Scalable and Continuous Media Streaming
on Peer-to-Peer Networks
Masahiro Sasabe
Osaka University, Japan
m-sasabe@ist.osaka-u.ac.jp
2003/9/19
P2P2003
1
Content
•
•
•
•
Backgrounds
Media Streaming on pure Peer-to-Peer Networks
Research Targets
Outline of Proposed Mechanisms
– Scalable Search Methods
– Block Retrieval Methods
• Simulation Results
• Conclusion and Future Work
2003/9/19
P2P2003
2
Backgrounds
- server-client model Load concentration
Client
Client
Client
Client
Media
Media
Media
Media server
server
High transfer delay
Diverse user demands
Client
Client
Client
Client
P2P is a network paradigm to solve these problems
2003/9/19
P2P2003
3
Backgrounds
- related works on P2P media streaming -
• Most of research works on P2P media streaming
construct an application-level multicast tree
– Effective for live-media streaming
– Not effective for on-demand media streaming
– The root of the tree is a single point of failure
• We discuss effective methods for on-demand
media streaming on pure P2P networks
2003/9/19
P2P2003
4
Media Streaming on pure P2P Networks
Flooding
Peer A
3. Forward
Peer B
Peer C
3. Forward
Peer D
3. Forward
4. Response
4. Response
No server
5. Request
Peer E
3. Forward
2. Query
1. Join
6. Transmit
New peer
Peer F
Search link
Data link
2003/9/19
Desired media stream
P2P2003
5
Media Streaming on pure P2P Networks
Problem:
Problem: Scalability
Scalability
Peer A
How
How can
can we
we get
get information
information
Peer about
B
about desired
desired media
media stream
stream
with
with less
less number
number of
of queries?
queries?
3. Forward
Peer D
Flooding
3. Forward
Peer C
3. Forward
4. Response
4. Response
No server
5. Request
Peer E
Search link
Data link
2003/9/19
3. Forward
2. Query
1. Join
6. Transmit
Problem:
Problem: Continuity
Continuity
Peer F
How
How should
should we
we determine
determine aa
provider
provider peer
peer for
for continuous
continuous
media
media play-out?
play-out?
P2P2003
New peer
Desired media stream
6
Research Targets
• We propose efficient mechanisms for scalable and
continuous media streaming
– Segmentation of media stream
• For efficient use of network bandwidth and cache buffer
– Scalable block-search method
• To solve the scalability problem of search on pure P2P networks
– Provider peer determination algorithm
• To achieve continuous media play-out
2003/9/19
P2P2003
7
Outline of Proposed Mechanisms
- per-group block search and retrieval Logical topology
Pd has blocks1, 4, 5
Block
Pd
Pc
Pc has blocks 6, 7
Pb has blocks 1, 3
Pb
Pa has block 2
Pa
Consumer
C
Blocks 1~4
Time
Play-out timing
1 block time
Round
2003/9/19
P2P2003
8
Advantage of Per-group Based Search
• There is a temporal order of reference in a media
stream
– A user watches a media stream from the beginning to
the end
• We can expect that a peer, which replies a
response message in the current round, has some
blocks of the next round
– A peer can estimate candidates of provider for blocks of
the next round from the search results in the current
round
2003/9/19
P2P2003
9
Scalable Search Methods
••
Full
Full Flooding
Flooding
–– Flooding
Flooding with
with static
static value
value of
of TTL
TTL
••
Limited
Limited Flooding
Flooding
–– Flooding
Flooding with
with limited
limited TTL
TTL based
based on
on
the
the search
search results
results at
at the
the previous
previous
round
round
••
Selective
Selective Search
Search
–– Send
Send queries
queries to
to particular
particular peers
peers
based
based on
on the
the search
search results
results at
at the
the
previous
round
previous round
Consider pros and
cons of them
••
FL
FL method
method
–– Combination
Combination of
of full
full flooding
flooding and
and
limited
limited flooding
flooding
••
FLS
FLS method
method
–– Combination
Combination of
of full
full flooding,
flooding, limited
limited
flooding,
flooding, and
and selective
selective search
search
2003/9/19
P2P2003
Desired media stream
10
Provider Peer Determination Algorithm
•
Every time a peer receives a response message,
it determines a optimum provider peer for every block in
current round.
1.
2.
Calculate a set of peers from which it can retrieve a block in time.
Select a peer from the set
a.
a. Select
Select aa peer
peer whose
whose estimated
estimated retrieval
retrieval time
time is
is the
the smallest
smallest among
among peers
peers
in
in the
the set
set （SF
（SF Method)
Method)
b.
b. Select
Select aa peer
peer with
with the
the lowest
lowest possibility
possibility of
of block
block disappearance
disappearance among
among
peers
peers in
in the
the set
set (SR
(SR Method)
Method)
Peer A
Cache
Peer B
Cache
Consumer
High Priority Low
2003/9/19
High Priority Low
P2P2003
11
Provider Peer Determination Algorithm (detail)
1.
1.
2.
2.
jj ←
← rr
Calculate
Calculate set
set S,
S, aa set
set of
of peers
peers
having
having block
block jj
a.
a.
b.
b.
3.
3.
4.
4.
