Self-Tuning Wireless Network Power Management

1. Self-Tuning Wireless Network Power Management Manish Anand Edmund B. Nightingale Jason Flinn Department of Electrical Engineering and Computer Science University of Michigan

2. Motivation • Wireless connectivity is vital to mobile computing • But, taxes limited battery capacity of a mobile device • Power management can extend battery lifetime • -However, it can negatively impact performance Manish Anand

3. 802.11 Network Power Management Network interface may be continuously-active (CAM) • Large power cost (~1.5 Watts) • May halve battery lifetime of a handheld Alternatively, can use power-saving mode (PSM) • If no packets at access point, client interface sleeps • Wakes up periodically (beacon every 100 ms) • Reduces network power usage 70-80% Manish Anand

4. Effect of Power Management on NFS Time to list a directory on handheld with Cisco 350 card • PSM-static: • 16-32x slower • 17x more energy • PSM-adaptive: • up to 26x slower • 12x more energy Manish Anand

5. What’s Going On? NFS issues RPCs one at a time ….. RPC requests RPC responses NFS Server Access Point Mobile Client Beacons 50ms 100ms 100ms • Each RPC delayed 100ms – cumulative delay is large • Affects apps with sequential request/response pairs • Examples: file systems, remote X, CORBA, Java RMI… Manish Anand

6. Outline • Motivation • Self Tuning Power Management • Design Principles • Implementation • Evaluation • Related Work andSummary Manish Anand

7. Know Application Intent Application: NFS File access Best Policy: Use CAM during activity period • Not enough network traffic to switch to CAM • Data rate is dependent on the power mgt. Beacon Period PSM CAM Manish Anand

8. Know Application Intent Application: Stock Ticker that is receiving 10 packets per second Best policy: Use PSM • Data rate is not dependent on power mgmt. Beacon Period PSM CAM • STPM allows applications to disclose hints about: • - When data transfer are occurring • - How much data will be transferred (optional) • - Max delay on incoming packets Manish Anand

9. Be Proactive Transition cost of changing power mode: 200-600 ms. Large transfers: use a reactive strategy - If transfer large enough, should switch to CAM - Break-even point depends on card characteristics - STPM calculates this dynamically Many applications (like NFS) only make short transfers: be proactive - Benefit of being in CAM small for each transfer - But if many transfers, can amortize transition cost - STPM builds empirical distribution of network transfers - Switches to CAM when it predicts many transfers likely in future Manish Anand

10. Respect the Critical Path Many applications are latency sensitive - NFS file accesses - Interactive applications - Performance and Energy critical Other applications are less sensitive to latency - Prefetching, asynchronous write-back (Coda DFS) - Multimedia applications (with client buffering) - Only energy conservation critical Applications disclose the nature of transfer: foreground or background Manish Anand

11. Embrace Performance/Energy Tradeoff Inherent tradeoff exists between performance and energy conservation STPM lets user specify relative priorities using a tunable knob Manish Anand

12. Adapt to the operating environment Must consider base power of the mobile computer Consider mode that reduces network power from 2W to 1W - Delays interactive application by 10% On handheld with base power of 2 Watts: - Reduces power 25% (from 4W to 3W) - Energy reduced 17.5% (still pretty good) On laptop with base power of 15 Watts: - Reduces power by only 5.9% - Increases energy usage by 3.5% - Battery lasts longer, user gets less work done Manish Anand

13. Outline • Motivation • Self Tuning Power Management • Design Principles • Implementation • Evaluation • Related Work andSummary Manish Anand

14. STPM Architecture User or Energy Aware OS Manish Anand

15. Transition to CAM STPM switches from PSM to CAM when: • Application specifies max delay < beacon period • Disclosed transfer size > break-even size • Many forthcoming transfers are likely To predict forthcoming transfers STPM generates an empirical distribution of run lengths Transfers >150ms >150 ms >150 ms Run Run Run Run Manish Anand

16. Intuition: Using the Run-Length History A good time to switch Switch when expected # of transfers remaining in run is high Manish Anand

17. Expected Time to complete a Run Expected time to execute transfers in PSM mode Expected to execute rest of the transfers in CAM mode Time penalty for making a PSM to CAM switch Manish Anand

18. Expected Energy to complete a Run • Energy calculation includes base power Manish Anand

19. Performance and Energy Tradeoff Calculate expected time and energy to switch after each # of transfers • What if these goals conflict? • Refer to knob value for relative priority of each goal! Manish Anand

