Preemptive Strategies to Improve Routing Performance of Native and Overlay Layers Srinivasan Seetharaman - College of Computing, Georgia Tech Volker Hilt - Multimedia Networking, Bell Labs Markus Hofmann - Multimedia Networking, Bell Labs Mostafa Ammar - College of Computing, Georgia Tech

Multi-Layer Interaction Service overlay networks offer enhanced services by forming a virtual network of specialized nodes They deploy independent routing schemes that are oblivious to underlying native network achieve a specific selfish objective Two main problems: Mismatch of routing objectives Misdirection of traffic matrix estimation

Repeated Game Model Player1: Overlay Routing (OR) Latency-optimized paths between nodes Reacts to changes in link latency by probing periodically, without concern for bandwidth Player2: Traffic Engineering (TE) MPLS-based scheme that solves a linear program (using GNU LP kit) to obtain optimal multi-paths using traffic matrix as input Minimize [ Max util = MaxaE ( Xa/Ca ) ]

Repeated Game model (contd.) Overlay Routing Overlay Link Latencies Overlay routes  Overlay layer traffic Native link delays  Traffic on each overlay link The frequency of looping depends on the probing mechanism Traffic Engineering Native routes  Background traffic Traffic Matrix

Illustration of OR vs TE 14ms C Shortest latency routes A 4ms 4ms B 5ms 10ms D OVERLAY 23ms NATIVE 2 E F 4 10ms 2ms C 3 3 4 2ms 4 3 Minimize (Max util) 2ms 2ms Numbers on each link represent the avail-bw 2ms 3ms A B 5 G H 4ms 2 2 3 3 3ms 3ms 6ms 2ms I 2 J 2 D 10ms 10ms Initial State

Illustration of OR vs TE (contd.) 14ms C Multihop paths A  B  C A  B  D A 6ms 4ms B 5ms 10ms D OVERLAY 23ms NATIVE 2 E F 4 10ms 2ms C 4 2ms 2 2 2ms 2ms 2ms 3ms A B 1 G H 4ms 2 2 1 2 6ms 3ms 3ms 2ms I 2 J 2 D 10ms 10ms Overlay traffic introduced Avail-bw changed

Illustration of OR vs TE (contd.) 14ms C Multihop paths A  B  C A  B  D A 4ms 5ms B 5ms 10ms D OVERLAY 23ms NATIVE 2 E F 2 10ms 2ms C 1 1 2 2ms 2ms 4 2 2ms 2ms 3ms SPLIT A B 3 G H 4ms 1 1 1 2 3ms 3ms 6ms 2ms I 2 J 2 D 10ms 10ms Latency changed After TE reacts

Illustration of OR vs TE (contd.) 14ms C Multihop paths A  B  C A  B  C  D B  C  D A 4ms 5ms B 5ms 10ms D OVERLAY 23ms NATIVE 2 E F 10ms 2ms C 1 1 4 2ms 2ms 2ms 2ms 3ms Traffic going from node A to D cannot be split into components that go to C and that which goes to D SPLIT A B 5 G H 4ms 1 1 3 3ms 6ms 3ms 2ms I 2 J 2 D 10ms 10ms After Overlay routing reacts Avail-bw changed

Simulation Results TE objective Overlay objective Overall stability Round Clear conflict Duration of suboptimality is longer for TE Lot of oscillations observed

Past research [Qiu-Sigcomm03] conducted a simulation study of scenarios where there is a conflict of objectives [Liu-Infocom05] analyzed the interaction between OR and TE to show existence of Nash equilibrium General conclusion: The system suffers from prolonged route oscillations and sub-optimal routing costs

Our goal .. is to propose strategies that obtain the best possible performance for a particular layer while steering the system towards a stable state.

Resolving Conflict – Basic Idea Designate leader / follower Leader will act after predicting or counteracting the subsequent reaction of the follower Similar to the Stackelberg approach

Resolving Conflict - Obstacles Incomplete information Unavailable relation between the objectives NP-hard prediction

Resolving Conflict - Simplification Assume: Each layer has a general notion of the other layer’s selfish objective Operate leader such that Follower has no desire to change  Friendly Follower has no alternative to pick  Hostile Constitutes a preemptive action Use history to learn desired action gradually. Follower is forced

Overlay Strategy - Friendly Native layer only sees a set of src-dest demands Improve latency of overlay routes, while retaining the same load pressure on the native network! Load-constrained LP C Overlay link Traffic (Mbps) A ® B A ® C 1 B ® C 2 E 1 B 1 D A

Overlay Strategy – Friendly (contd.) Acceptable to both OR and TE Stable within a few rounds

Overlay Strategy - Hostile Push TE to such an extent that it does not reroute the overlay links after overlay routing Send dummy traffic in an effort to render TE ineffective Dummy traffic injection C E 1 Unused overlay link AB B 1 D A - Unused by real overlay traffic - Non-overlapping with overlay links under use

Overlay Strategy - Hostile (contd.) TE can’t improve further Acceptable only to OR

Native Strategy - Friendly TE pays no attention to the length of the route! TE should balance load, while ensuring that the path length is almost the same! Hopcount-constrained LP Native route Next hop Total Hops A ® B E 2 A ® C D D ® E B All others Dest itself 1 C E 1 B 1 D A

Native Strategy - Friendly (contd.) Acceptable to both OR and TE Takes a bit longer to converge

Native Strategy - Hostile Dissuade overlay routing from using certain multihop paths Increase latency of native links that are heavily loaded, without any knowledge of overlay networks Load-based latency tuning Overused native link C E 1 1 B 1 D A

Native Strategy - Hostile (contd.) Disrupted overlay routing Takes a bit longer to converge

Preemptive Strategies: Summary We proposed four strategies that improve performance for one layer and achieve a stable operating point Inflation factor = Steady state obj value with strategy Best obj value achieved Inflation Leader Strategy Overlay TE Friendly: Load-constrained LP Hostile: Dummy traffic injection 1.082 1.023 1.122 1.992 Native Friendly: Hopcount-constrained LP Hostile: Load-based Latency tuning 1.027 1.938 1.184 1.072

Preemptive Strategies: Summary (contd.) Each strategy achieves best performance for the target layer within a few rounds with no interface between the two layers with all information inferred through simple measurements If both layers deploy preemptive strategies, the performance of each layer depends on the other layer’s strategy.