PPT - Networking Research Lab @ SYSU

On the Efficiency of
Collaborative Caching in
ISP-aware P2P Networks
Jie Dai……Hai Jin et al.
H.K.U.S.T. U.T. H.U.S.T.
IEEE Infocom, Shanghai, China, April 10-15, 2011
Presenter: Su Hu
Warm-up

P2P: Overlay network
For the first 4 downloaded pieces,
the pieces are selected at random
Warm-up

Challenges
ISPs
Tremendous data volume
Costly inter-ISP traffic
1) Not at the same layer
P2P Overlay
Application data
Internet access service
ISP Underlay
Warm-up

Challenges
ISPs
Profit
2) Users pay for bandwidth, why throttling ?
bandwidth definition, local loop,
ADSL architecture ……
Shared bandwidth
Outline
1.
2.
I.
II.
III.
IV.
V.
VI.
VII.
Warm-up
Abstract
Introduction
Related Work
Inter-ISP Traffic Model & Cache Allocation
Improving Cache with ISP Peering Agreement
α, q, β, η, ISP
Performance Evaluation
index etc.
Conclusion
Summary
Abstract

Why Collaborative Cache
1)
Reduce the inter-ISP traffic

Existing design ignores:
1)
Dynamic P2P traffic patterns, ISP peering, cache server
capacity ….
Analysis of resource allocation with awareness of InterISP traffic and ISP policies
2)
Abstract

Our work
1)
Characterize inter-ISP traffic patterns
Develop cache allocation framework focus on minimizing
inter-ISP traffic.
Incorporate both locality-aware/unaware & ISP peering
agreements
2)
3)

The research help us understand
1)
Traffic characteristics of existing P2P
Design of collaborative ISP cache mechanisms
2)
Outline
1.
2.
I.
II.
III.
IV.
V.
VI.
VII.
Warm-up
Abstract
Introduction
Related Work
Inter-ISP Traffic Model & Cache Allocation
Improving Cache with ISP Peering Agreement
Performance Evaluation
Conclusion
Summary
Introduction

Background 1: The Tussle
1)
P2P: 70% of the Internet traffic
Can ISP throttle P2P packets?
ISP want to maintain customer bases
2)
3)

Background 2: How to resolve it
1)
Disparity
Locality-aware peer selection: P4P TopBT
vulnerable due to the dynamic of P2P
2)
3)
Proximity-driven
biased neighbor
select
Introduction

Reduce access latencies
to web page
Our Solution- caching
Cache for P2P
Web cache
2) Collaborative Caching lead to win-win:
Inter-ISP
Redirect traffic to cache server at
1)
edges of ISP
Experiences of User
Reduce the latency of P2P
packet
Introduction

New Characteristics from web cache
Both Storage (cache hit
Mitigate the inter-ISP traffic
ratio) & bandwidth (server’s
1) Inter-ISP traffic pattern, collaboration
uploadingbetween
capacity)
constraints are important.
P2P & ISP
2) Cache server resource allocation
3) ISP peering agreements
The collaboration between ISPs
over the public Internet &
corresponding cache server
Introduction
Video distribution platform

Propose a Optimization framework
1)
Theoretical model of i-ISP traffic
ISP scales , channel popularity
2)
Resource allocation scheme
Reduce i-ISP traffic both locality-aware/unaware
peer selection
3)
The effects of ISP peering on our solution
Positive on mitigation i-ISP traffic
4)
Collaborative cache scheme tailored to ISP
peering
Outline
1.
2.
I.
II.
III.
IV.
V.
VI.
VII.
Warm-up
Abstract
Introduction
Related Work
Inter-ISP Traffic Model & Cache Allocation
Improving Cache with ISP Peering Agreement
Performance Evaluation
Conclusion
Summary
Related Work