20. Outline • Motivation • Self Tuning Power Management • Design Principles • Implementation • Evaluation • Related Work andSummary Manish Anand

21. Evaluation Client: iPAQ handheld with Cisco 350 wireless card Evaluate STPM vs. CAM, PSM-static, and PSM-adaptive: • NFS distributed file system • Coda distributed file system • XMMS streaming audio • Remote X (thin-client display) Run DFS workload to generate access stats for STPM • Use Mummert’s file system trace (SOSP ’95) • File system operations (e.g. create, open, close) • Captures interactive software development Manish Anand

22. Results for Coda Distributed File System Workload: 45 minute interactive software development activity Energy (Joules) Time (Minutes) STPM: 21% less energy, 80% less time than 802.11b power mgmt. Manish Anand

23. Results for Coda on IBM T20 Laptop Same workload as before: effect of base power on power mgmt strategies Time (Minutes) Energy (Joules) PSM-Static and PSM-Adaptive use more energy than CAM! Manish Anand

24. Results for XMMS Streaming Audio Workload: 128Kb/s streaming MP3 audio from an Internet server Effect of knowing application intent Power (Watts) • XMMS buffers data on client: • App not latency sensitive • PSM uses least power STPM: 2% more power usage than PSM-static – no dropped pkts Manish Anand

25. Related Work • Lu, Y.H., Benini, L., AND Micheli, G.D. Power-aware operating systems for interactive systems. IEEE Trans. on VLSI (April 2002) • Simunic, T., Benini, L., Glynn, P. and Micheli, G.D. Dynamic Power Management for Portable Systems. Mobile Computing and Networking (2000) • Kravets, R., and Krishnan, P.Application-driven power management for mobile communication. ACM Wireless Nets. (2000) • Shih’s Wake on wireless: (MOBICOM '02) • Krashinsky’s BSD Protocol:(MOBICOM '02) Manish Anand

26. Summary STPM adapts to: • Base power of mobile computer • Application networkaccess patterns • Relative priority of performance and energyconservation • Characteristics of network interface Compared to previous power management policies, we perform better and conserve more energy Manish Anand

27. Self-Tuning Wireless Network Power Management Manish Anand Edmund B. Nightingale Jason Flinn Department of Electrical Engineering and Computer Science University of Michigan

28. Expected Time to complete a Run Expected time to execute transfers in PSM mode Expected to execute rest of the transfers in CAM mode Time penalty for making a PSM to CAM switch Consider the case of switching before the 3rd transfer: Manish Anand

29. Results for tuning performance/energy Same workload as before: effect of tuning relative priorities CAM knob=100 • Decreasing the knob value never yields increased energy usage • Increasing the knob value never yields reduced performance PSM-static knob=95 knob=90 PSM-adaptive knob=80 knob=0-70 Manish Anand

30. Self Tuning Power Management STPM adapts to: • Base power of mobile computer • Application networkaccess patterns • Relative priority of performance and energyconservation • Characteristics of network interface Compared to previous power management policies, we perform better and conserve more energy Manish Anand

31. Results for Non Hinting Applications Running Mummert’s purcell trace on Coda Time (Minutes) Energy (Joules) STPM without hints: 16% less energy, 72% less time than 802.11b Power Management Manish Anand

32. Results for executing a web trace Result of executing a 45 minute BU web trace TIME ENERGY • CAM performs only 0.8% better than PSM-static while expending 62% more energy • STPM behaves like PSM-static when conserving energy and like CAM in presence of abundant energy Manish Anand

33. Results for Remote X (No Think Time) Energy (Joules) Time (Minutes) STPM uses less energy than CAM if think time > 6.5 seconds Manish Anand

34. Managing Other Devices with STPM STPM well-suited for power management when: • Performance / energy conservation tradeoff exists • Transition costs are substantial Consider disk power management: • Web browser, DFS, mobile DB cache data locally • Hard drive spins down for power saving • Significant transition cost to resume rot. latency • Faster, less energy to read small object from server • But, if many accesses, want to spin-up disk For what other devices can STPM be applied? Manish Anand

35. Expected Cost Calculation Manish Anand

36. STPM as a wireless power management strategy • Holistic solution • Application intent through hints • Proactive solution using run histogram • Nature of network transfer : foreground or background • Performance/Energy tradeoff with a tunable knob • Operating Environment: base power Manish Anand