260 likes | 351 Views
Turducken: Hierarchical Power Management for Mobile Devices. Jacob Sorber, Nilanjan Banerjee, Mark Corner, Sami Rollins†. University of Massachusetts, Amherst. http://prisms.cs.umass.edu/hpm. Batteries and Wireless: Two Achilles Heels. Imagine using a laptop like it was a desktop:
E N D
Turducken: Hierarchical Power Management for Mobile Devices Jacob Sorber, Nilanjan Banerjee, Mark Corner, Sami Rollins† University of Massachusetts, Amherst http://prisms.cs.umass.edu/hpm
Batteries and Wireless: Two Achilles Heels • Imagine using a laptop like it was a desktop: • your mail is always fresh • your file system is always consistent • your web is always prefetched • instant messages are instantaneous • Why isn’t this a reality? • Mobility leads to variable connectivity • Limited energy necessitates shutting laptop off • The result is we don’t even try!
Tension in Maintaining Availability • The laptop can remain active—available for communication • System lifetime is severely limited • The laptop can be suspended—extending lifetime • System is not available • Compromise: wake periodically to communicate • More frequent = more available, shorter lifetime • Less frequent = less available, longer lifetime • The tension remains despite current methods • Don’t know when to wake up • Systems lean toward longer lifetime • Goal: A system that is always available with extended lifetime
Solution: Hierarchical Power Management • Hierarchical Power Management • Combine a range of embedded platforms into a single device • Perform tasks on the tier that best fits the task • Download email using PDA • Download large movies with laptop • Sleep high-power tiers when not needed • Focus on non-interactive use • Provide power-efficiency when NOT in use • Remove the human bottleneck • Allow new laptop functionality
Outline • HPM Design Principles: • Combine the Right Platforms • Use the right device for the job • Decompose Tasks • Design of an HPM Laptop • Implementation of Turducken • Hardware • Software • Experimental Results • Related Work and Conclusions
Combine the Right Platforms • More platforms = more overhead • Size, weight, and complexity • Suspend power • Transition cost between platforms • Still need to support a large range of power modes • Use few devices separated by orders-of-magnitude in power
Use the Right Device for the Job 40W Laptop Tasks requiring large memory or storage space Very complex computation 2W PDA Power Needs Download Web Pages Send/Recv Mail 0.04W Sensor Simple computation and storage
Decompose Tasks • Not only separate platforms, but subdivide tasks • Tasks include common subtasks: • Checking for network connectivity • Service Discovery • Check for work to be done • Interact with user • Example: • Mote checks for a network, PDA checks for mail server • Laptop remains asleep until user reads the mail
Outline • HPM Design Principles: • Combine the Right Platforms • Use the right device for the job • Decompose Tasks • Design of an HPM Laptop • Implementation of Turducken • Hardware • Software • Experimental Results • Related Work and Conclusions
HPM Software Design • Take advantage of transparent proxies • Proxies periodically perform service discovery and fetch content • Provide it on demand to higher tiers in system • Many applications are already distributed • Lower tiers act as proxy for higher tiers • Little or no application recoding and invisible to users
Outline • HPM Design Principles: • Combine the Right Platforms • Use the right device for the job • Use the right number of devices • Design of an HPM Laptop • Implementation of Turducken • Hardware • Software • Experimental Results • Related Work and Conclusions
Hardware Implementation • Combines Laptop, PDA, and mote into a laptop • Laptop and PDA communicate via USB • PDA and mote via RS-232 • Wakeup using relays • WiFi detector allows mote to detect 802.11 beacons
Applications • Time Synchronization • Can be completely performed by the mote • Lowest tier periodically synchronizes with remote time server • Web Prefetching • More complicated application which requires at least PDA • Mote detects wireless networks • PDA periodically fetches updates to frequently accessed pages • E-Mail • Involves user interaction • Both incoming and outgoing updates • Laptop accesses mail via proxy on the PDA
Evaluating Turducken • How does HPM affect battery lifetime? • Show results for a variety of workloads • Vary the consistency of the application • Vary the availability of the network • Consider three configurations • Laptop • Laptop+Mote • Laptop+PDA+Mote • Experiments take a long time • Extrapolate from a set of measured data points
Effect of Consistency • Measure the impact of maintaining consistency • System wakes periodically—vary frequency • Network is always present • Time Synchronization (please see paper) • Web Prefetching • Laptop routes network traffic through PDA • PDA periodically refreshes expired pages • E-Mail • System wakes with constant rate to send mail (2 min/hr) • PDA periodically checks for new mail and sends pending mail
Results: Web Prefetching Lifetime (hours) Wake Interval (minutes) • PDA is the right device for the job • For always-on operation, Turducken lasts up to 10 times longer
Results: E-Mail Lifetime (hours) Wake Interval (minutes) • More interactive use reduces opportunities for Turducken to save power • Turducken still lasts up to 6 times longer
Power-By-Tier: Web Caching • Laptop+Mote shows cost of adding unnecessary platforms • Turducken replaces 400mW of laptop active power with 25mW on PDA • Laptop suspend power is the greatest limiting factor: try hibernation Laptop Laptop+Mote Full Turducken
Effect of Network Availability • Measure the impact of intermittent connectivity • Network is present with varying probability • Mote wakes at a fixed interval • Only wakes PDA when network is available • Only show results for Web Prefetching
Results: Web Prefetching Lifetime (hours) Probability of WiFi • With no network mote prevents laptop from waking • Turducken performs well over a variety of connectivity patterns
Related Work • Wake-On-Wireless [Shih, et al] • Low-power radio receives wakeup messages • Requires in-network support • Focus on when to wake up • Wireless optimizations [Anand, et al], Disk Spindown [Helmbold, et al], Multiple radios [Pering, et al], Multiple processor cores [Mayo & Ranganathan] • none provide orders-of-magnitude scaling of power
Conclusions • Always-on availability and extended battery life • Hierarchical Power Management • Integrate multiple independent platforms • Pass control between tiers • Suspend unneeded tiers • Use proxy-based approach • Little or no application recoding • Always-on operation for up to 10x as long
Turducken: Hierarchical Power Management for Mobile Devices Jacob Sorber, Nilanjan Banerjee, Mark Corner, Sami Rollins† University of Massachusetts, Amherst http://prisms.cs.umass.edu/hpm