1. Analyzing Cross-layer Interaction in Overlay Networks
Srinivasan Seetharaman
September 2007

2. Overlay Networks
Overlay networking helps overcome functionality limitations of the Internet by forming a virtual network over the native IP network that is:
Independent
Customizable
Complex machine with good processing power

3. Service Overlay Networks
Offer enhanced or new services by deploying intelligent routing schemes.
Overlay link
Relaying

4. Service Overlay Networks (contd.)
Characteristics:
Nodes and links are persistent
Perform overlay routing independent of native layer routing
Each Overlay path comprises one or more Overlay links, based on a certain selfish objective
Many types of services can be offered
Multicast (e.g. ESM, Overcast)
QoS (e.g. OverQoS, SON)
Security (e.g. DynaBone, SOS)
Better routes (e.g. RON, Detour, X-Bone)
… and much more

5. Cross-Layer Interaction
Performing dynamic routing at both overlay and native IP layers leads to:
Conflict due to mismatch or misalignment of routing objectives
Contention for limited physical resources
Functionality overlap (Both overlay layer and IP layer perform similar set of functions)
We demonstrate the impact in our work

6. Cross-Layer Interaction (contd.)
These issues are amplified in the presence of
Selfish motives and aggressive behavior
Lack of information about other layer
Increasing impact ( #overlays  |Traffic| )
P2P networks are a form of Overlay networks. But, they rarely relay data through other peers

7. Context Native network topology Attitude of native network
Intra-domain
Inter-domain
Attitude of native network
Restrictive
Oblivious
Cooperative

8. Thesis Organization . Oblivious Restrictive Cooperative Intra-domain
Interaction with failure recovery
[Chapter III]
OR vs Load balancing [Chapter V]
Overlay-friendly native network
[Chapter VII]
Inter-domain
OR vs Policy constraints
[Chapter IV]
BT vs Load balancing [Chapter VI]

9. INTERACTION BETWEEN FAILURE RECOVERY IN THE NATIVE AND OVERLAY LAYERS
Chapter III

10. Dual Rerouting
Each layer performs rerouting, with no knowledge of which layer leads to optimal restoration
Overlay rerouting C
OVERLAY1
LAYER E F H A G D A B A C E F
Functionality overlap and suboptimality H X
Failure B
NATIVE IP
LAYER D G
Native rerouting

11. Downside to Dual Rerouting
Overlap of functionality between layers causing
Unnecessary route changes (esp when connectivity in native network is very dynamic)
Increased probing overhead
Unawareness of other layer’s decisions leading to
Multiple simultaneous failures
Lack of flexibility and control

12. Tuning Dual Rerouting
Intra-domain (keepAlive-time = 1 sec, hold-time = 3 secs)
Dual Rerouting
Suppress overlay rerouting at 0.5 prob.
Defer overlay rerouting by secs
Native-only rerouting
Average route changes
125.08%
101.59%
109.85%
1.567
Stabilized inflation
100%
108.32%
1.202
Time when stable
113.7%
100.48%
107.33%
2.481
Peak inflation
114.22%
109.98%
110.73%

13. Further Improving Recovery
Adjust the functioning of native layer:
Tuning the native layer keepAlive-time:
keepAlive-time
keepAlive-time
This produces the best tradeoff between # of route changes, stabilization time and recovery time
Tuning

14. INTERACTION BETWEEN OVERLAY ROUTING AND TRAFFIC ENGINEERING
Chapter V

15. Repeated Non-Cooperative Game
Player1: Overlay Routing - Latency-optimized paths between nodes
Player2: Traffic Engineering - Optimal load-balanced routes
Overlay
Routing
Overlay Link
Latencies
Overlay routes 
Overlay layer
traffic
Native link
delays 
Say it in layman terms
Traffic on each overlay link
Traffic
Engineering
Native
routes 
Background traffic TM

16. Simulation Results TE objective Overlay objective Overall stability
Round

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

18. Resolving Conflict – Our Approach
Assume: Each layer has a general notion of the other layer’s selfish objective
Designate leader / follower
Operate leader such that
Follower has no desire to change  Friendly
Follower has no alternative to pick  Hostile
Use history to learn desired action gradually.
Follower is forced