••
••
ifif S=
S=φ
φ
Tf(j)
Tf(j)←
←Tp(j),
Tp(j), jj←
←j+1,
j+1, repeat
repeat Step2
Step2
ifif S
S≠φ
≠φ go
go to
to Step3
Step3
5.
5.
Derive
Derive set
set S’,
S’, aa set
set of
of peers
peers from
from
which
which aa provider
provider peer
peer can
can retrieve
retrieve
block
block jj by
by deadline
deadline Tp(j)
Tp(j)
max(Tf(j-1),
max(Tf(j-1),
Tnow+R(i))+B(j)/A(i)
Tnow+R(i))+B(j)/A(i) <=
<= Tp(j)
Tp(j)
a.
a.
b.
b.
ifif S’
S’ ≠φ
≠φ go
go to
to Step4
Step4
ifif S’
S’ == φ
φ
Tf(j)
Tf(j)←
←Tp(j),
Tp(j), jj←
←j+1,
j+1,
go
go back
back to
to Step2
Step2
2003/9/19
Determine
Determine provider
provider peer
peer P(j)
P(j) from
from
S’
S’
6.
6.
P2P2003
SF
SF (Select
(Select Fastest)
Fastest) Method
Method
SR
SR (Select
(Select Reliable)
Reliable) Method
Method
Derive
Derive the
the estimated
estimated completion
completion
time
time of
of retrieval
retrieval Tf(j)
Tf(j) and
and the
the time
time
Tr(i)
Tr(i) to
to send
send aa request
request for
for block
block jj
Tf(j)
Tf(j) == max(Tf(j-1),Tnow+R(i))
max(Tf(j-1),Tnow+R(i))
+B(j)/A(P(i))
+B(j)/A(P(i))
Tr(j)
Tr(j) == Tf(j)
Tf(j) –– R(P(j))
R(P(j)) -- B(j)/A(P(i))
B(j)/A(P(i))
jfjf jj == kN,
kN, finish
finish
otherwise
otherwise jj ←
← jj ++ 1,
1,
go
go back
back to
to Step2
Step2
12
Simulation Model
• Random network with 100 peers
• 40 media streams whose popularity follows a Zipf-like distribution with
α=1.0
• The inter-arrival time between two successive requests for the first
media stream follows the exponential distribution whose average is 20
minutes
• Cache replacement algorithm is LRU
Reference frequency Pi
82
55
18
20
Zipf distribution
21
26
37
15
45
94
63
52
40
32
14
64
44
46
30
66
99
12
76
11
17
77
9
1
Pi = α
i
4
81
10
28
2
87
60
31
92
27
73
35
91
13
98
58
88
62
71
70
33
93
19
65
8
23
39
22
6
50
84
83
59
38
42
16
61
49
89
1
95
69
78
47
74
72
24
86
7
36
68
34
54
80
5
85
43
48
56
25
53
51
79
41
57
90
67
29
97
96
0
75
3
High
2003/9/19
P2P2003
Media popularity
Low
13
Evaluation Criteria
• System scalability
– Average number of queries that a peer receives during
the simulation
• Continuity of media play-out
– Completeness of block retrieval
Number of retrieved blocks in time
Completeness =
Number of blocks in a media stream
2003/9/19
P2P2003
14
Simulation Results
- number of queries -
SF method
SR method
Number of Queries
Number of Queries
160000 Full flooding
140000 FL method
FLS method
120000
100000
80000
60000
40000
20000
0
0
160000 Full flooding
140000 FL method
FLS method
120000
100000
80000
60000
40000
20000
20000
2003/9/19
40000 60000
Time [sec]
80000
P2P2003
0
0
20000
40000 60000
Time [sec]
80000
15
Simulation Results
- completeness with 95% confidence interval SR method
1
1
0.8
0.8
Completeness
Completeness
70% of total requests
SF method
0.6
0.4
0.2 Full flooding
FL method
FLS method
0
0
5 10 15
High
2003/9/19
0.6
0.4
Full flooding
FL method
FLS method
0.2
20 25
Media
Popularity
30
0
35
Low
P2P2003
0
5
High
10
15
20 25
Media
Popularity
30
35
Low
16
Conclusion and Future Work
• Conclusion
– We discuss the media streaming on pure P2P networks.
– We proposed block search and retrieval methods for continuous and
scalable media streaming.
– Through simulation experiments, we have shown that the FLS
method can provide users with continuous media play-out without
introducing extra load on the system for popular media streams.
• Future Work
– Improve the completeness of unpopular media streams
•• We
We are
are now
now considering
considering an
an effective
effective cache
cache replacement
replacement algorithm,
algorithm,
which
which takes
takes into
into account
account supply
supply and
and demand
demand for
for media
media streams.
streams.
– Evaluate proposed mechanisms in more realistic situations where
network conditions and peer locations dynamically change.
2003/9/19
P2P2003
17
Appendix: Simulation Results
- LRU v.s. proposed cache replacement algorithm FLS method, SR method
1
1
0.8
0.8
Completeness
Completeness
FLS method, SF method
0.6
0.4
LRU
Proposed
0.2
0
0
5
10
High
2003/9/19
0.6
0.4
LRU
Proposed
0.2
15
20 25
Media
Popularity
30
0
35
Low
P2P2003
0
5
High
10
15
20 25
Media
Popularity
30
35
Low
18