SAAR: A Shared Control Plane for Overlay Multicast

Transcription

1 SAAR: A Shared Control Plane for Overlay Multicast Animesh Nandi Aditya Gangam,, Peter Druschel,, T.S. Eugene Ng, Ion Stoica, Hui Zhang, Bobby Bhattacharjee (MPI for Software Systems, Rice University, CMU, UC Berkeley, University of Maryland)

2 Overlay Multicast Endsystems cooperatively disseminate content Internet Animesh Nandi 2

3 Applications Webcast special events e.g. ESM, TMesh P2P-Internet Television e.g. Coolstreaming, PPLive, Sopcast Animesh Nandi 3

4 Why another Overlay Multicast paper? Architectural insight: - separate overlay, optimized for control - shared among many data overlays - benefits single-tree, multi-tree, mesh Animesh Nandi 4

5 Data dissemination T1 U.S T1 OC3 Multicast source OC3 T1 T1 Asia Overlay structure optimized for efficient data dissemination Animesh Nandi 5

6 Control Mechanism Employs control mechanisms to build/repair overlay structure A Animesh Nandi 6

7 ESM s Data/Control Data overlay optimized for latency/bandwidth Gossip to distribute membership info Probing to select parents A {B,C} D Probe Probe C {D} {C,E} B E {D,B} Does not scale to large groups and high membership churn Summary: efficient data plane structure but unscalable control Animesh Nandi 7

8 Idea: Decouple control/data into separate overlays Dataplane optimized for data dissemination Separate overlay provides efficient control Summary: Using separate overlays avoids efficiency tradeoffs Animesh Nandi 8

9 Sharing control overlay Shared overlay provides efficient control BBC CNN Channel switching does not affect control overlay membership Summary: Sharing reduces control overlay churn. Animesh Nandi 9

10 Neighbor Acquisition single-tree multi-tree mesh A A A Summary: Neighbor acquisition is fundamental control task Animesh Nandi 10

11 Anycast Primitive for Neighbor Acquisition anycast (groupid g,, constraint p,, objective function o) (Amongst the members of g satisfying p,, chooses the one that maximizes o) Constraint Objective function Summary: Anycast primitive used to build different overlays Animesh Nandi 11

12 Question: Can the control overlay efficiently implement this anycast primitive? Animesh Nandi 12

13 Idea: Use structured overlay Pastry key space Data planes optimized for efficient data dissemination Shared structured overlay provides efficient anycast service Animesh Nandi 13

14 SAAR: Per-channel control tree Root (Control-tree) channel member member of other channel Dataplane View Control plane view Control tree s structure is optimized for control and independent of data overlay structure. Animesh Nandi 14

15 Control tree formation Root (GID) channel member member of other channel Tree structure formed efficiently Animesh Nandi 15

16 SAAR: Aggregation Root (Control-tree) cap = 5 cap = 0 cap = 2 3 State variable cap := #spare capacity slots 0 cap = Control plane view channel member member of other channel Animesh Nandi 16

17 SAAR: Aggregation Root (Control-tree) min dep = 1 min dep = 4 min dep = 1 min dep = State variable dep := advertised tree depths 1 Control plane view channel member member of other channel Animesh Nandi 17

18 SAAR: Efficient search Anycast request cap = 5 min dep = 1 cap = 2 min dep = 1 - spare capacity - no spare capacity cap = 0 min dep = (cap=3 dep=4) (cap=2 dep=1) Aggregation enables efficient Depth-First-Search of control tree Animesh Nandi 18

19 Evaluation Modelnet 350 nodes heterogeneous forwarding bandwidths mean stay time of 5 minutes (exponential distribution) Metrics: channel switching delay streaming quality control overhead Animesh Nandi 19

20 SAAR enables low join delays ESM Animesh Nandi 20

21 SAAR enables low join delays ESM Initial single-tree join when not in SAAR control overlay Animesh Nandi 21

22 SAAR enables low join delays Multitree Singletree ESM Initial single-tree join when not in SAAR control overlay Animesh Nandi 22

