Efficient Internet Routing with Independent Providers

Transcription

1 Efficient Internet Routing with Independent Providers David Wetherall University of Washington

2 Credits Ratul Mahajan, Microsoft Research Tom Anderson, University of Washington Neil Spring, University of Maryland djw / nicta / dec

3 Independent providers affect routing Telstra Sprint ATT iinet Direct paths in the Internet aren t always short paths! djw / nicta / dec

4 Competing interests are fundamental A basic feature of the Internet: Providers carry traffic for their competitors Flows compete for available bandwidth Pervasive in other computing systems: Interference in wireless networks Freeloading in peer-to-peer systems , denial-of-service, djw / nicta / dec

5 A long-standing issue The most important change in the Internet architecture over the next few years will probably be the development of a new generation of tools for management of resources in the context of multiple administrations. -- David Clark, 1988 (DARPA design paper) djw / nicta / dec

6 and still an open issue Our position is that accommodating this tussle [in which different stakeholders vie to favor their particular interests] is crucial to the evolution of the network s technical architecture. -- David Clark et. al., 2005 (Tussle paper in TON) djw / nicta / dec

7 djw / nicta / dec

8 This talk 1. How does the existence of independent providers affect routing? An empirical study of BGP 2. Can we design routing protocols that work well with competing interests? WISER, a new protocol based on barter djw / nicta / dec

9 BGP Internet routing game Goal is to choose interconnection for traffic. Each provider: Selects the egress they prefer, e.g., nearest exit Agrees to accept packets from all ingresses Has little visibility into the workings of others djw / nicta / dec

10 Consequences of the BGP game 1. Asymmetric paths 2. Inefficient paths the price of anarchy 3. Unstable paths djw / nicta / dec

11 An example of inefficiency Seattle ATT San Francisco Sprint overload Paths are longer because ATT lacks insight into Sprint s network djw / nicta / dec

12 Estimating overall routing inefficiency 1. Get a detailed map of the Internet Provider topology, interconnection, traffic 2. Compute the routes selected by BGP Across multiple providers 3. Compute the most efficient routes 4. Find the difference between them Unfortunately, maps are hard to come by because of independent providers! djw / nicta / dec

13 Rocketfuel: an Internet mapping engine Probed Internet paths from monitors to many locations Merged found paths and annotated high-level maps Result is city-level maps of 55 diverse providers For our purposes: approximate latency from geography, traffic from population/gravity models Sprint, USA Telstra, Australia djw / nicta / dec

14 Price of anarchy for BGP cumulative % of of flows flows latency path length reduction inflation relative to social optimal % path length inflation Most paths are good but some are very poor djw / nicta / dec

15 Why inefficiency matters Poor paths (the tail of the curve) lead to operator intervention, which is undesirable. Costly, error-prone and difficult ( tweak and pray ) Ad hoc cooperation ( ) used in practice Likely to become worse as QOS takes off Even a small average inflation represents wastage Aim to carry traffic just as well with a smaller network and save money djw / nicta / dec

16 djw / nicta / dec

17 Overall goal for WISER, a new protocol optimum price of anarchy Route Efficiency social BGP Provider Freedom Want high efficiency and freedom for providers to make their own decisions djw / nicta / dec

18 Making WISER practical 1. Limited information disclosure: commercial interests 2. Independent optimization criteria: business strategies Use agnostic costs 3. Win-win outcomes: no one should lose or losers won t play Barter: ISPs lose on some flows and gain on others djw / nicta / dec

19 Approaches that don t work Unilateral optimization (automated tweak and pray ) Computing the social optimal Traditional mechanism design Using real money as the basis for optimization djw / nicta / dec

20 Example of barter with agnostic costs hops ms ms 2 2 hops ms 6 1 hop 4 3 Both providers win when each trades small losses on some traffic for bigger gains on other traffic djw / nicta / dec

21 WISER Internet routing game 1 7 D [1] D S 2 D [3] 1 D [11] 3 11 Goal is to select the interconnection for traffic 1. Providers determine the agnostic costs of internal paths 2. Downstream (receiver) announces its cost over each link 3. Upstream (sender) selects based on local and remote costs djw / nicta / dec

22 Incentinghonest cost disclosure honest disclosure inflated disclosure normalized costs Normalize costs so both providers have an equal say also helps with inherently unequal cost ranges djw / nicta / dec

