1. In VINI Veritas Realistic and Controlled Network Experimentation Andy Bavier Nick Feamster* Mark Huang Larry Peterson Jennifer Rexford Princeton University *Georgia Tech

2. How to Validate an Idea? Fixed, shared among many experiments Runs real routing software Exposes realistic network conditions Gives control over network events Carries traffic on behalf of real users Simulation Emulation Small-scale experiment Live deployment VINI

3. Scientific Value The most exciting phrase to hear in science, the one that heralds new discoveries, is not ‘Eureka!’ (I found it!) but ‘That’s funny …’ -- Isaac Asimov Move off the emulator, into the wild  Opportunity for more ‘that’s funny’ moments Avoid “Fallacy of Misplaced Concreteness”  Simulation and emulation are important tools  Modeling abstracts general properties from reality Philosophy: the devil may be in the details…  But insights and soundness are found there too

4. “Controlled Realism” Start with a controlled experiment Relax constraints, study effects Result: an operational virtual network that’s  Feasible  Valuable  Robust  Scalable, etc. Topology Actual network Arbitrary, emulated Traffic Real clients, servers Synthetic or traces Traffic Real clients, servers Synthetic or traces Network Events Observed in operational network Inject faults, anomalies

5. Overview VINI requirements  Fixed, shared infrastructure  Flexible network topology  Expose/inject network events  External connectivity and routing adjacencies Strategy for building VINI PL-VINI: prototype on PlanetLab Experimental results Timeline

6. Fixed Infrastructure Deploying VINI nodes in National LambdaRail, Abilene with Gigabit links

7. Shared Infrastructure Experiments given illusion of dedicated h/w

8. Flexible Topology VINI supports arbitrary virtual topologies

9. Network Events VINI exposes, can inject network failures

10. External Connectivity s c Experiments can carry traffic for real end-users

11. External Routing Adjacencies s c BGP Experiments can participate in Internet routing

12. PlanetLab  VINI Build VINI from PlanetLab, a global testbed for distributed services  Begun in 2002  700 nodes at 336 sites in 35 countries  600 projects and 2500 researchers  Serves 3-4 TB/day to ~1M clients MyPLC: PlanetLab software distribution  Anyone can run their own private PlanetLab

13. PlanetLab Experiments Simultaneous experiments in separate VMs  Each has “root” in its own VM, can customize Reserve CPU, network capacity per experiment Virtual Machine Monitor (VMM) (Linux++) Node Mgr Local Admin VM 1 VM 2 VM n … PlanetLab node

14. PL-VINI: Prototype on PlanetLab Feasible?  prototype on public PlanetLab Enable experiment: Internet In A Slice  XORP open-source routing protocol suite (NSDI ’05)  Click modular router (TOCS ’00, SOSP ’99) Clarify issues that a VINI must address  Unmodified routing software on a virtual topology  Forwarding packets at line speed  Illusion of dedicated hardware  Injection of faults and other events

15. XORP: Control Plane Goal: real routing protocols on virtual network topologies BGP, OSPF, RIP, PIM-SM, IGMP/MLD XORP can run in a PlanetLab VM XORP (routing protocols) PlanetLab VM

16. User-Mode Linux: Environment Interface ≈ network PlanetLab limitation:  Experiments cannot create new interfaces Run routing software in UML environment Create virtual network interfaces in UML XORP (routing protocols) UML eth1eth3eth2eth0 PlanetLab VM

17. Click: Data Plane Performance  Avoid UML overhead  Move to kernel, FPGA Interfaces  tunnels  Click UDP tunnels correspond to UML network interfaces Filters  “Fail a link” by blocking packets at tunnel XORP (routing protocols) UML eth1eth3eth2eth0 Click Packet Forward Engine Control Data UmlSwitch element Tunnel table Filters PlanetLab VM

