Networked Systems Boon Thau Loo University of Pennsylvania NSF ExCAPE Meeting 20 Aug 2013
Outline Summary (activities over past year) Research highlights Conclusion
Year 1 in Retrospect Original proposal focus: apply synthesis techniques in (1) Internet routing and (2) wireless network protocols Scope of work has expanded over the past year: From synthesizing protocol implementations to synthesizing policies and configurations Exciting new applications in Cloud computing and Softwaredefined Networking Common themes: Deal with management complexity in cloud and highly volatile networked environments Configurations need to be modified at runtime given changes in topologies, traffic flows, etc. Meeting optimization goals (e.g. SLAs) while not violating security or ISP business constraints
Main Participants Faculty: Rajeev Alur (Penn), Ras Bodik (UC Berkeley), Boon Thau Loo (Penn), George Pappas (Penn), Ufuk Topcu (Penn), Madhu Parthasarathy (UIUC) Postdoc: Alexander Gurney (Penn) Students: Behnaz Arzani (Penn), Shaon Barman (UC Berkeley), Sarah Chasins (UC Berkeley), Chen Chen (Penn), Salar Moarref (Penn), Anduo Wang (Penn -> UIUC), Shambwaditya Saha (UIUC), Yifei Yuan (Penn). Mode of communication/collaboration: Weekly small group meetings, larger group sync-ups, monthly webinar, Skype (Penn/India), qualifying exams (Yifei Yuan, Alur/Madhu/Loo), papers, demos.
Outline Summary (activities over past year) Three Research highlights Conclusion
#1: Routing Constraints and Traffic Disruptions Participants: Alex Gurney, Behnaz Arzani, Roch Guerin, Boon T. Loo Last year: we presented the Formally Safe Routing (FSR) toolkit -> automatically synthesize Internet routing policies that lead to convergence (i.e. safe) This year: two additional dimensions Traffic engineering, Incremental network repair SIGCOMM 12 demonstration and CoNEXT 13 submission Alex Gurney s 5 minute lunch talk
Routing Policy Routing policy often ends up being characterized in terms of preferences over paths path p should be better than path q I don t care to choose between p and q Universal language for supporting various business cases, operational concerns, and so on. Example abstraction: routing algebra (in original proposal), declarative networking programs
Traffic Engineering We have a network, represented as a (directed, weighted) graph. Routing will establish all-pairs shortest paths over this graph. We have a demand matrix of traffic volumes for each source and destination. The shortest-paths flow yields link loads (a traffic matrix). Our job: choose weights to minimize load.
Challenges and Initial Approaches Routing protocol must converge Minimize congestion network-wide These goals are potentially conflicting: Some choices of link weights (minimizing congestion) lead to convergence failure. How can we ensure convergence, while ensuring we get as close as possible to an optimal TE solution? Interactive repair: if convergence invariant fails (due to operator error or router failure), how to find a shortest sequence of repairs to minimize traffic impact? Is shortest sequence of repairs always better? Techniques we explore: Max-SMT formulation (Yices), embed path preferences inside existing TE optimization algorithms Evaluated on real network topologies on Emulab
#2: A Synthesis Approach Towards Automated Management of SDNs Participants: Anduo Wang, Solar Moarref, Ufuk Topcu, Boon Thau Loo, Andre Scedrov Software-defined networking: Centralized controller (programmable in software) Switches can be dynamically programmed: Flow table entries <pattern, action> Used in network virtualization in data centers Traffic isolation, dynamic load balancing OpenFlow standards. Gaining significant traction in industry WRiPE 13 workshop paper
