Scalable Application Layer Multicast Suman Banerjee Bobby Bhattacharjee Christopher Kommareddy ACM SIGCOMM Computer Communication Review, Proceedings of.

Scalable Application Layer Multicast

Suman Banerjee

Bobby Bhattacharjee

Christopher Kommareddy

ACM SIGCOMM Computer Communication Review , Proceedings of the 2002 conference on Applications, technologies, architectures, and protocols for computer communications,  Vol. 32 no. 4, Aug 2002

Outline

Introduction Hierarchical Membership Protocol Description Simulation Results Experiment Results

Introduction

Network layer multicast not widely deployed

Application layer multicast Do not change the network infrastructure Implement multicast forwarding functionality exclusively at

end-hosts

Introduction

NICE - NICE is the Internet Cooperative Environment

NICE is designed to Support large receiver sets Small control overhead Low latency distribution trees

Hierarchical Membership

Clients are assigned to different layers Each layer is partitioned into a set of clusters

of size between k and 3k – 1, where k is a constant parameter

All hosts belong to the lowest layer L0

The host with the minimum maximum distance to all other hosts in the cluster is chose to be the leader

Hierarchical Membership

Leaders of clusters of Li join layer Li+1

At most logkN layers Each host maintains state about all the cluste

rs it belongs to and about its super-cluster

Control Paths

Control paths Exchange periodic state refreshes For a host X, the peers on its control topology are

the other members of the clusters to which X belongs

E.g. For A0 – A1, A2, B0

For B0 – A0, A1, A2, B1, B2, C0

Data Paths

Data paths Source-specific tree Run the following algorithm

Data Paths

Host h received data from host p Forward the received packets to all clusters that h

belongs, except that of p

NICE Trees Analysis

Worst case control overhead of cluster-leader of the highest layer cluster O(k * logkN)

Average overhead

No. of hops on data path O(logkN)

Protocol Description

Assumption All hosts know the “Rendezvous Point” (RP) host RP is always the leader of the single cluster in the

highest layer RP interacts with other hosts on control path, but

not data path

New Host Joins

Join procedure Contact RP to get the cluster members of the

highest layer Loop until reach layer 0

Query the members of the returned cluster and find the closest one, X

Get the members of the child-cluster of X

New Host Joins

Overhead: O(k * logkN)

Latency: O(logkN) * RTT

New Host Joins

Cluster-leader may change as member joins or leaves

A change in leadership of a cluster C, in layer Lj Current leader of C removes itself from all layers

> Lj

Each affected layers choose a new leader The new leaders join their super-cluster

If the state of super-cluster is not locally available, contact RP

Cluster Maintenance and Refinement

Cluster-leader periodically checks the size of its cluster in layer Li

If the cluster size exceeds the 3k - 1 limit Split the cluster into two equal-sized clusters such that

the maximum of the radii among the two clusters is minimized

If the cluster size is under k The leader finds a closest host in layer Li+1 and merge

with it

Cluster Maintenance and Refinement

Each member, H, in any layer Li periodically probes all members in its super-cluster, to identify the closest member

If a host, J, that is closer to H is found, then H joins the cluster under the J

Host Departure and Leader Selection

Node H leaves Graceful leave

Send a leave message to all clusters it belongs Ungraceful leave

Other hosts detect the leave by not receiving the periodic refresh of H

If H is leader Each remaining member, J, select a new leader

independently Multiple leaders are resolved by the exchange of refreshes

Simulation Results

Simulation Results

Simulation Results

Simulation Results

Experimental Results

Experimental Results

Experimental Results

Experimental Results

Experimental Results

Experimental Results

Experimental Results

Conclusions

Proposed a new application layer multicast protocol Low control overhead Low link stress