19. Performance of Preemptive Strategies
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

20. CROSS-LAYER INTERACTION OF PERFORMANCE-AWARE OVERLAY APPLICATIONS
Chapter VI

21. BitTorrent File-Sharing
Popular file-sharing application that generates a large volume of Internet traffic
Characteristics:
Service capacity increases with demand
Centralized tracker regulating neighborhood
Dynamically change active peers by choke/unchoke protocol

22. Comparison to Overlay Routing
Data2
Data1 B1 B2 A3 A2 A1
Differences, Similarities AX BY

23. BitTorrent Protocol
Tit-for-tat based incentive for uploading decisions
Leecher: Unchoke the fastest uploaders
Seed: Unchoke the fastest downloaders
Popular strategy to improve performance
Optimistic unchoke: periodically look for faster peers
Feedback loop

24. BitTorrent Dynamics When bottlenecked on link L1 L2 L1
Choke
Choke C
When bottlenecked on link L1 L2
Unchoke
Request
Unchoke L1 X
Peer
TFT
Upload stats
Download stats
Status
Opt?
Interested
Pieces?
My interest
Load distribution across links is balanced
BitTorrent apps use all available b/w

25. BitTorrent Dynamics When NOT bottlenecked on link L1 L2 L1
Choke
Choke C
When NOT bottlenecked on link L1 L2
Unchoke L1 X
Fills up available bandwidth easily
Load distribution across links is unbalanced

26. Cross-Layer Interaction
Operating BitTorrent disrupts load balance and can result in high max util:
This can be a problem for background traffic
Objective of native layer:
Minimize ( Max Util.)
Objective of BitTorrent:
Minimize (Overall finish time)

27. Simulation Setup Pick 100 ASes with 60% of them being non-stub ASes
Upload bandwidth is heterogeneous with non-stub AS
Pick 100 ASes with 60% of them being non-stub ASes

28. Simulation Setup L1 D F F B C A
Peers per Torrent
Peers per Torrent per AS
Bandwidth constraints E L2
Generate 1-50 peers. Each associating with 1-3 torrents

29. Simulation Performance Metrics
Max util across access links = MaxaE ( Xa/Ca ), E is set of all links
X is the load, C is the capacity
Average finish time inflation of leechers = 1/Nl  ( ’i / i ) Nl is # of leechers
’ is finish time after strategy
Nl i=1

30. Reducing Impact – Traffic Engg
TE can be performed across inter-domain access links, in order to minimize (Max util)
Two flavors: Ingress / Egress
Determines which access link to pick for a certain destination or source IP address

31. Reducing Impact – Traffic Engg (contd.)
Performance of a random AS (Focus AS)
Results of focus AS
Applying TE does not make much difference

32. Reducing Impact – Tuning BitTorrent
Alter certain BitTorrent protocol components or tune the associated parameters
Minimal reduction of the max util
Significant inflation of finish time
Specifically, we tried each of the following:
Make peer selection random
Make piece selection random
Reduce duration of optimistic unchoking
Freeze list of unchoked peers after 10 mins
Tune the unchoking timers

33. Reducing Impact – Locality-awareness
Locality-based traffic management
Give priority to peers within AS
No change to BitTorrent clients
Also try caching of requests sent outside AS

34. Reducing Impact – Bandwidth Throttling
Limit bandwidth consumed by BT traffic
Popular strategy among most ASes
Involves lesser infrastructure cost

35. Cross-layer Conflict
Native layer and BitTorrent layer constantly retaliate to other layer’s disruptive behavior
Peers deploy BitTorrent Protocol Encryption to avoid detection by native layer
We develop two “friendly” BitTorrent strategies that achieve a mutually agreeable point by reducing peak load

