An Evaluation of Scalable Application-level Multicast Using Peer-to-peer Overlays

1
An Evaluation of Scalable Application-level
Multicast Using Peer-to-peer Overlays

• Miguel Castro, Michael B. Jones, Anne-Marie
  Kermarrec, Antony Rowstron , Marvin Theimer,
  Helen Wang and Alec Wolman
• Presented by Ricky Taing

Authors
2
Outline

• Motivations and Goals of Paper
• Overview of Overlay Networks
• Overview of p2p Multicast Implementations
• Experimental Methodology
• Results
• Conclusions

3
Motivations and Goals

• Lack of IP multicast adoption has increased
  development in app-layer multicast
• Different versions available to use, with
  different implementations
• Determine the best application layer multicast
  system from different overlays and implementations

4
P2P Overlay Networks

• Two main approaches
• Divide and conquer in a ring
• Pastry, Chord, Tapestry
• Cartesian hyper-space
• CAN

5
Pastry

• Routes by nodeID
• Circular 128-bit namespace
• Get to destination in log 2b N time
• b is configurable, usually b4 (hex)
• Each node maintains a leaf set
• pointers to l nodes closest to it

6
CAN

• Route in multiple dimensions
• Each node is assigned a particular zone
• Many optimizations
• Neighbor with lowest network delay
• Multiple nodes per zone
• Uniform partitioning
• Landmark based placement

7
Landmark based placement

• Set of well known landmarks
• ordered by distance
• placed into evenly sized bins
• Nodes with same landmark ordering end up close to
  each other

8
P2P Multicast Implementations

• Flooding
• Unique overlay-per-group
• Only nodes in group get group messages
• CAN
• broadcast to neighbors, use seq numbers
• Pastry
• Forwards copies to all nodes in routing table
• Notes levels, sends to greater levels

9
Tree-based

• Scribe
• Reverse path forwarding to create one tree per
  group
• Joins route messages to groupId (root)
• Registers if a node along the route is in the tree

10
Evaluation

• Simulation
• Five different topologies
• 5050 routers, 80000 end nodes
• Two sets of experiments
• Single multicast group, all nodes are members
• Large number of groups (1500)

11
Criteria

• Relative Delay Penalty
• RMD Ratio of Maximum delay between app and IP
  multicast
• RAD Ratio of Average delay between app and IP
  Multicast
• Link Stress
• number of packets sent over link

12
Criteria (2)

• Node Stress
• Number of nodes in a routing table
• Number of messages received by a join
• Duplicates
• Number received by end nodes

13
CANFlooding Results

• Landmark based and NDR (lowest network delay) was
  best
• Benefits from increased table state is uneven
• Link stress for 80,000 joins is huge
• Increases with state size
• Link stress is significantly lower for sent
  messages

14
Link stress CANFlooding
15
CANTree Results

• Landmark based assignment of nodes better
• Delay is better than flooding by a factor of 2 to
  3
• Link stress for joins and sends are relatively
  similar

16
Pastry Tart and Tangy

• varied b from 1 to 4
• TART
• Topology aware routing table construction
• default optimization
• nodes probe each other to estimate delay
• TOP
• Topology aware nodeId assignment
• currently random / distributed

17
PastryFlooding Results

• delay decreases as b increases
• delay of b4 is 50 lower than b1
• TART and TOP both decrease delay
• TOP reduces average link stress by factor of 3
  and max link stress by a factor of 30
• Large number of duplicates when b is large (16
  b4)
• due to holes in routing tables

18
PastryTree Results

• delay decreases as b increases, and with TART and
  TOP
• similar to flooding results
• TART reduces both max and avg link stress
• TOP reduces avg but increases max link stress
• Pastry with TART and without TOP is best for
  tree-based

19
Multiple multicast

• Tree based Pastry was best for delay
• Interesting is CAN Flooding average CDF curve
  not tightly bound
• Flooding is better in the 1500 group experiment
  than the single group

20
Max Delay Penalties
21
Average Delay Penalties
22
Evaluation

• For delay, Pastry is 20-50 better than CAN
• For average link stress, Pastry was 15 lower
• Max link stress CAN was 25 lower

23
Conclusion

• Separate overlays are only better if you want to
  limit nodes that route traffic
• More overhead for flooding approach
• Tree base can reuse the same overlay for multiple
  groups
• Less delay, joins and sends are more lightweight
• Multicast trees with Pastry have better
  performance than CAN