2/16

Interactions of Overlay Routing

● Overlay, overlay, overlay!○ Virtual network on top of IP network

● Vertical Interaction: Overlay vs. Underlay

○ Overlay: application-layer routing○ Underlay: IP-layer routing

● Horizontal Interaction: Overlay vs. Overlay

○ Multiple overlays share the same underlay network

● Internal Interaction: Overlay node vs. Overlay node○ Overlay nodes interact within a single overlay

3/16

Contents

● Motivation● Vertical Interaction Game● Traffic Engineering (TE)

○ MPLS, Oblivious, COPE● Overlay Routing

○ TE-Aware vs. Selfish● Simulation● Conclusion

4/16

Motivation: An Analogy

● Consider a transportation network● Traffic authority manages traffic lights

○ to make a good traffic flow○ to avoid traffic jam○ global view & manage the whole traffic

● Drivers choose their directions ○ to get to their destination ASAP○ local view & care about its own drive

Authority vs. Drivers

TE vs. Overlay

5/16

Vertical Interaction Game

TE : (Traffic Matrix) x (Topology) • (IP-layer Routing)

(IP Routing) x (Overlay Traffic) • (Overlay Latency)

Overlay: (Overlay Latency) x (Overlay Topology) • (Overlay Routing)

(Overlay Routing) x (Underlay Traffic) • (Traffic Matrix)

In game theory, a two-player non-cooperative repeated game

6/16

Traffic Engineering (TE)

● Input: traffic matrix, (static) topology● Output: physical routing (a.k.a. IP-layer, underlay)● {fs,d (l) | s,d: node, l: link }

○ source-destination (s, d) and link l○ fraction of demand (s, d) flowing through link l

● flow conservation law:○ for each router,○ total incoming traffic = total outgoing traffic

7/16

Various TE Techniques

● MPLS Traffic Engineering○ Adapt to the current traffic matrix (TM)○ Minimize the Maximum Link Utilization (MLU)

● Oblivious Routing○ Reasonable performance for all possible TMs○ Optimal oblivious ratio

● COPE○ Convex-hull-based Optimal TE with Penalty Envelope○ A hybrid approach○ Close-to-optimal performance for normal TM○ Performance penalty bound with unpredicted TM

8/16

Overlay Routing

● Input: link latency, (static) overlay traffic matrix● Output: logical routing (a.k.a. overlay, application-

layer)● {hps’t’ | s’, t’: overlay node, p: overlay path}

○ overlay node pair (s’, t’), path P○ fraction of demand (s’, t’) flowing through path P

● logical flow conservation sum of traffic in each path = total overlay demand

9/16

Overlay Routing

● Selfish Overlay Routing○ minimize total latency without any constraints

● TE-Aware Overlay Routing

○ (1) limit the selfishness so that overlay doesn’t bother TE■ if current latency > threshold, run limited optimizer■ minimize total latency■ given specific link is not overloaded • lower TE’s MLU

○ (2) overlay even helps out TE in some case■ if current latency < threshold, run load balancer■ minimize overlay traffic demand • reduce traffic demands■ given the current latency preserved

10/16

Overlay 1: Selfish

● Selfish Overlay Routing○ minimize total latency without any constraints

11/16

Overlay 2: TE-Aware

● Load-Balancer Limited-OptimizerReduce overlay traffic

Preserve current latencyDon’t overload a specific link

Reduce overlay latency

12/16

Simulation Settings

● 14-Node Tier-1 POP-to-POP Topology - 4-node overlay network

● Synthetic Traffic Matrix with Gravity Model - 10% of the whole traffic is controlled by overlay

13/16

MPLS vs. COPE under Selfish Overlay

● COPE is better than MPLS for both players○ Overlay latency variance: MPLS (75) >> COPE (1)○ TE MLU variance: MPLS (27%) >> COPE (5%)

● Adaptation may not be a good idea..

14/16

Selfish vs. TE-Aware on top of MPLS

● Stable TE-aware overlay○ Overlay latency variance: selfish (45) > TE-aware (15)

● Big MLU spikes with selfish overlay○ TE MLU variance: selfish (25%) >> TE-aware (5%)

15/16

Simulation Summary

85~86

(2)

80~95

(15)

85~86

(1)

67~107

(40)

overlay latency

51~55%

(4%)

50~53%

(3%)

50~55%

(5%)

50~77%

(27%)

TE’s MLU *

TE-aware on COPE

TE-aware on MPLS

Selfish on COPE

Selfish on MPLS

● *MLU: Maximum Link Utilization● 10% of traffic by overlay

16/16

Conclusion and Future Direction

● We improve “vertical interaction” between overlay routing and traffic engineering

○ Adaptation may not be good for TE in the presence of dynamic overlay traffic

○ Overlay has incentives to be TE-aware to get better performance

● We will enhance the model in several directions○ Vertical interaction in inter-domain level○ Vertical interaction + Horizontal interaction (multiple overlays)○ Overlay routing oblivious to underlay routing

18/16

Selfish vs. TE-Aware on top of COPE

● Both overlays work well on top of COPE

● (Selfish on MPLS) vs. (TE-aware on COPE)○ Overlay latency variance: 68~108 vs. 85~87○ TE MLU variance: 50%~77% vs. 51% ~55%

19/16

Implementation

● GAMS + Perl Scripts● GAMS: General Algebraic Modeling System

○ a language for linear, non-linear optimization problem formulation

○ interface to various LP, NLP solvers○ www.gams.com

● Condor for intensive computation○ takes hours to run the solve the optimization problems○ www.cs.wisc.edu/condor

20/16

Related Works

● Lili Qiu, Yang Richard Yang, Yin Zhang, and Scott Shenker. “On Selfish Routing in Internet-Like Environments”, SIGCOMM 2003.

○ first introduce the vertical interaction○ show MPLS has better interaction than OSPF

● Yong Liu, Honggang Zhang, Weibo Gong and Don Towsley, “On the

Interaction Between Overlay Routing and Traffic Engineering”, INFOCOM 2005.

○ formulate the interaction as a non-cooperative game○ prove the existence of Nash equilibrium○ with simulation, show the poor interactions

21/16

Vertical Interaction with OSPF

● From “Selfish Routing in the Internet-like Environments”

22/16

Vertical Interaction with MPLS● From “Selfish Routing in the

Internet-like Environments”

23/16

Vertical Interaction Game

TE determines the physical routing • link latency

Overlay changes the logical routing • traffic demand

24/16

TE1: Adaptive MPLS

25/16

TE2: Oblivious Routing

26/16

TE3: Hybrid - COPE