Fig. 1. Error recovery scheme. (a) NAKs originated from receivers. (1) Active router B receives a NAK from a downstream receiver. (2) When B finds that it is the first NAK with a different packet sequence number, B forwards the NAK to the upstream active router A. (3) A has the requested repair packet, it sends the repair packet only to the downstream link from which the NAK comes. (4) When B receives the repair packet, it buffers that packet and forwards the repair packet only to the link from which NAKs come according to the subscription information in the NAK record. (b) NAKs originated from active routers. (1) When active router B detects a packet sequence number gap, it sends a NAK to upstream nodes. (2) At the same time, B subcasts a FAKE_DATA packet to the subtree to suppress the NAKs generated by downstream active routers or receivers. (3) A has the requested repair packet, it sends the repair packet only to the downstream link from which the NAK comes. When B receives the repair packet, it subcasts the repair packet to the whole subtree.

Fig. 2. Congestion control and avoidance example. When there are buffers available in active routers, active router A dynamically selects the fastest receiver and slowest receiver from all the downstream receiver. It identifies the adjacent link from which feedback of the fastest receiver comes as the fastest link and the link from which feedback of the slowest receiver comes as the slowest link. Active router A forwards feedback from the fastest link to the source. It also adjusts the forwarding rate on all the adjacent downstream links so that each link will have different forwarding rates if the congestion level is different. Other active routers perform the same functions as the active router A.

Fig. 3. Packet format.

Fig. 4. Congestion window regulation algorithm in the source.

Fig. 5. Pseudo-code to regulate data packets at the source.

Fig. 6. Receiving window in the representative receiver.

Fig. 7. Pseudo-code to regulate repair packets.

Fig. 8. Pseudo-code to set the timer for data packets first sent.

Fig. 9. Congestion window regulation algorithm in an active router.

Fig. 10. Pseudo-code to regulate data packets at active routers.

Fig. 11. Analysis topology.

Fig. 12. Topology.

Fig. 13. AMCC competing with TCP, Droptail queue, No local recovery.

Fig. 14. Average throughput of AMCC flows and TCP flows, 6% loss in bottleneck, RTT = 240 ms.

Fig. 15. Adapting to different degrees of congestion.

Fig. 16. Interaction among multiple multicast flows, Droptail queue.

Fig. 17. Performance over multiple bottlenecks.

Fig. 18. Resisting the drop-to-zero problem.

Fig. 19. Per-node control, α=2.

Fig. 20. Topology.

Fig. 21. Inter-receiver fairness, α=100.

Fig. 22. Inter-receiver fairness, α=1000.

Operations at source, active routers and receivers in three operating modes