One Decoding Step

1
Placing Relay Nodes forIntra-Domain Path
Diversity

Meeyoung Cha (KAIST, http//an.kaist.ac.kr/mycha)
With Sue Moon, Chong-Dae Park and
Aman Shaikh
2
Placing Relay Nodes forIntra-Domain Path
Diversity

Meeyoung Cha (KAIST, http//an.kaist.ac.kr/mycha)
With Sue Moon, Chong-Dae Park and
Aman Shaikh
3
Routing Instability in the Internet

• Link and router failures are frequent.
• Routing protocols are used to detect such
  failures and route around them.
• Convergence time is in the order of seconds or
  minutes.
• End-to-end connections experience long outages.
• How to increase reliability and robustness of
  mission-critical services against temporary
  end-to-end path outages?

4
Path Diversity and Overlay Networks

• Take advantage of path diversity provided by the
  network topology.
• Overlay path use a node inside the network to
  relay packets over an alternate path that is
  different from the default routing path.
• ex) RON Anderson et al., SOSP 2001
• Detour Savage et al., IEEE Micro 1999
• Use disjoint overlay paths along with the
  default routing path to route around failures.

5
Objective of Our Work

• Previous work has focused on selecting good relay
  nodes assuming relay nodes are already deployed.
• As an ISP, we consider the problem of placing
  relay nodes well.
• Find a fixed set of relay nodes that offer as
  much path diversity as possible to all OD pairs.
• Under Assumptions
• Intra-domain setting Shortest Path First
  Routing
• Relays are simply routers with relaying
  capability
• Overlay paths use single relay nodes

6
Path Diversity Disjoint Overlay Path
ISP Network
Destination (egress router)
relays
default path
Origin (ingress router)
disjoint overlay path
Disjoint overlay path gives maximum robustness
against single link failures!
7
Impact of ECMP on Overlay Path Selection

• Completely disjoint overlay paths are often not
  possible.
• - Existing path diversity Equal Cost
  Multi-Paths (ECMP)

(AR Access Router, BR Border Router)
8
Partially Disjoint Overlay Path
We may need to allow partially disjoint paths.
r
overlay path
o
d
default path
Such overlap makes networks less resilient to
failures. We introduce the notion of penalty to
quantify the quality degradation of overlay paths
when paths overlap.
9
Penalty for Overlapped Links

• Impact of a single link failure on a path
• - prob. a packet routed from o to d encounters a
  failed link l
• Io,d,l P path o?d fails link l fails

10
Penalty Measures

• Consider overlay path (o?r?d) is used with
  default one (o?d).
• Penalty fraction of traffic carried on
  overlapped link

• Penalty of a relay r for OD pair (o,d)
• prob. both packets routed (1) directly from o to
  d and (2) indirectly from o to d via r encounter
  a single link failure
• Po,d(r) P both o?r?d and o?d fail
  single link failure
• Penalty of a relay set R of size k
• sum of minimum penalty of all OD pairs using
  relays in R ?o,d min( Po,d(r) r in R )

11
Placement Algorithms

• How to find a relay set R of size k with minimum
  penalty
• Optimal solution
• Exhaustive search, 0-1 Integer Programming (IP)
• Too expensive
• Greedy selection heuristic
• Start with 0 relays
• Iteratively make a greedy choice that yields
  minimal penalty
• Repeat until k relays are selected
• Local search heuristic
• Start with k set of random relays
• Repeat single-swaps if penalty is reduced

12
Evaluation Overview

• Performance evaluation
• What is the penalty of placement heuristics as we
  increase the number of relays?
• How far are our heuristics from optimal solution?
• Sensitivity to network dynamics
• Using three-month topology snapshots, we examine
  whether relays selected remain effective as
  topology changes.
• Using six-month network event logs, we calculate
  the fraction of traffic that is protected from
  failures by using relays.

13
Dataset

• We use an operational tier-1 ISP backbone
• and daily topology snapshots and event logs.
• Topology - 100 routers, 200 links
• Hypothetical traffic matrix
• - assumes equal amount of traffic between OD
  pairs
• For results on other topology (1 real, 3
  inferred, 6 synthetic), please refer to our
  technical report at
• http//an.kaist.ac.kr/mycha/docs/CS-TR-2005-214.
  pdf

14
Performance Evaluation
Optimal minimal penalty using k relays
Random randomly select k relays Degree select
nodes with highest degree first Lower bound
when all n nodes are used as relays
15
Sensitivity to Network Dynamics
5 of nodes are selected as relays
10 of nodes are selected as relays
Relay nodes by initial placement are nearly as
good as daily relocation relatively insensitive
to network dynamics.
16
Hypothetical Traffic Loss from Failure Event Logs
less than 1 of traffic lost for 92.8 failures
(failure events)
complete protection for 75.3 failures
17
Conclusions

• This is the first work to consider relay
  placement for path diversity in intra-domain
  routing.
• We quantify the penalty of using partially
  disjoint overlay paths and propose two
  heuristics for relay node placement.
• We evaluate our methods on diverse dataset.
• Relays by our method perform consistently better
  than other heuristics and are near-optimal.
• A small number of relay nodes (less than 10) is
  effective over the entire course of several
  months.
• Relays are relatively insensitive to network
  dynamics.

18
Further Works

• Relay architecture and practical considerations
• loose source routing option in routers/attaching
  servers to routers
• reflecting real traffic matrix
• Relay placement in inter-domain setting
• inter-domain routing is based on BGPs path
  selection
• very challenging AS path inference, AS path
  asymmetries, and realistic traffic matrix
  estimation
• Lower layer path diversity
• at physical layer, disjoint IP layer paths may
  run over the same optical fiber
• how to incorporate fiber map into our algorithm?

END
19
Dataset used for Evaluation
Abilene

• Real topologies
• Abilene (N11, E14, Degree24)
• Tier-1 ISP backbone (N100, E200, Degree210)
• Rocketfuel inferred topologies
• Exodus (N79, E147, Degree112)
• Ebone (N87, E161, Degree111)
• Tiscali (N161, E328, Degree129)
• Synthetic topologies (scale-free, random)
• Albert-Barabasi model (N50/100, E81/197,
  Degree217/225)
• Heuristically Optimal Model (N171, E440,
  Degree110)
• mesh, torus, ring (N64/64/64, E112/128/64,
  Degree24/4/2)

HOT
A mesh and a torus