Download - A Time Series-based Approach for Power Management in Mobile Processors and Disks

Transcript
Page 1: A Time Series-based Approach for Power Management in Mobile Processors and Disks

NUS.SOC.CS5248

A Time Series-based Approach for Power

Management in Mobile Processors and Disks

X. Liu, P. Shenoy and W. Gong

Presented by Dai Lu

Page 2: A Time Series-based Approach for Power Management in Mobile Processors and Disks

NUS.SOC.CS5248

Contents IntroductionTime Series based Power

Management Utilization Measurement Prediction Model Speed Setting Strategy

ImplementationEvaluationSummary

Page 3: A Time Series-based Approach for Power Management in Mobile Processors and Disks

NUS.SOC.CS5248

IntroductionMultimedia applications prevalent

on mobile devices 3G/4G wireless network

Devices more and more powerful Samsung SPH-V5400 hand phone is equipped

with a 1.5 GB micro driveEnergy is a scarce resource

Page 4: A Time Series-based Approach for Power Management in Mobile Processors and Disks

NUS.SOC.CS5248

Previous Work CPU

DVFS: Dynamic Voltage and Frequency Scaling Infer task periodicity by work-tracking

heuristic Assume implicit deadlines for interactive

applications Only periodic applications; assumes

applications tell OS their periods and work amount

DiskDRPM: Dynamic Rotations Per Minute Monitor disk request queue length On-disk cache impact not considered

Page 5: A Time Series-based Approach for Power Management in Mobile Processors and Disks

NUS.SOC.CS5248

Why DRPM? Power- RPM relation

Ke: spindle motor voltage R: motor resistance ω: angular velocity

Similar to DVS for processors (P~fV2)

Page 6: A Time Series-based Approach for Power Management in Mobile Processors and Disks

NUS.SOC.CS5248

Contents Introduction Time Series based Power

Management Utilization Measurement Prediction Model Speed Setting Strategy

Implementation Evaluation Summary

Page 7: A Time Series-based Approach for Power Management in Mobile Processors and Disks

NUS.SOC.CS5248

New WorkLow overhead

Prediction with simple statistical model in time series analysis

Processor + disk TS-DVFS + TS-DRPM

Different CPU scaling factor for different tasks Enable coexistence of MM and non-

MM applications

Page 8: A Time Series-based Approach for Power Management in Mobile Processors and Disks

NUS.SOC.CS5248

TS-PM enabled OS kernel

Page 9: A Time Series-based Approach for Power Management in Mobile Processors and Disks

NUS.SOC.CS5248

Prediction Model Box-Jenkins model in time series analysis

Assume a stationary process Statistical properties (mean, variance)

are essentially constant through time. Firs-order autoregressive process (AR(1))

predictor ũt = Φ1 ũt-1+at

Φ1: Correlation coefficient at: Error/ random shock

Sample Autocorrelation Function (SAC)

Page 10: A Time Series-based Approach for Power Management in Mobile Processors and Disks

NUS.SOC.CS5248

Prediction Model Cont.

Estimated demand:

Estimated mean:

Estimated constant( SAC):

TS-DVFS: one AR(1) for every task

TS-DRPM: a single AR(1)

Page 11: A Time Series-based Approach for Power Management in Mobile Processors and Disks

NUS.SOC.CS5248

Measuring utilization CPU

e: full-speed execution time

q: time quantum allocated to the task

Disk r: response time s: scaling factor

Page 12: A Time Series-based Approach for Power Management in Mobile Processors and Disks

NUS.SOC.CS5248

Speed Setting Strategy TS-DVFS

Two level CPU setting Interval T

Subinterval within T

Page 13: A Time Series-based Approach for Power Management in Mobile Processors and Disks

NUS.SOC.CS5248

Speed Setting Strategy TS-DRPM

Performance slow-down Pdiff[i] = a(1-h) × T × Rdiff[i]

Estimated utilization ûi = û + Pdiff[i]/ Th: hit ratea: arrival rateRdiff: rotational latency difference

Choose the lowest RPM level satisfying (ûi- ûmax) / ûmax ≤ threshold

Page 14: A Time Series-based Approach for Power Management in Mobile Processors and Disks

NUS.SOC.CS5248

Contents Introduction Time Series based Power Management

Utilization Measurement Prediction Model Speed Setting Strategy

Implementation Evaluation Summary

Page 15: A Time Series-based Approach for Power Management in Mobile Processors and Disks

NUS.SOC.CS5248

Implementation CPU

300-677 MHz, Transmeta Divide into 5 steps Mapping scaling factor to frequency level

Disk 3000-5400 RPM Divide into 5 steps Assumed power consumption level Trace driven simulation with DiskSim

Page 16: A Time Series-based Approach for Power Management in Mobile Processors and Disks

NUS.SOC.CS5248

Frequency and RPM Mapping

Page 17: A Time Series-based Approach for Power Management in Mobile Processors and Disks

NUS.SOC.CS5248

Contents Introduction Time Series based Power Management

Utilization Measurement Prediction Model Speed Setting Strategy

Implementation Evaluation Summary

Page 18: A Time Series-based Approach for Power Management in Mobile Processors and Disks

NUS.SOC.CS5248

TS-DVFSUp to 38.6% energy saving against LongRun

Page 19: A Time Series-based Approach for Power Management in Mobile Processors and Disks

NUS.SOC.CS5248

TS-DRPMUp to 20.3% saving against TPMperf (oracle)

Page 20: A Time Series-based Approach for Power Management in Mobile Processors and Disks

NUS.SOC.CS5248

Summary Time series statistical model TS-DVFS TS-DRPM

Comments General PM, no QoS measurement like deadline

miss rate Multiple rotational speed disk not commercially

available Increase the accuracy of profiling disk access

patterns. “Hit if response time < τ, otherwise miss.”

Page 21: A Time Series-based Approach for Power Management in Mobile Processors and Disks

NUS.SOC.CS5248

References Chameleon: Application Controlled Power

Management with Performance Isolation, X. Liu and P. Shenoy, Technical report 04-26, Department of Computer Science, University of Massachusetts

Forecasting and time series: an applied approach 3rd ed, Bowerman and O’Connell, Duxbury, 1993

Reducing disk power consumption in servers with DRPM, S. Gurumurthi, A. Sivasubramaniam and H. Franke, IEEE Computer, Dec 2003