23 Incentinghonest use of remote costs Upstream makes a virtual payment to the downstream a rough balance of payment per byte should emerge when both providers are sensitive to received costs Providers contractually stipulate this balance djw / nicta / dec

24 Extending WISER to many providers c3 = c1 l + path cost D [c3] c1 l D [c1] D [c4] S D D [c2] c2 l D [c5] Each pair of providers run WISER using exchange rates to convert costs between providers djw / nicta / dec

25 djw / nicta / dec

26 Evaluating WISER Efficiency Is the price of anarchy consistently low? estimate using the same maps Robustness to dishonesty Can a dishonest provider harm an honest one? study with different provider strategies Implementation Is the overhead low? Is it easy to deploy? prototype to flesh out details (XORP, SSFnet) djw / nicta / dec

27 Price of anarchy for WISER BGP anarchy WISER Wiser latency path length reduction inflation % length inflation BGP WISER cumulative % of of flows flow relative to social optimum WISER is close to the social optimum (y-axis) win-win strategy is efficient for real networks barter between each pair of providers works well djw / nicta / dec

28 WISER needs less provisioning % of ISPs 0 Wiser anarchy WISER BGP overprovisioning (%) cumulative % of proviers overprovisioning (%) Overprovisioningmeasures how much extra capacity providers need to handle failures WISER allows smaller networks to carry a given traffic load just as well djw / nicta / dec

29 Social optimum is not win-win % of ISPs ISP gain (%) Wiser optima average distance reduction (%) relative to BGP WISER Social cumulative % of providers Under WISER, some providers don t win as much, but they don t significantly either djw / nicta / dec

30 Barter on small sets is not sufficient anarchy flow-pair bar Wiser BGP BGP on flow pairs WISER cumulative % of flows % of flows path length inflation (relative to optimum) multiplicative inflation Bartering over many routes (WISER) helps providers trade small loses for large gains djw / nicta / dec

31 Ordinal preferences aren t efficient anarchy ordinal preferen Wiser BGP WISER with BGP MEDs WISER cumulative % of flows % of flows multiplicative path length inflation (relative to optimal) BGP MEDs(ordinal) aren t sufficient for efficient routing; agnostic costs in WISER (cardinal) are djw / nicta / dec

32 djw / nicta / dec

33 Robustness to dishonesty Honest behavior is not provably the best shortterm option unlike classic mechanism design But we expect cheating to be uncommon in the long-term Harm to reputations, contract termination Yet WISER should minimize the gains of dishonesty Study the impact of a dishonest provider djw / nicta / dec

34 Effects of dishonesty = all honest providers (WISER) = one dishonest provider (WISER) = one dishonest (no constraints) % of % of honest providers ISPs provider ISP gain gain (%) (%) provider ISP gain gain (%) % of % of dishonest providers ISP Honest provider in different scenarios. Its loss is small. Dishonest provider in different scenarios. Its gain is small. djw / nicta / dec

35 Implementation as a BGP extension c3 = c1 l + path cost D [c3] c1 l D [c1] D [c4] S D D [c2] c2 l D [c5] Costs are carried with BGP messages and used in the decision process. Border routers communicate to track exchange rates and log virtual payments. djw / nicta / dec

36 Low deployment barrier Implementation mirrors the current routing protocol; little disruption Nature of existing provider peering contracts is preserved Bilateral contracts with simple charging model Incrementally deployable start with a pair of providers who benefit djw / nicta / dec

37 Low overhead WISER BGP Wiser BGP messages per second cumulative % of % failures of experiments XORP prototype added <6% to BGP code and 15%-25% CPU The sample graph above shows that WISER sends a similar number of routing messages as BGP djw / nicta / dec

38 djw / nicta / dec

39 Conclusions optimum WISER Route Efficiency social BGP Provider Freedom WISER delivers high efficiency for independent providers in a practical protocol. It is an existence proof! Our future work is to move along the road to deployment We hope to do the same for other networking problems djw / nicta / dec

40 For more information Negotiation-based Routing Between Neighboring Domains, Mahajan, Wetherall and Anderson, NSDI Quantifying the Causes of Path Inflation, Spring, Mahajan, and Anderson, SIGCOMM Measuring ISP Topologies with Rocketfuel, Spring, Mahajan, Wetheralland Anderson, TON 2004 [Bennett Prize] Questions? djw / nicta / dec