Node Selection For a Fault-Tolerant Streaming Service On A Peer-to-Peer Network

1
Node Selection For a Fault-Tolerant Streaming
Service On A Peer-to-Peer Network

• Hyunjoo Kim, Sooyong Kang and Yeom H.Y.

2
Agenda

• Introduction
• P2P Service Model
• Server Selection Schemes
• Larger available outbound Bandwidth First (LBF)
• Smaller available outbound Bandwidth First (SBF)
• Playback Node First (PNF)
• Playback Node First with Prefetching (PNF-P)
• Simulation Results
• Conclusion

3
Introduction

• QoS degradation due to node failures is critical
  in streaming services
• Schemes to cope with node/link failures
• Redundant data
• Reallocation of data sending rate from sever
  nodes
• Special encoding for graceful quality degradation
• Service migration

4
Introduction

• The paper focus on server selection schemes
• Nodes are autonomous and have different bandwidth
• QoS depends greatly on the selection scheme
• The Fact
• Failure probability of a node currently being
  served is lower than that of a node not being
  served

5
P2P Service Model

• Assumptions
• A node can receive/transmit data from/to multiple
  nodes simultaneously
• Client node knows all the data placement
• CBR media objects which segmented into equal size

6
P2P Service Model

• When server receive a request, tell client the
  available outbound bandwidth
• Based on the selection scheme, tell the server
  whether it is selected
• If the server is not selected, it send update
  information to client when available bandwidth
  changes

7
P2P Service Model

• For each selected server, i
• Prefetch for td seconds
• Store tdri amount of data for each server
• Monitor the size of buffer to detect fault
• Fault detection threshold,
• sd is segment size, d is predicted migration time
• Replace the failed node by one or more server
  nodes

8
Server Selection Schemes

• Larger available outbound Bandwidth First (LBF)
• Smaller available outbound Bandwidth First (SBF)
• Playback Node First (PNF)
• Playback Node First with Prefetching (PNF-P)

9
Larger available bandwidth first (LBF)

• Select the nodes that can provide larger
  available outbound bandwidth first
• Minimize the number of servers needed
• Probability of encountering a node/link failure
  is smaller
• But individual node/link failure probability is
  still the same
• Fault recovery generally involve multiple nodes,
  so overhead is larger

10
Smaller available bandwidth first (SBF)

• Select the nodes that can provide smaller
  available outbound bandwidth first
• Maximize the number of servers needed
• Probability of encountering a node/link failure
  is larger
• But individual node/link failure probability is
  still the same
• Fault recovery generally involve one node only,
  so overhead is smaller

11
Playback Node First (PNF)

• The Fact
• Failure probability of a node currently being
  served is lower than that of a node not being
  served
• Based on LBF but nodes that are currently being
  served is selected first
• Node that has the longer remaining playback time
  is selected first

12
Playback Node First (PNF)
Request for 300Kbps Selection N0(150), N1(70),
N2(80)
13
Playback Node First with Prefetching (PNF-P)

• Node failure probability increases after the
  server finishes its own service
• Client node receive future data in advance into
  the storage
• Use PNF for selecting one or more servers for
  prefetching
• Prefetch data are transmit in reverse direction,
  from end of segment to the start

14
Simulation Results
Parameters Values
Number of initial servers 20
Outbound bandwidth 30 100 Kbps
Number of objects 20
Object playback rate 300 Kbps
Segment size 10 Kb
Object size 3600 seconds
Buffering time 5 seconds
Predicted migration time (d) 2 seconds
Request arrival rate Exp. Distr. w/ mean 0.012
Network congestion interval 10000 seconds
Network congestion duration 50 seconds
Normal state migration time Exp. Distr. w/ mean 0.1 sec
Congestion state migration time Exp. Distr. w/ mean 1 sec
MTBF 6250 200000 seconds
Simulation time 86400 seconds (24 hours)
15
Simulation Results
16
Simulation Results
17
Simulation Results
18
Simulation Results
19
Conclusion

• New node selection schemes for a service
  migration-based fault-tolerant streaming service
  on P2P networks.
• PNF and PNF-P decrease the average number of node
  failures and increase the probability of stable
  service.

20
Comments

• The Fact is the fact?
• Prefetching method is not good enough