1. 1 QoS Adaptive Group Communication Antonio Di Ferdinando, Paul D Ezhilchelvan and Isi Mitrani (with inputs from Jon Crowcroft and Panos Gevros)

2. Tapas’032 The Talk outline Generic Definition of a group Our approach to Adaptive QoS Provisioning First Step: Reliable Multicast QoS Adaptability Features Future work

3. Tapas’033 Group: A definition Set of distributed members that cooperate On the delivery of sent messages (notification of generated events) Varied forms of delivery semantics Members know each other directly or transitively – no unknown islands Group size can change dynamically, and be arbitrarily large or small Event generation: Anyone to all (n2n) Only one to all (one2n) A building block towards n2n (m2n) a fault-tolerant building block toward n2n

4. Tapas’034 Example A member can be An IP multicast router A core node in pub/sub context A B2B middleware process Member ‘knows’ relation

5. Tapas’035 Example A member can be An IP multicast router Attached to multicasting nodes A core node in pub/sub context Attached to publishers/ subscribers A B2B middleware process Supporting business processes Member End user

6. Tapas’036 QoS Adaptive QoS 1-to-1 communication + QoS guarantees 1-to-n communication + QoS Requirements QoS Adaptive GC 1-to-n communication + QoS guarantees QoS Exception Signal

7. Tapas’037 Reliable 1-to-n communication service If the sender does not crash, the sent message (or packet) is received by all functioning recipients If a functioning recipient receives a message, all functioning recipients receive it as well

8. Tapas’038 QoS Adaptive Reliable 1-to-n communication service If the sender does not crash, the sent message (or packet) is received by all functioning recipients Eventually with probability R, and within (delay) D with probability P D If a functioning recipient receives a message, all functioning recipients receive it as well Within (skew) S time after its reception with probability P S

9. Tapas’039 Building Block: Regular Groups Building a QoS adaptive GC is difficult So, divide and manage The unit of division is a regular group: Any two members of a regular group know each other directly Decompose the group into overlapping regular groups Rationale: Standard way to achieve scalability Tree formation algorithms

10. Tapas’0310 Group  collection of Regular groups Member ‘knows’ relation Member New ‘knows’ relation

11. Tapas’0311 Tree Formation  Identification of Regular groups An RG = A node + Immediate Children A Tree Tree Formation Algorithm

12. Tapas’0312 Reliable Multicast: First Step Processes can crash, crashed ones can recover Processes that never crash, are correct Messages between functioning processes: Can be lost with probability q If not lost, can suffer a delay whose probability distribution is known. A Regular group of pi, 0  i  n. p0 pi

13. Tapas’0313 Reliable Multicast Properties A message m (or a packet) sent by a correct process is delivered to all correct processes If a correct process receives m all correct ones receive as well

14. Tapas’0314 QoS Adaptive Reliable Multicast Properties A message m (or a packet) sent by a correct process is delivered to all correct processes Eventually with probability R, and within (delay) D with probability P D If a correct process receives m all correct ones receive as well Within (skew) S time after its reception with probability P S

15. Tapas’0315 The Protocol Sender Transmission followed by  retransmissions each with a gap of  : R is determined by group- size (n),  and q. P D is by the above,  and 1- 2-1 delay distribution. Receiver Expect the next retX, if any, within (  +  ) time If not arrived, act on behalf of the sender Given S, P S is additionally influenced by . The larger , fewer misjudgements – also other optimisation  tXretX

16. Tapas’0316 Approximate Evaluations with exponentially distributed delays D= 6 P D = 0.3975162002 P D,maj = 0.9924117041 D= 8 P D = 0.7938575743 P D,maj = 0.9999532124 D= 10 P D = 0.9437399031 P D,maj = 0.9999997923 D= 12 P D = 0.9850842340 P D,maj = 0.9999999990 D= 15 P D = 0.9984038496 P D,maj = 1.0000000000 with uniformly distributed delays: D= 6 P D = 0.1968759884 P D,maj = 0.9533320504 D= 8 P D = 0.7383674355 P D,maj = 0.9998670874 D= 10 P D = 0.9603616422 P D,maj = 0.9999999500 D= 12 P D = 0.9945734875 P D,maj = 1.0000000000 D= 15 P D = 0.9994561670 P D,maj = 1.0000000000

17. Tapas’0317 Work Plan Effectiveness of approximate evaluation against simulations API and Implementation Extracting q and 1-2-1 delay distribution from SLAs? Generalisation to groups