36. A. Limit # of parallel downloads
The unchoking protocol and their timeline is uncoordinated across neighbors
Average

37. A. Limit # of parallel downloads (contd.)
Reduces peak load from 0.94 to Finish time inflation is

38. B. Avoiding common neighbors
Problem is that two peers in same AS often contact same peer outside AS
Algorithm
Perform bilateral info exchange where each peer A finds out if its neighbor B has a neighbor C inside its own AS
If yes, toss a coin to determine if we can download from this peer B (Randomization acts as a load balancing strategy)

39. B. Avoiding common neighbors (contd.)
Reduces max util from 0.94 to 0.85 Finish time inflation is 1.187

40. ANALYZING INTER-DOMAIN POLICY VIOLATIONS IN OVERLAY ROUTES
Chapter IV

41. Inter-Domain Policy Violations
Two types of violations exist
Provider 2
Provider 1
Peer
Client 1
Legitimate native route $$ $ A
Client 3
Client 2
Overlay route B C
Transit violation
Exit violation

42. Measurement Results
Each transit violation has a corresponding exit violation upstream
Extent of exit policy violations in multihop paths
Violation Type
% paths
Next hop AS violated
72.05
Exit point violated
15.63
Total
87.68

43. Policy Enforcement by Native Layer
As ISPs become aware of the negative impact of overlays and commence filtering, this leads to
drastic deterioration in overlay route performance
commensurate with the number of ASes enforcing policy

44. Resolving Conflict
Overlay Service Provider (OSP) adopts a combination of the following strategies for achieving good legitimate paths:
Obtain transit permit from certain AS for $T
Add new node to certain provider AS for $N
Obtain exit permit from certain AS for $E

45. Illustration of Mitigation Strategy
With no filtering,
Tier-1 provider
Tier-2 provider
Stub customer 11 13
AS hosting overlay node
Cust-Prov relation 21 22 23
Peering relation 31 32 33
Transit violation

46. Illustration of Mitigation Strategy (contd.)
With filtering, we have no multi-hop paths
Tier-1 provider
Tier-2 provider
Stub customer 11 13
AS hosting overlay node
Cust-Prov relation 21 22 23
Peering relation 31 32 33

47. Illustration of Mitigation Strategy (contd.)
Option 1: Add new overlay node to provider AS 22 Option 2: Obtain transit permit from stub AS 32
Tier-1 provider
Tier-2 provider
Stub customer 11 13
AS hosting overlay node
Cust-Prov relation 21 22 22 23
Peering relation
Based on graph theoretic measurements, we obtain an optimal mix of expenses for a certain budget 31 32 33

48. Objective of Mitigation Strategy
For a certain budget, determine which ASes
to obtain transit permit from
to add new node to
to obtain exit permit from
… so as to achieve the best possible gain
Gain = Native route latency – Overlay path latency
Native route latency

49. Mitigation Results When all permit fee = P, new node fee = N Permit
Add new node

50. Summary of Cross-Layer Interaction
Overlays offer valuable services needed by end-systems. But, lead to complex cross-layer interaction with potentially detrimental effects
Layer awareness is essential to reduce negative effects and to improve performance of both layers. We propose simple strategies that achieve this goal in an effective manner.

51. Contributions of Thesis
Knobs for better control over the cross-layer interaction
Analysis and mitigation of the conflict in objective between native and overlay layers:
inter-domain: OR vs Policy enforcement
intra-domain: OR vs Traffic engg
Ways to improve coexistence between BitTorrent file-sharing and native layer
Framework for network layer support of overlay services

52. Future of Overlays Overlays are essential as…
Means for end-systems to collaborate
Environment for testing future innovations (GENI)
Architecture for Future Internet in the form of Network Virtualization
 Cross-layer interaction will affect performance. How best to design protocols and services in the future?
Two perspectives to the design problem

53. Future Work
Need to address the network impasse. How to tune the network for
.. the new breed of Internet applications? (e.g., file sharing)
…and new paradigms of communication? (e.g., wireless)
Which layer to implement a service at? For example, a service like multicast can be performed at both native layer and overlay layer!
Which layer to use for a particular scenario?
How can the other layer support this service?
Solving chicken and egg problem
Talk about IPTV