3 classes of ISP-friendly design
 Peer-driven
PPLive’s latency based mechanism, TCP ping
 ISP-driven
P4P: ISP advertise preferred paths to P2P app.
 Why
ISP caching?
Not impair the P2P robustness
Transparent to end user
Upon locality-aware system
Related Work

Existing P2P cache design
 Focus
on independent server cache
 Improving the byte hit ratio
 Ignore ISP collaboration & cache server bandwidth
constraint

Existing collaborative cache design
 Dan’s
work:
This paper: inter-ISP
Rate allocation among cache servers
traffic model, server
storage
Ignore inter-ISP traffic model, practical constraints
in and
real P2P
bandwidth constraints
,peer selection, ISP
peering
Outline
1.
2.
I.
II.
III.
IV.
V.
VI.
VII.
Warm-up
Abstract
Introduction
Related Work
Inter-ISP Traffic Model & Cache Allocation
Improving Cache with ISP Peering Agreement
Performance Evaluation
Conclusion
Summary
I-ISP Traffic Model & Cache Allocation
A.
Inter-ISP traffic model
 P2P
video streaming
 locality-aware
B.
locality-unaware
Optimization framework of allocation resource
 Inter-ISP
traffic mitigation
 Two sets of server strategies
 Collaboration
between P2P app. & cache server
I-ISP Traffic Model & Cache Allocation
A. Inter-ISP traffic model

Notation
 P2P video streaming
Assume
streaming length
is same, only
depend on
streaming rate
video channels
: number of concurrent users in P2P v system
: number of concurrent users in video channel i
: streaming rate of video channel i
: size of video channel I
: in-degree of individual peers
Assume peer outdegree equals indegree
I-ISP Traffic Model & Cache Allocation
A. Inter-ISP traffic model
 Notation
 Existing ISPs
: number of ISP in which peers view
video ?
ISP1 is most popular, ISPk is lest popular
: Storage capacity by cache server in ISP k
: uploading bandwidth by cache server in ISP k
: percentage of channel i stored in c server in ISP k
: uploading bandwidth to channel i by c server in ISP k
: number of concurrent users of channel i in ISP k
I-ISP Traffic Model & Cache Allocation
A. Inter-ISP traffic model

Probability that any
user view channel i
Channel popularity distribution
(1)
q
i

P2P object be accessed over long
term: Zipf-Mandelbrot distribution
the probability
the probability
ISP user distribution
(2)
Probability that any
user is in ISP k
β: different scenarios of ISP
user populations
β = 0, same user amount each ISP
higher the β, more unbalanced the ISP user
I-ISP Traffic Model & Cache Allocation
A. Inter-ISP traffic model
(3)
(4)