23 SAAR enables low join delays Multitree Singletree Mesh ESM Initial single-tree join when not in SAAR control overlay Animesh Nandi 23

24 SAAR: good streaming quality ESM Mesh Singletree Multitree Animesh Nandi 24

25 SAAR: low control overhead Singletree Mesh ESM Multitree Animesh Nandi 25

26 SAAR supports complex overlays Multi-tree, 5 stripes, 350 nodes, 2-min mean churn (exp.), RI=1.23 Animesh Nandi 26

27 More results in the paper Flash crowd conditions Higher churn (2 min mean, exp. dist) Control churn Planetlab experiments Animesh Nandi 27

28 Conclusions Control efficiency is key to performance of overlay multicast systems, especially tree-based systems. Anycast based on structured overlays enables powerful, efficient control for tree-based systems. Fast channel switching can be achieved with tree- based overlay multicast systems. THANK YOU Animesh Nandi 28

29 Backup Slides Animesh Nandi 29

30 Planetlab: : Join delays Animesh Nandi 30

31 Planetlab: : Streaming Quality Animesh Nandi 31

32 Flash crowd: Join Delays Animesh Nandi 32

33 Flash crowd: Tree Depths Animesh Nandi 33

34 eg.. Scribe s s Control/Data Pastry s Key space 128 Prefix-based Routing path 924 Root(GId=999) Low bandwidth subscribers Match prefix with GId Scribe has efficient control but inefficient data paths Animesh Nandi 34

35 SAAR Anycast: : Interior-node-disjoint Anycast requestor 617 control trees per CEM 302 Root(Tree 1 gid=333) spare capacity - overloaded Root(Tree 2 gid=888) 597 Summary: SAAR provides uninterrupted control in the face of control overlay churn. Animesh Nandi 35

36 Enabling multi-tree data plane Stripe Anycast group Data plane Control plane Attempts to build k interior-node disjoint stripe trees. Node anycasts to establish parents in each stripe, but joins only one anycast group corresponding to its primary stripe. Resource based primary stripe selection: groupaggregaterequest (groupid, state-variable) Same predicate and objective function as in single-tree for each stripe tree anycasts Summary : multi-tree policy expressed easily. Animesh Nandi 36

37 Enabling block-based data plane Oudegree Neighbors - request blocks From A A with D= 2 Data Plane buffermap Indegree neighbors - serve blocks to A Anycast group Control Plane Node anycasts for indegree-neighbors and employs swarming Group s s state variables g capacity, g load, g loss,, g buffermap, g coordinate Predicate: : (load < capacity) & // not overloaded (loss < threshold) // low loss coordinate, Objective function: cardinality(missingset Π g buffermap ) // recover missing blocks 1/g // establish a good mesh neighbor 1/g loss Summary: block-based policy expressed easily Animesh Nandi 37

38 Graphs from paper Animesh Nandi 38

39 Planetlab: : Join delays Animesh Nandi 39

40 Planetlab: : Streaming Quality Animesh Nandi 40

41 Multi-tree: Resource Balance Animesh Nandi 41

42 Multi-tree: #stripes recvd Animesh Nandi 42

43 Locality-awareness Animesh Nandi 43

44 Load-awareness Animesh Nandi 44

45 Depth-Optimization Animesh Nandi 45

46 Single-tree: Join Delays (milliseconds) Animesh Nandi 46

47 Single-tree: Continuity Animesh Nandi 47

48 Single-tree: tree-depths Animesh Nandi 48

49 Single-tree: Nodestress Animesh Nandi 49

50 Multi-tree: Join delays Animesh Nandi 50

51 Multi-tree: Streaming Quality Animesh Nandi 51

52 Mesh-based: Streaming Quality Animesh Nandi 52

53 Mesh-based: Path lengths Animesh Nandi 53