adaptive overlay topology for meshmesh--based p2p based
TRANSCRIPT
Adaptive Overlay Topology for Adaptive Overlay Topology for MeshMesh--Based P2PBased P2P--TV SystemsTV Systems
Richard Lobb (Univ. Canterbury, NZ)
Ana Paula Couto da Silva (Univ. Juiz de Fora, BR)
Michela Meo – NOSSDAV 2009
Ana Paula Couto da Silva (Univ. Juiz de Fora, BR)
Emilio LeonardiMarco MelliaMichela Meo
Politecnico di Torino, Italy
MeshMesh--BasedBased P2PP2P--TV TV SystemsSystems
• Distribute TV channels over the Internet in a P2P fashion
• Video is generated by a source and
Michela Meo – NOSSDAV 2009
• Video is generated by a source and received by users with short delay
• The video information is organized in small chunks that are individually and independently distributed by the participating peers
Michela Meo – NOSSDAV 2009
IP network
Overlay
Chunk 1 distribution
time
Michela Meo – NOSSDAV 2009
distributiontree
Chunk 2 distribution
tree
Chunk distribution processChunk distribution process
• Each peer maintains a (small) transmission window with chunks to be redistributed to other peersother peers
• A scheduling process decides which chunk to distribute to which neighboring peer
• The most critical phase of the chunk distribution is the initial one (when it is rare)
Michela Meo – NOSSDAV 2009
After a delayChunkSize/Bw1
1 more peer can contribute
P1
Chunk distribution processChunk distribution process
Q1
Michela Meo – NOSSDAV 2009
Q1
After a delay ChunkSize/Bw1
1 more peer can contribute
P1
Chunk distribution processChunk distribution process
Q1
Michela Meo – NOSSDAV 2009
After the same delayChunkSize/Bw1
3 more peers can contribute
P2P2 has 3 times the Bw of P1
Q1
Q1
Q2 Q3
Upload Upload bandwidthbandwidth matters…matters…
• Peers that can contribute more to the chunkdistribution (high bw peers) should be
�favored during the scheduling process�favored during the scheduling process
Michela Meo – NOSSDAV 2009
Upload Upload bandwidthbandwidth matters…matters…
• Peers that can contribute more to the chunkdistribution (high bw peers) should be
�favored during the scheduling process�favored during the scheduling process
Michela Meo – NOSSDAV 2009
High bw peers should bepreferentially served first
SchedulingScheduling isis notnot enoughenough
• Peers that can contribute more to the chunkdistribution (high bw peers) should also
�Have many neighbors (they can use their bw)
High bw, small out-degree:Too few neighbors to distribute the
chunk to, the bw is not well exploited
�Have many neighbors (they can use their bw)
�Be close to the source
�Be well connected to each other, to speed up the distribution between high bw peers
Michela Meo – NOSSDAV 2009
Low bw, large out-degree:Large neighborhood is useless for
distribution (little bw), while it is costly to manage
SchedulingScheduling isis notnot enoughenough
• Peers that can contribute more to the chunkdistribution (high bw peers) should also
�Have many neighbors (they can use their bw)�Have many neighbors (they can use their bw)
�Be close to the source
�Be well connected to each other, to speed up the distribution between high bw peers
Michela Meo – NOSSDAV 2009
Overlay construction and maintenance are
crucial
Our proposal key pointsOur proposal key points
• Overlay maintenance algorithm that
�Automatically adapts the out-degree (neighborhood size) to the actual capacity of (neighborhood size) to the actual capacity of the peer to contribute to chunk distribution
�Automatically makes high bw peers highly connected and close to the source
�Does not require the explicit estimation of peer bw
Michela Meo – NOSSDAV 2009
Set a time window
Every time window……compute the fraction of used links
Use this value as an indication of peer capacity
Michela Meo – NOSSDAV 2009
• Large? (the peer can contribute more)
Increase out-degree
• Small? (the neighborhood is too large)
Decrease out-degree
Use this value as an indication of peer capacityto contribute to chunk distribution
Set a time window.
