1. Snowbird: Interactive Resource-Intensive Applications Made Easy H. Andrés Lagar-Cavilla * Niraj Tolia † ‡, Eyal de Lara *, M. Satyanarayanan ‡ & Dave O’Hallaron ‡ * University of Toronto, † HP Labs, ‡ Carnegie Mellon University Middleware, November 2007

2. Bimodal Applications Interactive –Cognitive phase Resource-Intensive –Crunch phase Digital Animation Scientific Computing Engineering Design Bio/Pharma Video Editing …. Maya (digital animation)

3. Dichotomy Crunch (computation) – Short completion time – Remote computing resources Cognitive (interaction) – Crisp interactive performance – User attention

4. Execution Alternatives –Thick Clients (Desktop PCs) Cognitive: excellent interactive performance × Crunch: resource constrained –Thin Clients (VNC, Remote Desktop) Crunch: use of remote resources (CPU, Data) × Cognitive: latency and jitter impact interactions –Custom Applications Pipeline: placement over different nodes × Requires significant developer resources

5. Execution Alternatives Cognitive: excellent interactive performance Crunch: use of remote resources (CPU, Data)

6. Snowbird: Agent Abstraction Applications encapsulated within an agent –Agent: processes, libraries, IPC, OS, config data… –Migration: performance goals achieved by morphing a thin client for resource intensive crunch phases a thick client for highly interactive cognitive phases

7. Bimodal Applications Made Easy Develop apps as monolithic blocks –Don’t worry about what executes where –Agent is migrated to satisfy each phase Seamless and transparent behavior Legacy support –Different OS’s (and versions/features) –Different languages (Fortran!) –No need for recompilation, relinking, etc… –Closed source apps just work

8. Implementation

9. Design Criteria VM-based migration –x86 interface most widely deployed –Transparently support OS, lang, etc… Internet scale –WAN migration –Long fat pipes: 50 Mbp/s…, 50-200 ms RTT Use of graphics HW acceleration a must –For the cognitive mode of bimodal apps Server-less design (P2P) –All hosts symmetric –Can execute anywhere

10. Components VMM: suspend, resume, live migration –Xen 3.0.1 Interaction-aware migration manager –Transparent Support for 3D acceleration in VMs –Vital for crisp interaction Virtual disk that maximizes locality –WAN area migration

11. Agent Profiles: Migration Manager FSM that models an agent’s behavior Provided by expert users, admins, or developers –Default system-wide profile available Only deployment additional effort CPU Intensive cycles.org Net Intensive data.edu Interaction Intensive usr.home Net > 4 Mbit/s CPU > 95% FPS 15 What’s this?

12. Interaction-aware Novelty in our approach Measure the quality of the interactive response We use frames per seconds –More in sync with operations in bimodal apps –Stretch object, rotate, zoom, etc… –Pure latency not enough Ample space for future work

13. Frames Per Second Non work-conserving (VNC): FPS = 2/latency Local: FPS = n/latency Work-conserving (X): FPS = n/(latency*k)

14. The Rest VMGL: support for 3D acceleration in VMs –Coming live to this conference in 5 mins –Follow up: VEE 2007, >4K downloads WANDisk: virtual disk maximizes locality –Minimizes WAN communication –Simplifies state synchronization More details in paper

15. Evaluation

16. Benchmarks Maya (closed source)ADF (closed source) QuakeViz Kmenc15 Broad set of domains –Scientific Computing, Bio, Video Editing, Animation Closed and open source Straightforward installation Able to use generic profile on all four Partitioned mode for comparison –ADF & Maya

17. Methodology Crunch + cognitive benchmarks Performing “crunch” experiments is easy Replaying long interactive traces is not –Can’t expect a user to do it –Must be able to compare results VNC-Redux: record and replay interactive user sessions –Record input and screen state –This is matched during replay for accuracy

18. Experimental Setup Thick: No virtualization, on User Desktop –User Desktop: UP with graphics acceleration Thin: No virtualization, on Compute Server –Compute Server: 4-way SMP –100 Mbit/s WAN, RTT: 33, 66, and 100 ms Partitioned –App-specific developer-brewed: Maya & ADF Snowbird –Agents are initially launched on User Desktop –Xen+Migration Manager+VMGL+WANDisk

19. Results: Crunch Phase Snowbird’s crunch performance –Much better than thick –Comparable to thin/partitioned 370…Ouch!

