A BGPbased Mechanism for LowestCost Routing - PowerPoint PPT Presentation

1 / 28
About This Presentation
Title:

A BGPbased Mechanism for LowestCost Routing

Description:

V (c) = Tij Ik(c;i,j)ck. i,j. N k N. Model : Limitations. per-packet cost model ... Computation at node i in single stage : O(nd x degree(i)) Static environment ... – PowerPoint PPT presentation

Number of Views:25
Avg rating:3.0/5.0
Slides: 29
Provided by: joh9202
Category:

less

Transcript and Presenter's Notes

Title: A BGPbased Mechanism for LowestCost Routing


1
A BGP-based Mechanism for Lowest-Cost Routing
  • By Feigenbaum,Papadimitriou,Sami,Shenker
  • Presented by Namrata Rastogi

2
Presentation Layout
  • BGP Review
  • Comparison with Related work
  • Pricing Mechanism
  • Game Model
  • Distributed Price Computation
  • Open Problems and Conclusions

3
BGP Recap-I
  • Internet viewed as collection of Autonomous
    systems (AS).
  • Backbone of high bandwidth lines and fast routers
  • Hooked to these backbones are regional networks
  • Routing Algorithm within AS OSPF
  • Interior Gateway protocol
  • Routing between ASes BGP
  • Exterior Gateway protocol

4
BGP Recap-II
  • EGP, IGP and Multiple AS

AS 100
AS 200
Router A
Router B
EGP
EGP
IGP
Router D
Router C
AS 300
5
BGP Recap-III
  • Why do we need different protocols for
    inter-domain routing and intra-domain routing ?
  • Policies involving political,security and
    economic considerations
  • Example traffic starting or ending at IBM
    should not transit Microsoft.
  • AS willing to carry traffic for some who are its
    neighbors or have paid .
  • Policy based routing
  • Policies are manually entered into each BGP
    router (not part of protocol)

6
BGP Recap-IV
  • Path Vector protocol
  • Instead of maintaining cost to each destination ,
    BGP router keeps track of exact path used.
  • Periodically sends each neighbor the exact path
    it is using instead of just the cost to a
    particular destination. (incremental updates)

7
BGP Example
  • Information F receives from its neighbors about
    destination D
  • From B I use BCD -From E I use EFGCD
  • From G I use GCD
  • From I I use IFGCD

C
B
A
Solves count-to-infinity problem (how?)
D
G
F
E
H
I
J
8
Related work
  • Nisan and Ronen
  • Formulation of VCG pricing for routing
  • Hershberger and Suri
  • Centralized mechanism design
  • incentive issues in route selection
  • compute LCP and set of payments to links
  • lacked in simulation of real-world routing
  • Starting point of this work need for
    distributed algorithmic mechanism design (DAMD)

9
Why incentives?
  • AS incurs a per-packet cost for carrying traffic
  • Compensation payment for transit traffic
    (price)
  • Current implementation of inter-domain routing
    framework has no notion of incentives.
  • BGP computes LCP given a set of AS costs.
  • AS could lie !!
  • Lower than truthful
  • Higher than truthful
  • Pricing scheme such that no incentive to lie
    about costs.

10
Strategyproof Pricing Mechanism
  • Pricing scheme should be strategyproof
  • AS that carry no transit traffic receive no
    payment.
  • Member of VCG class of mechanisms.

11
More Realistic Model 3 ways
  • Strategic agents
  • Links intra-domain routing
  • Nodes inter-domain routing
  • computes LCP for all source destination pairs
  • Solve N2 LCP instances
  • Distributed algorithm for computation BGP
  • Backward compatibility with existing internet
    protocol

12
Network Model Game
  • Routing Game

Strategic agents AS nodes
How to play? -node has to report transit cost
V (c) ? Tij ? Ik(ci,j)ck
k ? N
i,j ?N
Only way to minimize the overall cost is
incentivize the players to report true costs
VCG pricing!
Centralized vs Distributed Mechanism Design
13
Model Limitations
  • per-packet cost model
  • Transit cost administrative in nature
  • BGP general policy based routing
  • LCP not commonly used policy
  • no transit restrictions
  • no general path costs in BGP currently
  • our model assumes actual LCP (not based on
    number of AS hops)

14
Assumptions
  • biconnected graph
  • cost independent of which neighbor k received the
    packet from and which neighbor k sends the packet
    to
  • nodes carry no transit traffic receive no transit
    payment
  • Cost of only transit node counted
  • At most one link (bidirectional) between two ASs
  • Intra-domain issues ignored

