A Distributed and Robust SDN Control Plane for Transactional Network Updates

Transcription

1 A Distributed and Robust SDN Control Plane for Transactional Network Updates Marco Canini (UCL) with Petr Kuznetsov (Télécom ParisTech), Dan Levin (TU Berlin), Stefan Schmid (TU Berlin & T-Labs) 1

2 Network Policy Specification Routing Load Monitoring Access Waypoint Balancing Control Centralized Network Policy Controller Platform 2

3 Network Policy Specification forward from X to Y Routing Load Monitoring Access Waypoint Balancing Control Centralized Network Policy Controller Platform X 1.<Match, Action> 2.<Match, Action> 3.<Match, Action> 4.<Match, Action> 5.<Match, Action> Y 3

4 Network Policy Specification X Routing Load Monitoring Access Waypoint Balancing Control Centralized Network Policy Composition Controller Platform Y Policy composition assembles data plane updates as a semantically sound set of rules 4

5 Policy Composition Review Foster 11, Monsanto 13: Modular, parallel and sequential composition Srv Load Balancing >> Routing U Monitoring Ferguson 13: Policy trees for multi-authorship Composition Routing Waypoint Reitblatt 12: Consistent network updates 5

6 Conflicting Policies dst = H fwd(x) Routing Load Balancing Conflict X!= Y In the general case, policies might conflict Examples: Overlapping domains and same precedence Scarce flowtable resources Goal: Must avoid conflicting policies! dst = H fwd(y) 6

7 Centralized Network Control? Routing Load Monitoring Access Waypoint Balancing Control Centralized Network Policy Controller Platform Fully centralized Inadequate availability, scalability and responsiveness 7

8 Distributed Network Control Routing Load Monitoring Access Waypoint Balancing Control Centralized Network Policy Controller Platform Composition? Consistency and concurrency Faults and asynchronous communication 8

9 Now, consider policy composition in the distributed control plane... Routing Monitoring Waypoint Policy Controller Platform Routing Monitoring Waypoint Policy Controller Platform Switch Reader- Writer Model State Distribution Model Control Logic Factorization 9

10 Why should the programmer care? We believe the programmer should not! Enter Software Transactional Networking Let a dedicated component implement a general solution to all hard-to-solve, low-level concurrency and fault tolerance issues to policy composition apply(π) ack / nack(reason) Transactional Interface 10

11 Consistency: Linearizability of updates apply(π 1 ) ack apply(π 1 ) ack p 1 p 1 apply(π 2 ) ack apply(π 2 ) ack p 2 apply(π 3 ) nack p 2 p 3 p 3 sw 1 sw 1 sw 2 sw 2 sw 3 Original history We don t control traffic! 11 sw 3 Manipulate the network as though there is no concurrency Sequential equivalent

12 Software Transactional Networking: Consistent Policy Composition (CPC) Routing Monitoring Waypoint apply(π 1 ) ack apply(π 2 ) nack(reason) CPC Interface 1. All-or-nothing semantics 2. Tolerate up to f controller node crash failures 3. Non conflicting policies eventually installed and at least one policy commits (among conflicting ones) 4. Ensure policy updates affect traffic as a sequential composition of their policies 12

13 Conceptualizing CPC Routing Monitoring Waypoint Reliable but asynchronous channel CPC Node CPC Node Every controller node receives and participate in installing every policy update 13

14 Conceptualizing CPC Routing Monitoring Waypoint apply(π) ack CPC Node CPC Node Policy π unique tag τ 1. Internal ports match on tag τ 2. Ingress ports apply tag τ 14

15 Conceptualizing CPC Routing Monitoring Waypoint CPC Node CPC Node RMW RMW RMW RMW RMW Atomic read-modify-write primitive RMW RMW 15

16 RMW model Controllers access ports with atomic read-modify-write primitive RMW(g,v): read current state v apply and return g(v,v ) RMW(g,v) Controller g(v,v ) Intuition: do not update if policy update conflicts with currently installed policy In the paper: Theorem: 1-resilient read-write CPC is impossible 16

17 FixTag: upper bound algorithm Operation: 1. Unique tag per path 2. Broadcast policy π to all other controllers 3. Update ingress ports in predefined order 4. add rule to tag all packets matching dom(π) with the tag corresponding to the path π (i) for ingress port i Upsides: wait-free (tolerates all failure patterns) Controllers only synchronize through the data plane Downsides: tag complexity linear in # possible policies and paths May grow super-exponential in the size of the network 17

18 Can we lower the tax complexity? No, if we get no feedback from the network Tag τ cannot be reused if a packet tagged with τ is still in flight Suppose, we can correctly evaluate the set of active tags Correct (but asynchronous) oracle Single-controller scenario: one bit is enough! Upon policy update π i, wait until (i mod 2)-traffic is over, and use tag i mod 2 Two or more controllers: inherent price of concurrency? Between constant and super-exponential? Yes, if controllers coordinate the use of tags 18

19 ReuseTag: linear complexity Proportional to the level of resilience: Up to f failures: f+2 tags needed (proved optimal) Controllers use replicated state machine that imposes a total order on the policy updates and ensures coordinated use and reuse of tags All requests are serialized, even non-conflicting ones 19

20 Summary Software Transactional Networking framework for consistent policy composition (CPC) in distributed SDN control planes Transactional interface to manipulate the network as though there is no concurrency Policies compose or conflict (and abort) Formal model of the problem is in the paper Two CPC algorithms FixTag ReuseTag: f+2 tags (minimal number) 20

21 Backup 21

22 Acknowledgements Petr Kuznetsov Dan Levin Stefan Schmid Supported by ARC grant 13/ from Communauté française de Belgique, and European Union s Horizon 2020 ENDEAVOUR project (grant agreement ) 22