Inter-ISP traffic rate model
1. Locality-unaware peer
(n-c)Evenly selected, Neighbors
decides mainly by ISP user
selectionnumbers
m:number of neighbor in same ISP
Hyper-geometric distribution
I-ISP Traffic Model & Cache Allocation
A. Inter-ISP traffic model
M defectives in N, extract n
samples, and the
probability of k defectives
H(n , M , N)
p(x=k) = C(k , M) * C(n-k , N-M) / C(n , N)
k= max(0 , n-N+M) , …… , min(n , M)
N – xi M – xik n – din
p2p streaming
server is the external sources.
(5)
I-ISP Traffic Model & Cache Allocation
A. Inter-ISP traffic model
Inter-ISP generate by channel I in ISP k:
(6)
1) more popular channel more inter-ISP traffic
2) ISPs have similar scales,
3) ISPs have widely different scales,
I-ISP Traffic Model & Cache Allocation
A. Inter-ISP traffic model
 Inter-ISP traffic rate model
(n-c)
Give priority to nearby
peer (evaluate by the
ISP peer in)
2. Locality-aware peer selection
:number of persistent external links
i-ISP traffic per peer
i-ISP traffic per peer
(7)
I-ISP Traffic Model & Cache Allocation
A. Inter-ISP traffic model
Locality-aware
Locality-unaware
: 30
: 5-10
1.
= 80%, both have similar inter-ISP traffic
2.
-> 0 , both coefficients values -> 1
3. the left coefficients is always larger than the right
I-ISP Traffic Model & Cache Allocation
B. Cache resource allocation mechanisms
Inter-ISP traffic rate for ISP k:
Peers in any channel
are evenly distributed
along the channel ?
(8)
Maximize
Subject to:
Minimize
Subject to:
≤
(9)
(10)
I-ISP Traffic Model & Cache Allocation
B. Cache resource allocation mechanisms
Theorem 1
For max i-ISP mitigation, optimal resource allocation:
(11)
(12)
(13)
Continuous knapsack, solution:
Non-decreasing with index
Use greedy algorithm , give storage as needed for
channel with higher priorities, (11)
I-ISP Traffic Model & Cache Allocation
B. Cache resource
mechanisms
Achieve allocation
upper of
as min (
Theorem 1
Proof:
,
) using (12) , (13)
Maximize
Subject to:
(14)
I-ISP Traffic Model & Cache Allocation
B. Cache resource allocation mechanisms
Theorem 1 Remark:
Design guidelines of collaborative cache mechanism:
1. P2P system parameters:
number of users, channel popularity, file size,
streaming rate of channel
2. ISP cache server needs to collaborate with P2P app.
Precisely indentify the
content requests of
P2P packets needs
help of P2P app.
Reduce end-to-end
latencies,
Mitigate i-ISP prevents
throttling by ISP
I-ISP Traffic Model & Cache Allocation
B. Cache resource allocation mechanisms
Algorithm 1:
Optimization-based Collaborative Cache framework for
i-ISP mitigation
Population-based I,
Concurrent
users
1. P2P app. actively transmits system
states to
ISPx.
cache server.
2. Compute
,
, allocate
,,
as
,
3.
4.
Cache server cut request to external, if average
uploading rate to channel
, satisfy the request
Monitor P2P states, adjust resource according to T1.
Outline
1.
2.
I.
II.
III.
IV.
V.
VI.
VII.
Warm-up
Abstract
Introduction
Related Work
Inter-ISP Traffic Model & Cache Allocation
Improving Cache with ISP Peering Agreement
Performance Evaluation
Conclusion
Summary
Improve Cache with ISP Peering Agreement
A. ISP Peering Agreements

Concept
 ISPs
provide free connectivity to transit user
Free i-ISP
 Alleviate costly transit traffic
traffic is not

2 positive outcomes
need to cache
 Large
group of traffic-free candidate neighbor
 Strategically select P2P content to store and deliver

ISP peering relation is Reflexive & Symmetric
(15)
symmetric Matrix E
Improve Cache with ISP Peering
Agreement
Only peers in peering ISP
B. Impact of ISP Peering

help to mitigate i-ISP
traffic, no collaboration
(5) servers
between cache
Not-full collaboration between peering ISPs
 Cache
server not deliver to peers of peering ISP
 Locality-unaware peer selection
(16)
(17)
Improve Cache with ISPCompared
Peering
Agreement
to (6),
B. Impact of ISP Peering

here need to also subtract the
probability of being peering ISP
Not-full collaboration between peering ISPs
 Locality-unaware
peer selection (cont.)
(18)
 Locality-aware
peer selection
i-ISP traffic per peer
i-ISP traffic per peer
Multiply
not
(19)
Improve Cache with ISP Peering Agreement
B. Impact of ISP Peering
 For
both scenarios i-ISP traffic reduced due to
expansion of free neighbor candidates.
(18)
(19)
Improve Cache with ISP Peering Agreement
C. Improving cache with ISP Peering

Full collaboration between peering ISPs
 The
bottleneck
Cache server not
One ISP’s cache server can’t
store
only
serve whole
for peersP2P object
in own ISP, but also
-- Cache server bandwidth
utilization insufficient
to peering ISPs
 Peering: combine of global cooperative cache
 Peering-based full collaboration
: bandwidth assigned by
to
for channel i
<-----Upload rate for i
rate of i-ISP can be intercept(
)
Improve Cache with ISP Peering Agreement
C. Improving cache with ISP Peering

