Transcript ORALS--finals--oct31/Nov3--2007

Overlay Network Creation and Maintenance
with Selfish Users
Georgios Smaragdakis
Dissertation committee members: Azer Bestavros,
Nikolaos Laoutaris,
John Byers
Overlays & Neighbor Selection
Overlay node
Overlay links
Internet
Transit
ISP
Transit
ISP
Access
ISP
Access
ISP
Access
ISP
Overlay applications:
overlay routing,
p2p file sharing,
content distribution..
Focus on service
quality!
2
Challenges
v2
v1
v5
v4
v6
v7
v8 that
 What
gain
p1=[v2v3v4is
v5v6the
v7v8v9performance
]
can be achieved by a selfish node?
p8=[v1v2v3v4v5v6v7v9]
 Whatvis
3 the impact of selfish neighbor
selection to overlay network
p3=[v1v2v4v5v6v7v8v9]
performance?
v9
p9=[v
node
Selfish
What
are the implications of
selfish
1v2v3v4v5v6v7v8]
neighbor selection to system design?
3
Outline
Implications
Selfish Neighbor
Selection
to
Overlay Routing
Implications to File Sharing
Implications to
Service Provisioning
4
Implications
Selfish Neighbor
Selection
to
Overlay Routing
Implications to File Sharing
Implications to
Service Provisioning
5
Selfish Neighbor Selection (SNS)
 Constraints that need to be addressed in a
realistic model for overlay networks:




Bounded degree
Preference vectors
Realistic network distance
Link directionality
 Fundamentally different from other models that
have been proposed for other networks.
[Fabrikant et al.,PODC’03; Chun et al., Infocom’04 …]
6
Optimal Neighbor Selection
vi: choose k neighbors, s.t.
min C i ( S ) 
 pij  d S (vi , v j )
w
v j Vi
over all siSi
vi
u
G-i=( V-i , S-i )
Set of residual nodes
Set of residual wiring
vi’s residual network
7
SNS & Facility Location
Uniform link weights, and uniform preference
 k-median on asymmetric distances
8
k-median
k-median:
Find a subset I of F and a function σ:CI
to min ( Σi,j sjcij ) such that |I| ≤ k
F: set of
facilities
C: set of clients,
cij: cost connecting
client jfacility I
sj: demand of node j
9
Uncapacitated Facility Location
Uncapacitated Facility Location (UFL):
Find a subset I of F and a function σ:CI
to min ( Σi fi + Σi,j sjcij )
F: set of
facilities
fi: cost to
open
facility
C: set of clients,
cij: cost connecting
client jfacility I
sj: demand of node j
10
SNS & Facility Location
Uniform link weights, and uniform preference
 k-median on asymmetric distances
Since the wiring
cost is the same
w
Non-uniform link weights, and uniform
vi
preference
 ILP formulation
min Ci ( S ) 
p
v j Vi
ij
u
can be
 d S (vw,u
i,vj )
obtained from
k-median on
reversed distances
11
Local Search (LS)
vi: choose k neighbors
min Ci ( S ) 
p
v j Vi
ij
 d S (vi , v j )
w
over all siSi
vi
[Arya et al,STOC’01]
u
G-i=( V-i , S-i )
Set of residual nodes
Set of residual wiring
vi’s residual network
12
SNS : the Game
Game <V,{si},{Ci}>
 V : set of n players (nodes)
 {si}: strategies available to vi (wirings),
choose k out of n to connect
 {Ci}: set of costs for vi
min C i ( S )   pij  d S (vi , v j )
v j Vi
 Best response of a node: node’s optimal wiring
Outcome: S, the global wiring
 A stable wiring is a pure Nash equilibium
 Using iterative best response
Fundamentally different from selfish routing
13
SNS : Equilibria
Uniform Preference
n=15
Skewness of preference
k=2
k=3
k=8
k=11
k (Link density)
In-degrees are highly skewed
even under uniform preference !
 Quality-based
“preferential attachment”
14
SNS : Efficiency
Performance of ILP & LS is close to
Utopian!
Link density
Skewness of
preference
Link density
Skewness of
preference
 Theoretical results showed in the worst case the cosial
cost can be bad
[Laoutaris, Poplawsi, Rajaraman, Sundaram, Teng,PODC’08]
15
SNS : Trace-Driven Evaluation
How we assign the distance:
 Synthetically using BRITE
 Empirically from PlanetLab
 Empirically from AS-level maps [Routeviews]
Neighbor Selection Strategies:




k-Random heuristic
k-Closest heuristic
k-Regular heuristic
k-Best Response
Control parameter:
 Bound on out-degree k (link density)
16
Connecting on a k-Random graph
BRITE (n=50)
0 2 3
5
k
11
22
PlanetLab (n=50)
0 2 3
5
11
k
22
AS-Level (n=50)
0 2 3
5
11
22
k
If your neighbors are naïve, it pays to be selfish!
17
Connecting on a k-Closest graph
BRITE (n=50)
0 2 3
5
11
22
PlanetLab (n=50)
0 2 3
k
5
11
k
22
AS-Level (n=50)
0 2 3
5
11
22
k
“Greed is not good”
If your neighbors are greedy, it pays to be selfish!
18
Connecting on a k-Regular graph
BRITE (n=50)
0 2 3
5
k
11
22
PlanetLab (n=50)
0 2 3
5
11
22
k
AS-Level (n=50)
0 2 3
5
11
22
k
“Common pattern is not good”
If your neighbors have the same wiring pattern,
it pays to be selfish!
19
Connecting on a Best Response graph
BRITE (n=50)
0 2 3
5
k
11
22
PlanetLab (n=50)
0 2 3
5
11
22
k
AS-Level (n=50)
0 2 3
5
11
22
k
The BR graph is highly optimized!
If your neighbors are selfish, it is OK to be naïve!
20
SNS vs. Heuristics: Social Cost
Macroscopic view:
Focusing on the social welfare
(k=2)
k-Random/BR
k-Closest/BR k-Regular/BR
BRITE
1.44
1.53
3.61
PlanetLab
2.23
1.48
3.84
AS
2.04
1.90
4.78
The network is better off with selfish nodes!
21
Real-Time Applications
Min-Max Best Response
Worst delay
in the overlay:
0 2 3
5
11
22
k
22
SNS with Variable Degree
Real-time applications
100 links
Variable degree
through LS:
 Swap 1 link
 Add 1 link
 Drop 1 link
120 links
Application requirement
(Performance when k=5, n=50
i.e. 250 links)
23
Implications
Selfish Neighbor
Selection
to
Overlay Routing
Implications to File Sharing
Implications to
Service Provisioning
24
Basic design of EGOIST:
Link state protocol
Measurements of distance to candidate neighbors
Wirings according to chosen strategy
Re-wirings every T second
A newcomer bootstraps by connecting to arbitrary
neighbors
25
EGOIST : Performance
Best
Response
26
EGOIST: Passive Measurements
Passive measurements based on virtual
coordinates (pyxida system) with minimal
cost
27
EGOIST: Other Metrics
 End-to-end available bandwidth (pathchirp)
with minimal measurement overhead
 CPU load (loadavg)
28
EGOIST: Marginal Utility of Rewiring
BR
Lazy BR (threshold = 10%)
There exists a performance knee (k=3 or 4)
 Re-wirings could be reduced with lazy BR

29
quality
Connectivity
Efficiency Index
EGOIST: Effect of Churn
 Connectivity is guaranteed (in T/n time)
 HybridBR (a connected ring is maintained)
delivers much of the efficiency of BR
30
quality
Connectivity
Efficiency Index
EGOIST: Effect of Churn
 BR and Hybrid BR dominate all the other
heuristics
 HybridBR pays off at high churn
31
EGOIST : Other Work
 CPU and memory load is very low
 Robust to cheating
 Scalability


