self-tuning wireless network power management

Self-Tuning Wireless Network Power

Management

CSC 547 - Fall 2013, University of ArizonaSumin Byeon

Problems

• Wireless card shortens battery life

• Obvious solution: turn it off!

• Power management can degrade performance

• Critical for latency sensitive applications

• No one-fits-to-all solution

Previous Work

• Power Saving Mode (PSM) in 802.11

• Will refer this as “PSM-static”

• Does not adapt to power characteristics of the network and mobile devices

• Does not adapt to usage patterns

Solution

• Self-tuning power management module that adapts to

• Usage patterns

• Network interface characteristics

• System (i.e., hardware)

Design Principles

• Know application intent

• Be proactive

• Respect the critical path

• Embrace the performance/energy tradeoff

• Adapt to the operating environment

Know Application Intent

• Not knowing intent of applications - Leads to either too conservative or too extravagant power management

• Allow applications to disclose “hints” - STPM can work more efficiently

Be Proactive

• Reactive strategy - Cost of transition between modes must be low

• Transition time 200-600ms; exceeds perception threshold (50-200ms)

• STPM performs cost-benefit analysis based on hints

• Not always possible

Respect the Critical Path

• Perception threshold - generally 50-200ms

• Foreground transfer - latency sensitive, interactive, synchronous

• Background transfer - latency tolerable, asynchronous

Performance-Energy Tradeoff

• Inherent tradeoff between performance and energy conservation

• Static threshold won’t work

• STPM provides mechanism to adjust priorities

• Is performance your priority? Or energy conservation?

Adapt to Environment

• Global optimization

• Excessive power saving on network interface may lead to inefficient energy usage on other components

• Consider base power usage

Implementation

• Implemented as Linux kernel module

• Inputs:

• Base power and current tradeoff between energy conservation and performance

• Device-specific power usage characteristics and transition costs

API

Power Cost Analysis

1. Base power

2. Power usage for each mode

3. Translation cost between modes

4. Power usage for data transfer in each mode

Algorithm

• Assume network card only supports two modes: CAM and PSM

Transition to CAM

• Delay tolerance is less than the maximum latency of PSM (i.e., beacon interval)

• Forthcoming transfer will be large enough

• Forthcoming transfer will be large enough, and STPM expects that there will be enough subsequent short transfers

Transition to PSM

• No transfers in progress

• No application specified delay tolerance

• Network card will be idle long enough

General Model

• Some network cards support more than two modes

• Number of possible strategies proportional to square of number of modes

• Calculates the lowest cost policy that transitions to each mode

• Calculates the lowest cost hybrid policies then make a further transition at some later time

Evaluation

• Question: How much did STPM improve the power efficiency? What was the impact on the performance?

Experiments

• Coda (distributed file system)

• NFS (Network File System)

• Xmms (audio streaming)

• Thin client using remote X

Coda

NFS

Xmms

Remote X

Effect of Think Time

Potential Weakness

• In their experiments, they made modifications on applications to emit hints

• Unrealistic to modify every single application in the world

• How would have STPM behaved if applications were not modified? (i.e., no hints)

Conclusion

• Power management may degrade performance, and even increase total energy consumption

• Power management must be tuned to reflect application intent and characteristics of network interface card

• Can’t expect users to manually tune power management algorithm

Conclusion (cont.)

• Results show that STPM improves both performance and energy conservation

• Future work: adapt STPM to other network cards and different system components (e.g., disks)