A Paymentbased Incentive and Service Differentiation Mechanism for P2P Streaming Broadcast

1
A Payment-based Incentive and Service
Differentiation Mechanism for P2P Streaming
Broadcast

• Guang Tan and Stephen A. Jarvis
• Department of Computer Science,
• University of Warwick, United Kingdom
• June, 2006

2
Motivation

• Bandwidth-demanding
• Free-riding problem

Goals

• Encourage contribution and discourage free-riding
  via ServDiff
• Achieve higher average media quality

3
Related Work
Incentive ServDiff Mechanisms for

• P2P Streaming Related Applications
• Taxation model (Chu Zhang. SIGCOMM-PINS04)
• Score-based mechanism (Habib Chuang. IWQoS04)
• Pricing model (Wang Li. INFOCOM05)
• General P2P Applications
• Score-based mechanism (Kazaa)
• Reputation systems (Kamvar et al. EigenTrust.
  WWW03)
• Auction-based model (Semret et al. JSAC00)

4
Payment-based auction model

• Multiple substreams
• Periods of fixed length (e.g., 3 min)
• Bidding for substream parents for next period in
  each period
• Payment in Points happens at the beginning of
  next period
• Bonus for serving zero-point peers

• A secure and efficient payment protocol (e.g.,
  bank servers)
• An approximate time synchronization protocol
  (NTP)

Assumptions
5
Payment-based auction model
6
Basic Protocol
Virtual overlay construction

• Peers submit bids to the root
• Highest bidders win the root
• Failed peers choose new targets from the winners
  and re-submit bids to the new targets
• Some peers win the new targets
• The same process continues until all peers have
  found their next-period parents, or they randomly
  find parents with best-effort

7
Parent Selection Strategies (1)
Shortest Path (SP) Strategy A peer selects a
parent from the candidates that makes the
accumulated service latency the smallest.
Advantage Small latency, simplicity
Disadvantage A well located peer may attract
most peers, resulting a highly unbalanced (tall)
tree.
8
Parent Selection Strategies (2)
Balanced Tree (BT) Strategy A peer selects a
candidate parent probabilistically. Given a set
of candidates, the probability of one peer being
picked is in proportional to its number of out
slots.
Advantage Balanced and short tree (small loss
rate) and simplicity.
Disadvantage No Nash Equilibrium.
9
Parent Selection Strategies (3)
Shortest Path Balanced Tree (SP-BT) Strategy A
peer first selects a parent using the SP
strategy. If it fails to win a slot on that
parent, it uses the BT strategy to select a
parent.
Advantage Short tree and Nash Equilibrium.
Disadvantage Relatively complex.
10
Security Issue
A fraction (e.g., 20) of root slots for
non-incentive service
Non-incentive trees

• The non-incentive trees make the attack difficult
• The fraction of non-incentive root slots
  tradeoff between incentive (thus performance) and
  security

11
In-Session Utility Maximization
Purpose Maximize the expected media quality in
each period.
Model To find a best-reply in a game of
incomplete information. Maximize the expected
utility by planning bids for different substreams
under the constraint of a certain number of
points (earned in last period).
12
In-Session Utility Maximization

• uij utility of substream j of peer i
• Ui collective utility of peer i
• bij bid price for substream j by peer i
• Dij mapping from bid price to data loss rate
• Lij mapping from bid price to substream latency
• Ci peer is total number of points (to be spent
  for bidding)

Unknowns that need to be estimated!
13
In-Session Utility Maximization
Problem solving by

• Static even allocation strategy

• Allocate points evenly to all substreams.
• Advantage simple.
• Disadvantage no Nash equilibrium.

4
6
5
5
5
4
5
2
14
Off-Session Point Accumulation
Purpose Maximize individual wealth in each
period (and indirectly increase the systems
bandwidth supply).
Active mode (maximizing utility)
Inactive mode (disconnected from the overlay)
Half-active mode (maximizing wealth)
15
Off-Session Point Accumulation
Model Maximize expected income in terms of
points by buying service of some substreams and
selling them to others.

• bj bid price for substream j
• oi out slots for substream j
• bij bid price for substream j by peer i
• Ej mapping from bid price to expected income in
  terms of points
• W total number of points
• O total number of out slots

Ej(.) needs to be estimated!
16
Off-Session Point Accumulation
Theorem A peer can maximize its expected income
in terms of points by buying a single (arbitrary)
substream and selling that substream using all of
its out slots.
Implication Since an off-session peer
contributes all of its out slots while consuming
only one slot from othters, the systems
bandwidth supply is increased.
Problem solving

• Estimate Ej(.) using a peers own and others
  recent history information
• Solving for optimal solution (i.e., the bid for a
  substream) in O(W) time

17
Simulation Effectiveness of Incentive
Utility vs. Bandwidth
Tree level number vs. Bandwidth
18
Simulation Effectiveness of Incentive
Average utility of all peers with and without
incentive
19
Simulation Effect of Period Length

• Incentive does not significantly increase
  protocol overheads because
• Period length in the order of minutes
• Short tree

20
Simulation Effect of Period Length
The longer the period, the less chances the tree
has to be optimized
21
Simulation Parent Selection Strategies
The effect of parent selection strategies on
overall system performance depends on the factor
of latency/loss rate in the utility
22
Simulation Off-Session Point Accumulation
Some typical peers wealth over time
Change from utility maximization mode to point
accumulation model (session ending time)
23
Simulation Off-Session Point accumulation
Systems resource increases as more peers choose
to stay online and contribute after the normal
session services.
24
Thank you!