Networking Seminar Stanford University. Madan Jampani 3/12/2015

Transcription

1 Networking Seminar Stanford University Madan Jampani 3/12/2015

2 Can SDN control plane scale without sacrificing abstractions and performance?

3 Control Plane Data Plane

4 Simple yet powerful abstraction Global Network View Scaling strong consistency Managing Complexity

5 What is the simplest controller? Observe Program

6 Abstractions Single Instance Performance Correctness Availability Scale

7 How to improve availability?

8 How to improve availability? Primary Standby

9 How to improve availability? Standby Primary

10 Abstractions Single Instance Primary/ Standby Performance Correctness Availability Scale

11 Why you might need to scale? Primary Standby

12 Why you might need to scale? Primary Standby

13 Why you might need to scale? Primary Standby

14 Why you might need to scale? Primary Standby

15 Fully Distributed Control Plane

16

17

18

19

20

21

22 Abstractions Single Instance Primary/ Standby MultiInstance Performance Correctness Availability Scale

23

24

25

26 Abstractions Single Instance Primary/ Standby Multi Instance Performance Correctness Availability Scale

27 Tell me about your slice?

28 Abstractions Single Instance Primary/ Standby Multi Instance Performance Correctness Availability Scale

29 Tell me about your slice? Cache

30 Abstractions Single Instance Primary/ Standby Multi Instance Performance Correctness Availability Scale

31 Can we build a complete solution? Let s look at prior art...

32 Write Topology State Topology Events Distributed Topology Store

33 Read Topology State Distributed Topology Store

34 Read Topology State Cache Distributed Topology Store

35 Read Topology State Cache Distributed Topology Store

36 Can we build a solution that meets all criteria?

37 Topology as a State Machine Current Topology apply event Events are Switch/Port/Link up/down Updated Topology

38

39

40 E

41 E E E

42 Pros Simple Fast Cons Dropped messages Reordered messages

43 E

44 F

45 F E

46 We have a single writer problem

47

48

49 Switch Mastership Terms C1 C2 C We track this in a strongly consistent store

50 Switch Event Numbers C1 C2 C

51 Partial Ordering of Topology Events Each event has a unique logical timestamp (Switch ID, Term Number, Event Number)

52 E (S1, 1, 2)

53 E (S1, 1, 2) E (S1, 1, 2) (S1, 1, 2) E

54 F (S1, 2, 1)

55 F (S1, 2, 1) F (S1, 2, 1)

56 F (S1, 2, 1) F (S1, 2, 1) E (S1, 1, 2)

57 F (S1, 2, 1) F (S1, 2, 1) E (S1, 1, 2)

58 To summarize Each instance has a full copy of network topology Events are timestamped on arrival and broadcasted Stale events are dropped on receipt

59 There is one additional step... What about dropped messages?

60 Did you hear about the port that went offline?

61

62 Anti-Entropy Lightweight, Gossip style, peer-to-peer Quickly bootstraps newly joined nodes

63 Abstractions Single Instance Primary/ Standby Multi Instance Performance Correctness Availability Scale

64 A model for state tracking If you are the observer, Eventual Consistency is the best option View should be consistent with the network, not with other views

65 Hiding the complexity EventuallyConsistentMap<K, V, T> Plugin your own timestamp generation

66 What about other Control Plane state... Switch to Controller mapping Resource Allocations Flows Various application generated state

67 What about other Control Plane state... Switch to Controller mapping Resource Allocations Flows Various application generated state In all case strong consistency is either required or highly desirable

68 Consistency => Coordination 2PC All participants need to be available Consensus Only a majority need to participate

69 State Consistency through Consensus LOG LOG LOG

70 State Consistency through Consensus LOG LOG LOG LOG LOG LOG LOG

71 Scaling Consistency

72 Scaling Consistency

73 Scaling Consistency

74 Consistency Cost Atomic updates within a shard are cheap Atomic updates spanning shards use 2PC

75 Hiding Complexity ConsistentMap<K, V>

76 To Conclude SDN at scale is possible if we can take care of state management Abstractions are important

77 Thank you