18. Resource Isolation Issue: Forwarding packets in user space  PlanetLab sees heavy use  CPU load affects virtual network performance PropertyDepends OnSolution ThroughputCPU% receivedPlanetLab provides CPU reservations LatencyCPU scheduling delay PL-VINI: boost priority of packet forward process

19. Intra-domain Route Changes s c 1176 587 846 260 700 639 1295 2095 902 548 233 1893 366 Watch OSPF route convergence on Abilene 856

20. Ping During Link Failure Link downLink up Routes converging Abilene RTT: 73ms

21. TCP Throughput Zoom in Link downLink up

22. Arriving TCP Packets Slow start Retransmit lost packet PL-VINI enables a user-space virtual network to behave like a real network on PlanetLab

23. Attracting Real Users Could have run experiments on Emulab Goal: Operate our own virtual network  Carrying traffic for actual users  We can tinker with routing protocols We expect that:  PlanetLab services will subscribe to VINI network architectures to access Gb/s  Experiments will advertise routes via BGP

24. Timeline PL-VINI PlanetLab Resource resv CPU priority You are here NLR-VINI Abilene-VINI Japan-VINI PCs VINI OS MyVINI Xen Exchange traffic with ISPs 2007 NLR-VINI Abilene-VINI PCs PlanetLab OS MyPLC Gigabit layer 2 eBGP uplinks to friendly ISPs Fall 20062008 NLR-VINI Abilene-VINI Japan-VINI ???-VINI Other GREN PC + FPGAs, NPs Create layer 2 “on the fly” Other features?

25. The End Questions?

26. The End URL: http://www.vini-veritas.net Questions?

27. Backup slides

28. Conclusion VINI = evolution of PlanetLab Installing VINI nodes in NLR, Abilene Download and run Internet In A Slice MyPLC  MyVINI as code diverges  Build, run, modify your own VINI  We expect there to be many VINIs http://www.vini-veritas.net

29. Timeline Conclude with a timeline instead? Like the one for Gibson. Experiments on the top, infrastructure on the bottom, “You are here.” Today: IIAS, PL-VINI Next: RCP, VINI-NLR What other experiments?

30. Ongoing Work Improving realism  Exposing network failures and changes in the underlying topology  Participating in routing with neighboring networks Improving control  Better isolation  Experiment specification

31. Performance is bad User-space Click: ~200Mb/s forwarding Can do a lot with 200Mb/s  20 experiments can have dedicated 10Mb/s nationwide networks Improving performance is ongoing work  Allow experiments to load custom Click modules into the VINI kernel

32. PL-VINI Summary Flexible Network Topology Virtual point-to-point connectivityTunnels in Click Unique interfaces per experimentVirtual network devices in UML Exposure of topology changesUpcalls of layer-3 alarms Flexible Routing and Forwarding Per-node forwarding tableSeparate Click per virtual node Per-node routing processSeparate XORP per virtual node Connectivity to External Hosts End-hosts can direct traffic through VINIConnect to OpenVPN server Return traffic flows through VININAT in Click on egress node Support for Simultaneous Experiments Isolation between experimentsPlanetLab VMs and network isolation CPU reservations and priorities Distinct external routing adjacenciesBGP multiplexer for external sessions

33. PL-VINI / IIAS Router XORP: control plane UML: environment  Virtual interfaces Click: data plane  Performance Avoid UML overhead Move to kernel, FPGA  Interfaces  tunnels  “Fail a link” XORP (routing protocols) UML eth1eth3eth2eth0 Click Packet Forward Engine Control Data UmlSwitch element Tunnel table

34. What’s New with VINI? Integration of routing w/Internet Better isolation Real topologies Inject events

35. “Controlled Realism” Control:  Reproduce results  Methodically change or relax constraints Realism:  Long-running services attract real “customers”  Forward high traffic volumes (Gb/s)  Robustly handle unexpected events Topology Actual network Arbitrary, emulated Traffic Real clients, servers Synthetic or traces Traffic Real clients, servers Synthetic or traces Network Events Observed in operational network Inject faults, anomalies