Automatic Monitoring for Interactive Performance and Power Reduction


Click here to start

Table of contents

Automatic Monitoring for Interactive Performance and Power Reduction

Overview

Research contributions

Response time

Episode classification

A utilization trace

Slide 7

Mouse movement

Interactive episodes

Interactive episodes can include idle time

Finding interactive episodes

Tracking interactive episodes

Communication between tasks

Does multiprocessing improve interactive performance?

Why use TLP?

Initial results

Workloads and TLP results

Methodology

Response-time improvement over uniprocessor

Background activity: MP3 playback

Time above the perception threshold

Characteristics of Interactive Episodes

Why bother?

Power Density!

Dynamic Voltage Scaling

Processors supporting DVS

Small performance reduction = big energy savings

The key: performance-setting algorithm

Producer and consumer episodes

Cumulative interactive episode length distribution

Slide 31

Performance-setting strategy for interactive episodes

Performance-setting for interactive episodes

Performance-setting algorithm

Advantages

Performance-setting during the Acrobat Reader benchmark (200ms p.t.)

Performance-setting during the Acrobat Reader + MP3 benchmark (200ms p.t.)

Hardware assumptions

Energy factors (no MP3)

Energy factors with MP3 playback

Changes in cumulative episode lengths as the result of performance scaling (Xemacs 50ms p.t. )

Desired improvements

Applicability to other environments

Conclusions

Future work

PowerPoint Presentation

The performance gap

Computing the performance factor for interactive episodes

Performance scaling

Energy-delay (no MP3)

Energy-delay (MP3)

Author: Krisztian Flautner

E-mail: manowar@engin.umich.edu

Homepage: www.flautner.com