15
Pricing Mechanism
  • The following two requirements for pricing
    mechanism uniquely determines the mechanism we
    must use.
  • strategyproof
  • nodes with no transit traffic receive no payment
  • payments as per-packet price depending on source
    and destination
  • Pk ? Tijpijk
  • where pijk is the per-packet price paid to
    node k for each transit packet from i to j.

i,j ? N
16
Theorem 1.
  • When routing picks lowest cost paths, and the
    network is biconnected, there is a unique
    strategyproof pricing mechanism that gives no
    payment to nodes that carry no transit
    traffic.The payments to transit nodes are of the
    form
  • pk ?i,j? N Ti,jpijk
  • where pijk ckIk(ci,j) ?Ir(ck?i,j)cr -
    ? Ir(ci,j)cr
  • Proof ?

r?N
r?N
17
Claims of proof
  • Payments do not depend on traffic matrix.
  • prices pij are zero if LCP between i and j does
    not traverse k.
  • prices depend on source and destination unlike
    costs
  • Payment to a node k for a packet from i to j is
    determined by the cost of LCP and the cost of the
    lowest-cost path that does not pass through k.

18
AS Graph sample calculation
Traffic X to Z LCP XBDZ Payment to
D (depends on LCP and lowest-cost k avoiding
path) Lowest-cost D avoiding path is XAZ
(transit cost 5) pd cD 5 - 3 3 Pb
cB 5 - 3 4
Z Cz4
D cD1
A cA5
Y cY3
B cB2
X cX2
19
BGP Computational Model
  • Every node i stores lowest-cost AS path for
    each j.
  • A router stores
  • O(nd) AS numbers , where d is the diameter of the
    network.
  • O(n) path costs
  • Routing table updates (incremental)
  • Computation of a single router infinite
    sequences of stages
  • complexity measures
  • number of stages for convergence
  • total communication

20
Assumptions
  • Nodes run synchronously
  • BGP converges within d stages of computation.
  • Each stage involves O(nd) communication on any
    link.
  • Computation at node i in single stage O(nd x
    degree(i))
  • Static environment

21
Distributed Price Computation
Loop-free Tree formation
Z Cz4
Z Cz4
D cD1
D cD1
A cA5
A cA5
Y cY3
Y cY3
B cB2
B cB2
X cX2
X cX2
Tree T(Z)
AS Graph
22
Distributed Price Computation
  • 4 cases

Z cz4
  • a is is parent in T(j)
  • pijk lt pajk

ii)a is is child in T(j) pijk lt pajk ci
ca
D cD1
A cA5
Y cY3
iii)a is not adjacent to i in T(j) and k is on
P(ci,j) pijk lt pajk ca c(a,j) - c(i,j)
B cB2
X cX2
iv)a is not adjacent to i in T(j) and k is not
on P(ci,j) pijk lt ck ca c(a,j) - c(i,j)
Tree T(Z)
23
Lemma 1
  • Let ib be the first link on P-k(ci,j).Then, the
    corresponding inequality (1) to (4) attains
    equality for b.

P-k(ci,j) lowest-cost k avoiding path
Pk(ci,j) LCP from i to j for vector of
declared costs c b neighbor of i on
P-k(ci,j)
24
Theorem II
  • Our algorithm computes the VCG prices correctly,
    uses routing tables of size O(nd) (i.e., imposes
    only a constant-factor penalty on the BGP
    routing-table size), and converges in at most (d
    d) stages (i.e.,imposes only an additive
    penalty of d stages on the worst case BGP
    convergence time).

25
Open Problems
  • AS implement distributed algorithm different
    algorithm.
  • Overcharging
  • Convergence time

26
Comments-I
  • This paper considers distributed resource
    allocation problems
  • Users are distributed
  • Resources are distributed
  • Computation is distributed
  • Focus on Network complexity
  • Assume that the economic mechanism involves a
    distributed computation carried out over a
    network.
  • Network complexity measures the computational and
    communication efficiency of the distributed
    algorithm

Internet is the biggest and most successful
distributed system, making it a natural source
of DAMD problems.
27
Comments-II
  • Field still not mature
  • No particularly compelling example yet
  • Some isolated results, but no coherent framework
  • Many fundamental issues unresolved
  • Four examples
  • Congestion Game
  • Alternate Path Game
  • Inter-domain Routing Game
  • Multicast Cost Sharing Game

28
Comments-III
  • Common Features of such games
  • Little information about infrastructure and
    other players.
  • Dynamic Environment
  • Asynchronous
  • Computational Constraints
  • Communication/computation costs are important
  • Mechanisms should have low network complexity
  • DAMD raises two types of issues
  • Game-theoretic issues in distributed systems
  • Distributed computational issues in resulting
    economic
  • mechanisms
Write a Comment
User Comments (0)
About PowerShow.com