1. An Active Reliable Multicast Framework for the Grids M. Maimour & C. Pham ICCS 2002, Amsterdam Network Support and Services for Computational Grids Sunday, April 21st, 2002 http://www.ens-lyon.fr/LIP/RESAM Action INRIA-RESO

2. 2 Outline Motivations behind (reliable) multicast Use of active networks : the DyRAM protocol DyRAM main services Simulation results Conclusion

3. 3 From unicast… Problem Sending same data to many receivers via unicast is inefficient. Sender data Receiver data

4. 4 …to multicast on the Internet. Sender data Receiver Problem Sending same data to many receivers via unicast is inefficient. Solution Using multicast is more efficient

5. 5 At the routing level, IP Multicast efficiently delivers packets to all the receivers subscribed to a multicast session but without any reliability guarantees. Reliability (including flow and congestion control) is to be addressed at the transport level. Reliable multicast

6. 6 Data replications Database updates Code & data transfers Data communications for distributed applications (collective & gather operations, sync. barrier) Data replications Database updates Code & data transfers Data communications for distributed applications (collective & gather operations, sync. barrier) Reliable multicast: a big win for grids Multicast address group 224.2.0.1 224.2.0.1 SDSC IBM SP 1024 procs 5x12x17 =1020 NCSA Origin Array 256+128+128 5x12x(4+2+2) =480 CPlant cluster 256 nodes

7. 7 Reliable multicast strategies End-to-end solutions : Only the end hosts (the source and/or the receivers) are involved. Problem : the lack of topology information at the end hosts. In-network solutions : Some intermediate nodes (router/server) are involved in the recovery process.

8. 8 Active networking solutions Active routers are able to perform customized computations on incoming packets: cache of data, feedback aggregation, filtering, subcasting, …

9. 9 The DyRAM framework for grids (Dynamic Replier Active Reliable Multicast) In order to enable distributed grid applications, main design goals are : low recovery latency using local recovery low memory usage in routers : local recovery is performed from the receivers (no cache in routers) low processing overheads in routers : light active services

10. 10 DyRAM loss recovery strategy : main active services DyRAM is NACK-based … Global NACK suppression Early packet loss detection Subcast of repair packets Dynamic replier election

11. 11 Global NACKs suppression NACK4 data4 NACK4 only one NACK is forwarded to the source

12. 12 Early loss packet detection NACK4 A NACK is sent by the router data3 data4 data5 The repair latency can be reduced if the lost packet could be requested as soon as possible These NACKs are ignored!

13. 13 Replier election A receiver is elected to be a replier for each lost packet (one recovery tree per packet) Load balancing can be taken into account for the replier election

14. Replier election and repair subcast IP multicast DyRAM DyRAM R1R1 R2R2 R3R3 R4R4 R5R5 R6R6 R7R7 0 1 2 10 NAK 2,@ NAK 2 from link 1 NAK 2 from link 2 NAK 2 Repair 2 D0D0 D1D1 NAK 2

15. core network Gbits rate 1000 Base FX active router 100 Base FX source The backbone is very fast so nothing else than fast forwarding functions. Nacks suppresion Subcast Loss detection A hierarchy of active routers can be used for processing specific functions at different layers of the hierarchy. Any receiver can be elected as a replier for a loss packet. Nacks suppression Subcast Replier election The DyRAM framework for grids

16. 16 Some simulation results Network model and metrics used Local recovery from the receivers DyRAM vs. ARM (cache in routers) DyRAM : early lost packet detection

17. 17 Network model 10 MBytes file transfer Source router

18. 18 Metrics Load at the source : the number of the retransmissions from the source. Load at the network : the consumed bandwidth. Completion time per packet (latency).

19. 19 Local recovery from the receivers (1) Local recoveries reduces the end- to-end delay (especially for high loss rates and a large number of receivers). #grp: 6…24 4 receivers/group p=0.25

20. 20 Local recovery from the receivers (2) As the group size increases, doing the recoveries from the receivers greatly reduces the bandwidth consumption 48 receivers distributed in g groups  #grp: 2…24

21. 21 DyRAM vs ARM ARM performs better than DyRAM only for very low loss rates and with considerable caching requirements

22. 22 DyRAM: early lost packet detection #grp: 6…24 4 receivers/group The end-to-end latency is decreased when the early lost packet detection is enabled

23. 23 Conclusions Reliability on large-scale multicast is difficult. Active services can provide more efficient solutions for reliable multicast related problems. Main DyRAM design goal is reducing the end- to-end latencies using active services which are keeped as light as possible making DyRAM more suitable to grid applications.