Multicast for Video Streaming EE290T Spring 2002 Puneet Mehra [email protected] IP Multicast Overview Semantics Approach 1 -> Many or Many -> Many Build tree connecting source and receivers on Current Infrastructure in Net  Group Addressing – provides flexibility Packets delivered throughout tree. Dynamic changes to tree Receivers/senders unaware of each other New Receiver -> graft path onto tree Receiver leaving -> pruning path from tree Uses UDP – so no reliability Challenges Efficient routing of data to receivers Video Multicast Over Net Issues in Multicast over Best Effort Approaches to Multicast Fixed Frame Rate – regardless of delay/jitter Losses – degradation, possibly ungraceful Heterogeneity of receivers QoS resource reservation for Multicast Adaptive Rate Control Techniques for Rate Adaptation Single Stream Video Multicast Replicated Stream Video Multicast Layered Video Multicast Single Stream Video Multicast Only send 1 stream to all receivers. Pros: Cons: Easy To Implement Ignores Receiver Heterogeneity Feedback Implosion INRIA Video Conferencing System Feedback Problem handled through probabilistic receiver response Tradeoff granularity of control vs B/W efficiency Efficiency Tradeoff in Single Stream Approach Replicated-Stream Video Multicast Destination Set Group (DSG) Pros: Small # of video streams of varying quality sent to different multicast groups Intra-stream Rate control to adjust stream rate by receivers Inter-stream protocol used by receivers to switch streams deals with heterogeneity – more fair Scalable since receiver-driven Cons: Network carries redundant info Layered Video Multicast Receiver-Driven Layered Multicast (RLM) Send different “layers” to multicast groups, and receiver subscribes as needed -> scalable solution Congestion -> layer dropping Spare B/W -> layer adding Receivers conduct group join experiments and share info with others. Layered Video Multicast Cont. Layered Video Multicast with Retrans. (LVMR) Improve reception w/in a layer by retransmission Deal w/ congestion using Hierarchical Rate Control Hierarchical Rate Control (HRC) Congestion info distributed at both sender/receivers Intelligent partitioning of info -> concurrent experiments w/ less overhead Use hierarchy to only inform those who need to know about an experiment – affected regions Collaborative layer drop – better approach to congestion Error Control in Video Multicast Pure FEC ARQ – From LVMR Local Recovery - designated receivers at each level in tree help w/ rtx. of pkts -> lower latency Don’t rtx packets past deadline Receivers can trade reliability/latency by picking parent with desired attributes Multicast Routing [2,3] Routing – construct efficient tree from source to receivers Theoretical Results  Steiner Tree – minimize total cost of a multicast tree. NP-Complete. So use heuristics to provide a “good” approx. to Steiner Tree. Constrained Steiner Tree – impose b/w delay constraints on links to receivers. Also NPComplete. So must use heuristics All practical algorithms based on shortest path tree – minimize sum of weights on links along each path from source to receiver Intra-Domain Routing Source-based Routing Tree rooted at source Dense-mode routing – works best when topology densely populated with receivers Core-based Approach Select a Rendezvous Point (RP) to root the tree Sparse Mode Routing – More efficient than dense mode when few, wide-spread receivers Dense Mode Protocols Distance Vector Multicast Routing Protocol Uses broadcast & prune technique to build reverse shortest path trees (RSP) Steps: A lot of state info kept in ALL routers in net. Multicast extensions to OSPF Src bcasts pkt on Lan. Local router fwds pkt on all ifaces If pkt received on RPF iface, then it is forwarded. Leaf routers send prune toward src if no attached receivers Prune message forwarded to source, and send own prune if receive prune message on all ifaces. Uses IGMP locally, then floods info along with link state to net. PIM-DM Less complex than DVMRP since no RPF check is done. More inefficient as a result Tree Construction in DVMRP • S = Source. Black Circles = Receivers • Periodically flood net w/ datagrams • Leaf routers send prune toward source if there are no group members on leaf subnet • Final Tree is shown in (d). Core-Based Routing General Approach A core, or rendezvous point (RP) is configured for a multicast group Info about the RP & mapping from group to RP is discovered by routers using bootstrap protocol (also finds alternate RP in case of failure) Receivers explicitly join tree -> contact RP Src sends data to RP which sends down tree More efficient since state only kept in routers on path from src/receivers to RP. Examples CBT – Core-Based Trees PIM-SM – Protocol Independent Multicast/Sparse Mode Tree construction in CBT The Join Process for a new node •Receiver Contacts Local Router • Router sends JOIN_REQUEST to the core router • When msg reaches on-tree router, a JOIN_ACK is sent back • every router receiving JOIN_ACK updates state information • Periodically send echo-request to parent router. If echo not received in time, then router sends quit-notification upstream and deletes state information. Inter-Domain Routing Probs w/ multicast described Large flat topology -> complexity and instability since no BGP-like protocol No mechanism to build hierarchical mcast routing Solution – Immediate Future Introduce Hierarchy – multi-protocol extensions to BGP (MBGP) Each router only knows topology of its own domain & how to reach other domains Used to determine next hop for a host Inter-Domain Routing Cont. What if you have a src in one domain & receivers in others? Multicast Source Discovery Protocol When src registers w/ RP -> a source active (SA) msg is sent to MSDP peers Prevent loops w/ per-RPF flooding (ie: if msg received on correct iface -> flood) If MSDP is aware of local group members (use IGMP), then it will send a join to the src Long-Term Inter-Domain Proposals Border Gateway Multicast Protocol Bidirectional shared trees between domains with single root. Need strict allocation of addresses among domains. Address Allocation Protocols Multi Address Set Claim – Helps allocate addresses dynamically across domains GLOP – a “glop” of addresses statically allocated among domains Problems Deploying IP Multicast  Complexity Makes old routers useless disrupts ISP router migration model (routers generally migrate from core to edge) Domain Independence Can’t put it in core routers Hardware more difficult to manage (probs w/ firewalls) ISPs don’t want to rely on remote RPs Don’t want to be RP for non-customers Security – anyone can send/listen Address Allocation – anyone can pick a class D addr. References  “Video Multicast over the Internet.” Xue Li et al. IEEE Network. 1999.  “The Evolution of Multicast: From the MBone to Interdomain Multicast to Internet2 Deployment.” Kevin Almeroth. IEEE Network. 2000.  “Multicast Routing and Its QoS Extension: Problems, Algorithms, and Protocols.” Bin Wang and Jennifer C. Hou. IEEE Network. 2000.  “Deployment Issues for the IP Multicast Service and Architecture.” Christophe Diot et Al. IEEE Network. 2000.