Every time window……compute the fraction of used links
Use this value as an indication of peer capacity
Control the no. of unused links
Michela Meo – NOSSDAV 2009
• Large? (the peer can contribute more)
Increase neighborhood
• Small? (the neighborhood is too large)
Decrease neighborhood
Use this value as an indication of peer capacityto contribute
UUU HL αα <<
no. of used links no. of unused links
Control the no. of unused links
Implicit estimation of bandwidthImplicit estimation of bandwidth
• Fraction of used links is employed to adapt the out-degree to the capacity of the peer to contribute the peer to contribute
• The peer out-degree is then used as an implicit estimation of its bandwidth
�For scheduling
�When choosing/dropping peers
Michela Meo – NOSSDAV 2009
Growing the neighborhoodGrowing the neighborhood
• New peers are added choosing among the set of neighbors’ neighbors
• Within this set, peers are chosen with a probability proportional to the desirability
• Within this set, peers are chosen with a probability proportional to the desirability function
�The desirability function is monotonic increasing with out-degree
�Out-degree is thus used as an implicit measure of the peer upload bandwidth
Michela Meo – NOSSDAV 2009
Need to exchange the list of neighbors and
their out-degree
Growing the neighborhoodGrowing the neighborhood
• At peer p, neighborhood size L(p) increases by
� kL(p): k is the growth factor
• Initially, L(p) is small • Initially, L(p) is small
�Low bw peers are not congested
• Initially, k=ki
k linearly decreases to kf < ki
�High bw peers quickly grow their neighborhood
Michela Meo – NOSSDAV 2009
Shrinking the neighborhoodShrinking the neighborhood
• Cull from the neighborhood a number of links within the set of unused links, so that
• Set UU HL αα +=• Set
Michela Meo – NOSSDAV 2009
UU2
=
no. of desired unused links
UUU HL αα <<
no. of unused links no. of used links
Chunk schedulerChunk scheduler
• Choose the latest useful chunk (latest chunk needed by some neighbor)
• Send chunk to neighbor q with probability • Send chunk to neighbor q with probability proportional to the desirability function
Michela Meo – NOSSDAV 2009
∑∈
=
),(
)(
)()(
pcNr
rDqD
qp
desirability function
)()( qLqD =
set of neighbors needing the chunk
Scenario & evaluationScenario & evaluation
• Discrete event simulator
Download P2PTVSim from www.napa-wine.eu
• 10,000 chunks (1000s at 10 chunk/s rate)• 10,000 chunks (1000s at 10 chunk/s rate)
• N=10,000 peers, partitioned in 4 classes
� Class 1 (10%): Bw=5Mbps
� Class 2 (40%): Bw=1Mbps
� Class 3 (40%): Bw=0.5Mbps
� Class 4 (10%): Bw=0Mbps
Michela Meo – NOSSDAV 2009
E[Bw]=1.1Mbps(video rate=1Mbps)
Other parametersOther parameters
• Chunk size 0.1Mb, 10 chunks/s
• Playout delay=5s (50 chunks)
• Start with L(p)=10• Start with L(p)=10
• Time window size: 50 chunks
• ki=0.4, kf=0.1, after 750 chunks
• aL=0.1, aH=0.3
Michela Meo – NOSSDAV 2009
THE ADAPTIVE OVERLAYTHE ADAPTIVE OVERLAY
Michela Meo – NOSSDAV 2009
degr
ee
80
100
120
140
Bw=5Mbps
NeighborhoodNeighborhood
Michela Meo – NOSSDAV 2009
Out
-degr
ee
Chunk0 2000 4000 6000 8000 10000
20
40
60
Bw=1Mbps
short adaptationtime (2 min)
NeighborhoodNeighborhood
Per class: percentage of links to other classes
Michela Meo – NOSSDAV 2009
Peers prefer to be connected to high bw peers (10% of total only)
PERFORMANCEPERFORMANCE
Michela Meo – NOSSDAV 2009
DelayDelay
0.8
1.0
1.2
Michela Meo – NOSSDAV 2009
High bw peers are favored
Chunk
Dela
y
0 500 1000 1500 2000 2500 30000
0.2
0.4
0.6
LossesLosses
600
800
1000of
loss
es
Michela Meo – NOSSDAV 2009
0
200
400
0 50 100 150 200 250 300
Num
ber
Chunk
Only during the initial transient
ImprovementImprovementDela
y[s
]
1.0
1.5
High bw peers - more neigh.- highly connected
Michela Meo – NOSSDAV 2009
Dela
y
Degree, k10 20 30 40 50 60 70 80 90 100
0
0.5
Tot no. links smaller of a
factor 4
ChurningChurning
60
70
80
• 50% of the peers disconnect after a random time in [10,100]s
Robust to churning
Michela Meo – NOSSDAV 2009
Degr
ee
Chunk0 2000 4000 6000 8000 10,000
10
20
30
40
50
60
ChurningChurning
0.8
1.0
Michela Meo – NOSSDAV 2009
Chunk
Dela
y
0 500 1000 1500 2000 2500 30000
0.2
0.4
0.6
ConclusionsConclusions
• Effective in adapting the overlay to peer heterogeneity
� Neighborhood size and topology
• Robust to varying conditions •� Churning
� Available bw variations
• Simple and does not need bw estimation
• Requires limited exchange of information
� List of neighbors and their out-degree
Michela Meo – NOSSDAV 2009
Impact Impact ofof loadload
dela
y
6
7
8
9
10Loss, dmax=5Delay, dmax=5Loss, dmax=10Delay, dmax=10
Michela Meo – NOSSDAV 2009
00.9 1.0 1.05
Los
ses/
dela
y
Load0.95 1.1
1
2
3
4
5
6
Sensitivity to parametersSensitivity to parameters
ααααH
Michela Meo – NOSSDAV 2009
ααααL
Delay decreases with large no. unused links