Scalable Applicaiton Layer Multicast

The (at least currently correct) assumption in this paper is that we are not going to have in network multicast anytime soon. Given this, the authors develop an application level multicast scheme that tries to get as close to 'ideal' in network performance.

The approach this paper takes is to build an overlay network out of participating applications and route multicast traffic using this overlay. The protocol they develop for this purpose is called NICE.

In NICE nodes are arranged in levels. Each level is grouped into nodes that are 'near' each other (distance defined by network distance and latency), and each group is assigned a 'leader' who is the graph theoretic center of the nodes in the group. This leader also participates in the next layer up with all the leaders of the groups below and a leader of that group is picked in the same manner. In this recursive fashion a tree is built which can be used to forward packets. To multicast a packet the leader of a group simply forwards to everyone else in the group. If any of those nodes are leaders of groups below them they will take the same action in those groups until everyone has the packet.

The authors evaluate the stress (how many extra packets need to go over a link compared to an in network scheme) and stretch (how many more hops packets take vs. in network) and also compare themselves to Narada, another application level multicast scheme. They come out ahead of Narada, but do impose significant overhead vs. in network schemes. The stress of roughly 1.6 in steady state would be a non-trival performance hit. Even worse the path length is much longer in NICE than in IP Multicast (about 23 hops vs. 15) which means many more chances for bottleneck links in the wide area.

I was also concerned about the join protocol. When a new node joins it finds a cluster close to it to join. If it is the center of the cluster is becomes the new leader. In the face of churn, many new leaders could generate a large amount of control traffic and an unstable tree. There was not really any evaluation of the robustness of NICE in the face of churn.