via topological sampling
via layered architecture
 Applications including multi-player P2P
games, real-time traffic over IP etc.
32
Implications
Selfish Neighbor
Selection
to
Overlay Routing
Implications to File Sharing
Implications to
Service Provisioning
33
Modern File Sharing Systems
 Parallel upload/
download
Internet
Seeder
Transit
ISP
Transit
ISP
Access
ISP
Access
ISP
Access
ISP
- Swarming
 Local scheduling
- Local Rarest First
 Flat connectivity
- Choke/unchoke
Leecher
Overlay node
34
n-way Broadcast
Internet
 Synchronization
- Distributed databases
- Backups
 Batch parallel
processing
- The files have to be
received by all nodes
before the next step
of processing begins
35
Preliminary Solutions
 n co-existing swarms
(-) Stress of physical links
(-) Exchange of multiple chunks in parallel overpartitions
the uplink capacity [Tian et al., ICPP’06]
 End-system multicast (mesh)
[SplitStream, Bullet]
(-) Creates an overlay for each swarm
(-) No coordination among swarms
(-) Monitor overhead
36
Design Strategies for n-way Broadcast
 Joint optimization of upload/download
while participating in many swarms
 Data Agnostic
- Keeps swarming and local scheduling
 Bandwidth-Centric
- Max-flow to approximate swarming behavior
[Massoulie et al., Infocom’07]
 Bounded Degree
37
Reducing the Average Download Time
Objective: Minimize the average download time
Max-Sum:
 Neighbor selection strategy of node vi:
max (sum (MaxFlow(vi, vj)), for all vj
38
Reducing the Download Time
Objective: Minimize the total download time
Max-Min:
 Neighbor selection strategy of node vi:
max (min (MaxFlow(vi, vj)), for all vj
39
Optimized Graphs and Swarming
 Formation of stable graphs
 Each node strives to improve both the
upload and download flow
 Performance of swarming on optimized graphs
- Max flow might not be realizable
40
Performance Evaluation
Max-Sum
Max-Min
Delivery Time
Node ID
Naive
Selfish Upload:
Protects the uplink
File ID
capacity
of the slow node File ID
 Improves the
download time in the
system
File ID
 Flattens distribution time!
 Guarantees synchronization!
 Comparable average download
time
41
Other Work: File Searching
Best response: max #nodes reached
4
Bootstrap
1
Server
3
6
5
2
selfishly
TTL of scoped flooding is 2
 Maximum Coverage Problem
42
Implications
Selfish Neighbor
Selection
to
Overlay Routing
Implications to File Sharing
Implications to
Service Provisioning
43
Server Selection
Hardware server
44
Centralized Deployment
Generic Service Host
Software server
Demand change
e.g. Flash crowd,
time-of-day
45
effect
Dynamic Service Deployment
Generic Service Host
Software server
Demand change
e.g. Flash crowd,
time-of-day
46
effect
Distributed Service Migration (DSM)
“ring” nodes
r-ball
(r=2)
 Solve k-median or UFL
in an r-ball
 ..BUT nodes outside the
r-ball are totally neglected

Iterate until
convergence
47
DSM: Properties
Convergence:
Migration only if the cost of facilitating the
demand decreases at least be a%, converges in
O(log1+a n) steps
 We can control the speed of convergence by tuning a
Limited horizon view requirement:
 r regulates the trade-off between scalability and
performance
48
DSM: Evaluation
Similar results for UFL under different
cost functions to open and maintain the server
49
Dynamic vs. Static Deployment
Static
deployment
DSM
DSM
Dynamic
deployment
50
Conclusions
 What is the performance gain that can
be achieved by a selfish node?
 Selfish nodes can reap substantial performance gain.
 What is the impact of selfish neighbor
selection to overlay network
performance?
 Surprisingly, the evolving graphs have also good
performance!
51
Conclusions
 What are the implications of selfish
neighbor selection to system design?
 Selfish wiring strategies are easily realizable
 Selfish wiring behavior can be used towards
distributed overlay network creation and
maintenance
 Selfish wiring must be a component of any system
to protect it from abuse
 Selfish wiring behavior can be used for efficient
dynamic service provisioning
52
Thank You!
http://csr.bu.edu/sns
http://csr.bu.edu/dfl
53