective links. 14 Routing as a Game (2) • A node that is activated in a certain round can – Read update messages announcing routes. – Send update messages announcing routes. – Choose a neighboring node to forward traffic to. • The gain of node i from the game is: – vi(R) if from some point on it has an unchanging route R. – 0 otherwise. (can be defined as the maximal gained path in an oscillation as well). • a node’s strategy is its choice of a routing protocol. – Executing BGP is a strategy. 15 Approximating Social Welfare 1 / 2 O n • Theorem: Getting an approximation to the optimal social welfare is impossible unless P=NP even in Gao-Rexford settings. (Improvement on a bound achieved by [Feigenbaum,Sami,Shenker]) • Theorem: It requires exponential communication 1 to approximate social welfare up to O n 16 Manipulating in The Protocol • A node is said to deviate from BGP (or to manipulate BGP) if it does not follow BGP. • We want nodes to comply with the alg. Otherwise, suffer a loss when they deviate • Which forms of manipulation are available to nodes? – – – – – Misreporting preferences. Reporting inconsistent information. Announcing nonexistent routes. Denying routes. … 17 No Optimal Protocols • Theorem: Any routing protocol that: 1. Guarantees convergence to a solution for any timing with any preference profile 2. Resists manipulation Must contain a (weak) dictator: A node that always gets its most preferred path. (Simple to prove using a variant of the Gibbard-Satterthwaite theorem) 18 • Suppose node 1 is a weak dictator. • If it wants some crazy 6 path, it must get it. • This feels like an unreasonable protocol. 5 4 3 7 2 1 d 19 Is BGP Incentive-Compatible? • Theorem: BGP is not incentive compatible even in Gao-Rexford settings. m1d m12d 12d 1d 1 m m1d m12d 12d 1d 1 m d d 2md 2d 2 without manipulation 2md 2d 2 with manipulation 20 Can we fix this? • We define a property: – Route verification means that an AS can verify that a route is available to a neighboring AS. • Route verification is: – Achievable via computational means (cryptographic signatures). – An important part of secure BGP implementation. 21 Incentive Compatibility • Theorem: If the “No Dispute Wheel” condition holds, then BGP with route verification is incentive-compatible in expost Nash equilibrium. • Theorem: If the “No Dispute Wheel” condition holds, then BGP with route verification is collusion-proof in ex-post Nash equilibrium. 22 Open Questions • Characterizing robust BGP convergence (“No dispute wheel” is sufficient but not necessary). • Does robust BGP convergence with route verification imply incentive compatibility? • Can network formation games help to explain the Internet’s commercial structure? • Maintain incentive compatibility if the protocol is changed to deal with attacks and other security issues? • How do congestion and load fit in? 23 Conclusions • Our results help explain BGP’s resilience to manipulation in practice. – Manipulation requires extensive knowledge on network topology & preferences of ASes. – Faking routes requires manipulation of TCP/IP too. – Manipulations by coalitions require Herculean efforts, and tight coordination. • We show that proposed security improvements would benefit incentives in the protocol. • Work in progress: other natural asynchronous games. – “Best Reply Mechanisms” with Noam Nisam and Michael Schapira 24