Challenges and Initial Approaches Given SDN configuration 1, migrate to configuration 2: Determine sequence of routers to update Need to avoid routing loops, and preserve correctness invariants (e.g. traffic isolation, deny certain classes of traffic) Solve as a reachability problem in model checker nusmv Output: an ordering of rule updates Given existing SDN configuration 1, update configurations in response to network state changes Solve as two-player temporal logic game Control logic= routing path rule (environment player)+ access control rule (system player) Find a winning strategy for access-control rules against all path changes Winning strategy: invariant preserving Dealing with state explosion: Abstraction based on grouping by nodes or flows
#3: Automatic Bandwidth Allocation in Data Center s Network Data Centers X 1 X 2 X 3 1G 600M 500M 450M S 1 S 2 S 3 S 4 Yifei Yuan, Anduo Wang, Rajeev Alur, Boon Thau Loo
Automatic Bandwidth Allocation in Data Centers Data Center s Network Virtual Network X 1 V 1 400M 400M X 2 X 3 V 2 V 3 1G 600M 500M 450M S 1 S 2 S 3 S 4
Automatic Bandwidth Allocation in Data Centers Data Center s Network Virtual Network X 1 V 1 10G 10G 400M 400M X 2 X 3 V 2 V 3 1G 600M 500M 450M S 1 S 2 S 3 S 4
Automatic Bandwidth Allocation in Data Centers Data Center s Network Virtual Network X 1 V 1 400M 400M X 2 X 3 V 2 V 3 1G 600M 500M 450M v 1 v 3 v 2 S 1 S 2 S 3 S 4
Automatic Bandwidth Allocation in Data Centers Data Center s Network Virtual Network X 1 V 1 400M 400M X 2 X 3 V 2 V 3 1G 600M 500M 450M v 1 v 3 v 2 S 1 S 2 S 3 S 4
Challenges and Initial Approaches NP-complete problem Existing heuristics are inefficient but may not find a solution (even if it exists) Our approach: SAT/SMT solving Abstraction and refinement for scalability Exploit hierarchical structure of data centers (Tree, Fat-tree) FMCAD 13 paper
Outline Summary (activities over past year) Research highlights Conclusion
ExCAPE Inspired Research Use of formal methods and programming languages has been catching on in the networking community: Domain-specific languages: Frenetic, Declarative Networking Formal methods: Formally Safe Routing toolkit Cornell Summer School on Formal Methods and Networks Workshop on Rigorous Protocol Engineering (WRiPE), 2011 onwards Network management as a synthesis problem is a new frontier to be explored We are the first group to explore this concept, and this work is made possible by ExCAPE.
ExCAPE Inspired Themes Network reconfiguration as a reactive synthesis problem: Environment: Network state (e.g. topology) changes System: routing protocol and configurations Strategy: Figure out a sequence of changes to System Applicable to BGP and SDN challenge problems Distributed/autonomous/asynchronous nature of networked systems may require rethink of traditional methods Solvers: Max-SMT (Yices) for the BGP problem, and Z3 for data center Optimization problem given logical constraints Dealing with state explosion: Flor or hierarchical abstractions (BGP, SDN, Data center)
Publication/Dissertation Highlights Automated Synthesis of Reactive Controllers for Software-Defined Networks. Anduo Wang, Salar Moarref, Ufuk Topcu, Boon Thau Loo and Andre Scedrov. 3rd International Workshop on Rigorous Protocol Engineering (WRiPE), 2013. On the Feasibility of Automation for Bandwidth Allocation Problems in Data Centers. Yifei Yuan, Anduo Wang, Rajeev Alur, and Boon Thau Loo. Formal Methods in Computer-Aided Design (FMCAD), 2013. Route Shepherd: Stability Hints for the Control Plane. Alexander J.T. Gurney, Xianglong Han, Yang Li, and Boon Thau Loo. ACM SIGCOMM Conference on Data Communication (demonstration), Helsinki, Finland, Aug, 2012.
Dissertation/Qualifying-Exams Automated Formal Analysis of Internet Routing Configurations. Anduo Wang University of Pennsylvania Ph.D. dissertation, 2013. (Advisor: Boon Thau Loo, committee member: Rajeev Alur) Disruption-free Network Migration Yifei Yuan UPenn s PhD qualifying exam, 2013. (Advisor: Rajeev Alur, committee chair: Boon Thau Loo, committee member: Madhu Parthasarathy
Plans for Year 2 Continue our current research directions Many projects in early stages Many papers and tool development in the pipeline Evolve into Ph.D. dissertation topics. Co-advising would be a plus. Cross-layer synthesis (routing, data center networking, cloud scheduling, web scripting)