virtualization
play

<virtualization> Feedback-directed emulation IP address - PDF document

Emulab: Network Testbed Large-scale Virtualization in the Emulab Network Testbed Mike Hibler, Robert Ricci, Leigh Stoller, Jonathon Duerig, Shashi Guruprasad, Tim Stack, Kirk Webb, Jay Lepreau 2 The Basic Idea: Whats Wrong? Use


  1. Emulab: Network Testbed Large-scale Virtualization in the Emulab Network Testbed Mike Hibler, Robert Ricci, Leigh Stoller, Jonathon Duerig, Shashi Guruprasad, Tim Stack, Kirk Webb, Jay Lepreau 2 The Basic Idea: What’s Wrong? Use virtualization to Too small perform network experiments using fewer physical resources. Inefficient ... and do this in a way that: Solution? is transparent to applications Virtualize and Multiplex preserves experiment fidelity 3 Not Just Picking a VM Technology Challenges Opportunities • Fidelity • Closed world • Can re-run • Preserve network experiments Complete Virtual Network topology Experimentation System 5 6

  2. Full System • Virtualization technology • Host and network • Resource mapping <virtualization> • Feedback-directed emulation • IP address assignment • Scalable control system • Routing table calculation 7 Start: FreeBSD jail • Namespace isolation • Virtual disks • We added network virtualization: • Ability to bind to multiple interfaces • New virtual network device ( veth ) • Separate routing tables 9 10 What does it mean to <mapping> make a good mapping?

  3. Good Mapping assign • Pack well • Solves an NP-hard problem • Use resources efficiently • Pack both nodes and links • Specifying packing criteria • Avoid scarce resources • Do it quickly • Paper: [Ricci+:CCR03] • Critical path for creating an • Based on simulated annealing experiment • We extended for virtual nodes 13 14 Resource-Based Packing • Use quantities we can directly measure • Resource-based system Assigning Quickly “This virtual node uses 100 MHz of CPU” “This physical node has 3 GHz of CPU” • Works well for heterogenous virtual and physical nodes 15 Small Topologies Virtual Topologies 17 18

  4. Prepass Scaling With Prepass 200 19 20 Mapping Quality <feedback> 21 How do I know how tightly I can pack my Closed, repeatable virtual nodes? world I don’t!

  5. The Plan Picking Initial Packing • Pick a packing • Pick a packing • Start one-to-one • Run experiment • Run experiment • Possibly with a subset of topology • Monitor for artifacts • Monitor for artifacts • Start tightly packed • If artifacts found: • If artifacts found: • Optimistically assume low usage • Re-pack • Re-pack • Repeat • Repeat 25 26 Monitoring for Artifacts Re-Packing • CPU near 100% • Measure resource use • Significant paging activity • Feed into resource-based packing • Disk utilization 27 28 Feedback in a Nutshell • Rely on packing, not isolation • Discover packing factors empirically <numbers> • Re-use between experiments 29

  6. kindex: Packing Factors Feedback Case Study Transactions Response Round Per Second Time (s) Bootstrap: 74 physical 2.29 0.43 Round 1: 7 physical 1.85 0.53 Round 2: 7 physical 2.29 0.43 Deployed Use Conclusion • Creation time: 7 minutes for 100 nodes • Virtualization increases Emulab’s capacity • 5,125 experiments • Transparently • 296,621 virtual nodes • Preserves fidelity • 32% of Emulab nodes virtual • Requires solving several challenging problems • 5.75 average packing factor • Proven useful in production • Average: 58 nodes, max: 1,000 nodes www.emulab.net 33 34 Related Work • Virtual Network Emulation • ModelNet, DieCast • Virtual Machines • Xen, VMWare, vservers, OpenVZ, NET <end/> • Network Virtualization • NetNS, OpenVZ, Trellis, IMUNES • Feedback-based Mapping • Hippodrome 36

  7. Minimal Effective ModelNet Virtualization • Applications can only run on edge nodes: • Application transparency single-homed only • More basic virtualization • Application fidelity • No artifact detection • System capacity • No feedback system • Emulab has richer control framework • Scales to much larger interior networks 37 38 Application Transparency Application Fidelity • Physical results � Virtual • Real applications • Simulation results • Virtual machines • Virtual node interference • Keep most semantics of unshared machines • Perfect resource isolation • Simple processes • Detect artifacts and re-run • Preserve experimenter’s topology • Full network virtualization 39 40 System Capacity • Low overhead • In-kernel (vservers, jails) • Hypervisor (VMWare, Xen) • Don’t prolong experiments • DieCast 41

Recommend


More recommend