Full collaboration between
Any request to i can
be served
if sufficient bandwidth
peering
ISPs
(20)
Maximize
Subject to:
Peering, resource, limit aik to
serve max
, propose a
distributed collaborative cache
scheme in algor 2
Upper bound,
Centralized solution,
inappropriate for practice
(21)
Improve Cache with ISP Peering Agreement
Algorithm 2:
An ISP Collaboration-based Distributed Cache
framework for i-ISP mitigation
1. Cache server announce surplus bandw and storage
to peering ISPs.
2. After announce of
, sorts channel in descending
order of
,first channel ,
, bandw
request to
3. Upon receive r from ,
allocates and confirm
4. After confirm of ,
evicts content confirm,
reallocate to such
, broadcast surplus info
to peering ISP.
Outline
1.
2.
I.
II.
III.
IV.
V.
VI.
VII.
Warm-up
Abstract
Introduction
Related Work
Inter-ISP Traffic Model & Cache Allocation
Improving Cache with ISP Peering Agreement
Performance Evaluation
Conclusion
Summary
Performance Evaluation
A. Trace-Driven Analyses
Statistical result of measurement on UUSee:
Number of channels: 993
(channel 100 has 100 users at peak time)
Number of concurrent users: 100000
To fit the cure of peak time users:
α = 0.78 q = 4
= 30
η=5
B. Evaluation of Inter-ISP Traffic Pattern
Factors: P2P content popularity, ISP popularity
L-A(locality-aware) & L-U(locality-unaware)
Performance Evaluation
B. Evaluation of Inter-ISP Traffic Pattern
Fig.1.
Performance Evaluation
B. Evaluation of Inter-ISP Traffic Pattern
η
Fig.2.
Performance Evaluation
Fig.3.
Performance Evaluation
C. Evaluation of Collaborative Cache Mechanisms
Fig.4.
Performance Evaluation
C. Evaluation of Collaborative Cache Mechanisms
Fig.5.
Performance Evaluation
C. Evaluation of Collaborative Cache Mechanisms
Fig.6.
Performance Evaluation
D. Evaluation of ISP Peering Agreements
= 10 3 Peering Scenarios
1) Scenario 1:
1/2 3/4 … 9/10 extreme unbalanced
2) Scenario 2:
1/6 2/7 … 5/10
still has original property
3) Scenario 3:
1/10 2/9 … 5/6
extreme balanced
Performance Evaluation
D. Evaluation of ISP Peering Agreements
Fig.7.
Performance Evaluation
D. Evaluation of ISP Peering Agreements
Fig.8.
Performance Evaluation
D. Evaluation of ISP Peering
Fig.9.
About
percentage of
ISPs, so it
Agreements10
can’t reach 1
Outline
1.
2.
I.
II.
III.
IV.
V.
VI.
VII.
Warm-up
Abstract
Introduction
Related Work
Inter-ISP Traffic Model & Cache Allocation
Improving Cache with ISP Peering Agreement
Performance Evaluation
Conclusion
Summary
Conclusion





Propose an inter-ISP traffic model
Develop a cache resource framework under
resource constraint and peering agreement
Put forward guidelines for cache storage and
bandwidth allocation design
Strategy to improve collaborative cache under
ISP peering
Future work: improving user experience
Summary







Review P2P overlay and challenge with ISP
Review other existing ISP-friendly design
Give the notation used in this slide
Propose the inter-ISP traffic model
Give the Cache resource allocation mechanisms
Improve cache mechanisms with ISP peering
Evaluation of our collaborative cache mechanism
Good Points



Propose the probability model, summarize the
formulation of traffic under every strategy,
formulate the optimization problem
Rational performance analysis based on
experience data
Next : how to improve and implement it?