Fig. 1. Architectural choices for multicast (left to right): (a) infrastructure implements control plane as well as data forwarding, (b) end-hosts implement both control plane as well as data forwarding, (c) our architecture, where infrastructure nodes implement data forwarding, and end-hosts implement the control plane.

Fig. 2. (a) The receiver R inserts trigger (id, R). (b) The sender sends packet (id, data).

Fig. 3. Basic primitives. (a) Packet replication: Every packet (id, data) is forwarded to each receiver R_i that inserts trigger (id, R_i). (b) Anycast: The packet matches the trigger of receiver R2. id_pid_s denotes an ID of size m, where id_p represents the prefix of the k most significant bits, and id_s the suffix of the m − k least significant bits. (c) Trigger chains: Replace the receiver address in the second field of a trigger with another trigger ID.

Fig. 4. Joining algorithm.

Fig. 5. (a) R₇ wants to join an existing multicast tree. Joinable nodes are represented by empty circles; full nodes are represented by black circles. (b) The resulting tree obtained by running the algorithm in Fig. 4.

Fig. 7. The tree of triggers corresponding to the join operation shown in Fig. 5.

Fig. 6. Joining multicast protocol.

Fig. 8. Timing of messages for distance computation.

Fig. 9. Timing of refresh messages and refresh acks.

Fig. 10. Pseudocode of the recovery procedure.

Fig. 11. (a) The latency stretch of the multicast tree and (b) the maximum stress caused by the overlay multicast tree on the links of the underlying network.

Fig. 12. Maximum and average number of refreshes per refresh period.

Fig. 13. (a) Total number of duplicates as a fraction of total number of drops and (b) cumulative distribution of time to recover for 1024 receivers.

Fig. 14. (a) CDF of the stretch and (b) scatterplot of multicast latency vs. IP latency for PlanetLab experiments.

Fig. 15. CDF ratio of number of duplicates to number of losses at a node.

Number of repairs as a fraction of drop rate, and maximum number of anycast requests per sequence number for rel-random and rel-loss (standard deviation in parentheses)

Results from experiments on Millennium showing the depth of the tree and the number of query_distance messages processed (standard deviations in parentheses)