1 A Case For End System Multicast Yang-hua Chu, Sanjay Rao and Hui Zhang Carnegie Mellon University

2 Unicast Transmission End Systems Routers Gatech CMU Stanford Berkeley

3 IP Multicast No duplicate packets Highly efficient bandwidth usage Key Architectural Decision: Add support for multicast in IP layer Berkeley Gatech Stanford CMU Routers with multicast support

4 Key Concerns with IP Multicast Scalability with number of groups –Routers maintain per-group state –Analogous to per-flow state for QoS guarantees –Aggregation of multicast addresses is complicated Supporting higher level functionality is difficult –IP Multicast: best-effort multi-point delivery service –End systems responsible for handling higher level functionality –Reliability and congestion control for IP Multicast complicated Deployment is difficult and slow –ISP’s reluctant to turn on IP Multicast

5 The billion dollar question... Can we achieve efficient multi-point delivery, without support from the IP layer?

6 End System Multicast Stanford CMU Stan1 Stan2 Berk2 Overlay Tree Gatech Berk1 Berkeley Gatech Stan1 Stan2 Berk1 Berk2 CMU

7 Scalability –Routers do not maintain per-group state –End systems do, but they participate in very few groups Easier to deploy Potentially simplifies support for higher level functionality –Leverage computation and storage of end systems –For example, for buffering packets, transcoding, ACK aggregation –Leverage solutions for unicast congestion control and reliability Potential Benefits

8 What I hope to convince you of... End System Multicast is a promising alternative approach for multi-point delivery –Narada: A distributed protocol for constructing efficient overlay trees among end systems –Simulation and Internet evaluation results to demonstrate that Narada can achieve good performance Consider applications with small and sparse groups –Around tens to hundreds of members

9 Performance Concerns CMU Gatech Stan1 Stan2 Berk1 Berk2 Duplicate Packets: Bandwidth Wastage CMU Stan1 Stan2 Berk2 Gatech Berk1 Delay from CMU to Berk1 increases

10 What is an efficient overlay tree? The delay between the source and receivers is small Ideally, –The number of redundant packets on any physical link is low Heuristic we use: –Every member in the tree has a small degree –Degree chosen to reflect bandwidth of connection to Internet Gatech “Efficient” overlay CMU Berk2 Stan1 Stan2 Berk1 High degree (unicast) Berk2 Gatech Stan2 CMU Stan1 Stan2 High latency CMU Berk2 Gatech Stan1 Berk1

11 Why is self-organization hard? Dynamic changes in group membership –Members may join and leave dynamically –Members may die Limited knowledge of network conditions –Members do not know delay to each other when they join –Members probe each other to learn network related information –Overlay must self-improve as more information available Dynamic changes in network conditions –Delay between members may vary over time due to congestion

12 Berk2 Berk1 CMU Gatech Stan1 Stan2 Narada Design CMU Berk2 Gatech Berk1 Stan1 Stan2 Step 1 Source rooted shortest delay spanning trees of mesh Constructed using well known routing algorithms – Members have low degrees – Small delay from source to receivers “ Mesh”: Richer overlay that may have cycles and includes all group members Members have low degrees Shortest path delay between any pair of members along mesh is small Step 2

13 Narada Components Mesh Management: –Ensures mesh remains connected in face of membership changes Mesh Optimization: –Distributed heuristics for ensuring shortest path delay between members along the mesh is small Spanning tree construction: –Routing algorithms for constructing data-delivery trees –Distance vector routing, and reverse path forwarding

14 Optimizing Mesh Quality Members periodically probe other members at random New Link added if Utility Gain of adding link > Add Threshold Members periodically monitor existing links Existing Link dropped if Cost of dropping link < Drop Threshold Berk1 Stan2 CMU Gatech1 Stan1 Gatech2 A poor overlay topology

15 The terms defined Utility gain of adding a link based on –The number of members to which routing delay improves –How significant the improvement in delay to each member is Cost of dropping a link based on –The number of members to which routing delay increases, for either neighbor Add/Drop Thresholds are functions of: –Member’s estimation of group size –Current and maximum degree of member in the mesh

16 Desirable properties of heuristics Stability: A dropped link will not be immediately readded Partition Avoidance: A partition of the mesh is unlikely to be caused as a result of any single link being dropped Delay improves to Stan1, CMU but marginally. Do not add link! Delay improves to CMU, Gatech1 and significantly. Add link! Berk1 Stan2 CMU Gatech1 Stan1 Gatech2 Probe Berk1 Stan2 CMU Gatech1 Stan1 Gatech2 Probe

17 Used by Berk1 to reach only Gatech2 and vice versa. Drop!! An improved mesh !! Gatech1 Berk1 Stan2 CMU Stan1 Gatech2 Gatech1 Berk1 Stan2 CMU Stan1 Gatech2

18 Narada Evaluation Simulation experiments Evaluation of an implementation on the Internet

19 Performance Metrics Delay between members using Narada Stress, defined as the number of identical copies of a packet that traverse a physical link Berk2 Gatech Stan1 Stress = 2 CMU Stan2 Berk1 Berk2 CMU Stan1 Stan2 Gatech Berk1 Delay from CMU to Berk1 increases

20 Factors affecting performance Topology Model –Waxman Variant –Mapnet: Connectivity modeled after several ISP backbones –ASMap: Based on inter-domain Internet connectivity Topology Size –Between 64 and 1024 routers Group Size –Between 16 and 256 Fanout range –Number of neighbors each member tries to maintain in the mesh

21 Simulation Details Simulator –Packet-level and event-based –Models propagation delay of physical links –Does not model queuing delay and packet loss Individual Experiment Description –All group members join in random sequence in first 100 seconds –No change in group membership after 100 seconds –One sender picked at random and multicasts data at constant rate

22 Delay in typical run 4 x unicast delay1x unicast delay Waxman : 1024 routers, 3145 links Group Size : 128 Fanout Range : for all members

23 Naive Unicast Native Multicast Narada : 14-fold reduction in worst-case stress ! Stress in typical run

24 Variation with group size Waxman model: 1024 routers, 3145 links Fanout Range:

25 Variation with topology model Waxman ASMap Mapnet

26 Implementation Status Implemented and ported to Linux and Sun Available as library that can be compiled with applications Examining how applications written with IP Multicast API can be used without source-code modification

27 Internet Evaluation 13 hosts, all join the group at about the same time No further change in group membership Each member tries to maintain 2-4 neighbors in the mesh Host at CMU designated source Berkeley UCSB UIUC1 UIUC2 CMU1 CMU2 UKY UMas s GATech UDe l Virginia1 Virginia2 UWisc

28 Narada Delay Vs. Unicast Delay Internet Routing can be sub-optimal (ms) 2x unicast delay1x unicast delay

29 Related Work Yoid (Paul Francis, ACIRI) –More emphasis on architectural aspects, less on performance –Uses a shared tree among participating members More susceptible to a central point of failure Distributed heuristics for managing and optimizing a tree are more complicated as cycles must be avoided Scattercast (Chawathe et al, UC Berkeley) –Emphasis on infrastructural support and proxy-based multicast To us, an end system includes the notion of proxies –Also uses a mesh, but differences in protocol details

30 Conclusions Proposed in 1989, IP Multicast is not yet widely deployed –Per-group state, control state complexity and scaling concerns –Difficult to support higher layer functionality –Difficult to deploy, and get ISP’s to turn on IP Multicast Is IP the right layer for supporting multicast functionality? For small-sized groups, an end-system overlay approach –is feasible –has a low performance penalty compared to IP Multicast –has the potential to simplify support for higher layer functionality –allows for application-specific customizations