20. Better interactivity than thin clients Is Snowbird any worse than a thick client? > 20 FPS is ok, < 8 FPS is unusable Results: Cognitive Phase

21. Take Home Messages Bimodal applications –What they are and why they matter Thin clients are not almighty –There is no replacement for local interaction Best of both worlds: thick and thin clients –Necessary in an Internet world with remote computing resources VM-based app migration is feasible –And with many advantages Future trends align well…

22. Futurism More bandwidth: cheaper VM migration What about latency? –The earth is not shrinking –Speed of light is not increasing –More routers, overlays, firewalls –Toronto-London UK: ~109ms –Toronto-LA: ~84ms Insurmountable obstacle for thin clients

23. Questions? Thanks H. Andrés Lagar-Cavilla andreslc@cs.toronto.edu Niraj Tolia, Eyal de Lara, Satya & Dave O’Hallaron U of Toronto, HP Labs, Carnegie Mellon

24. Backup

25. VMGL: 3D Acceleration in VMs OpenGL virtualization Hardware specs closed, unavailable Focus instead on software standards –OpenGL -> cross-platform –Direct3D -> MS-only Intercept GL calls and forward them to the host. Proprietary driver renders there. More details: VEE 2007

26. Open GL for X11 Apps

27. VMGL for X11 Apps

28. WANDisk: Virtual Disk foo.toronto.edu bar.cmu.edu Chunk Table Chunk Misses baz.europe.org

29. Why Another Storage System? Exploits Snowbird characteristics P2P model –No server interposition –Single-writer: simple metadata, no locks Minimizes WAN talk –Locality: persistent replicas –Differential transfers: rsync –On-demand fetching

30. Experimental Setup Thick: No virtualization, on User Desktop –User Desktop: UP with graphics acceleration Thin: No virtualization, on Compute Server –Compute Server: 4-way SMP –100 Mbit/s WAN, RTT: 33, 66, and 100 ms Partitioned –App-specific developer-brewed: Maya & ADF Snowbird –Agents are initially launched on User Desktop

31. ADF Migration Time (secs) Latency (ms)DetectionMigrationPause 3312.5624.9 6611.5626.2 10013.164.96.7

32. Applicability of Snowbird Morphing time: our implementation Speedup: application & resources C: Crunch phase time locally

33. Snowbird Limitations Parallelism up to SMP level –What about cluster-scale? SSE, 3DNow! –i.e. x86 is not that uniform Overlapping phases –Hysteresis, priority in migration manager Very short phases –Cost/benefit analysis

34. Our Current Interests VM support for large parallel tasks –Relevant to commodity computing –Migration, the cloud, etc… How to measure interactive performance –Thin clients, desktop consolidation, VMs –Unknown effects for modern (3D-heavy) GUIs

35. Really Backup Backup

36. Keywords Of This Presentation Thin clients –Remote execution Interactive Performance –Thick clients Virtual machine migration –Application migration Bimodal applications –What they are and why they matter

37. Frames Per Second Non work-conserving: same latency, less frames Local: FPS = frames/latency Work-conserving: same frames, more latency

38. What We Need System support for bimodal applications Combine best of both worlds –Thick client, local execution Interaction –Thin client, remote execution Computation Make development easy

39. Results: Crunch Phase Snowbird’s crunch performance –Much better than thick –Comparable to thin/partitioned

40. Talk Pointers Be more explicit demo: thin, frames Proxy FPS & input in migr manager