Energy Efficiency through Burstiness

Athanasios E. Papathanasiou and Michael L. ScottUniversity of Rochester, Computer Science Department

Rochester, NY 14623, U.S.A.

(WMCSA'03) 5th IEEE Workshop on Mobile Computing Systems and Applications

Presented by Hsu Hao Chen

outline

Introduction Design Guidelines Prototype Experimental Evaluation Conclusion

Introduction Smoothness:

OS resource management policies traditionally employ buffering

Maximize overall throughput and minimize the latency of individual requests.

A little energy efficiency Burstiness:

Can improve energy efficiency Without a significant impact on performance

Power-efficient devices

Energy consumption parameters for various disks

Linux File System Behavior(1/2)

CD copy (1359 sec) Disk Idle Time: 1191 seconds 92%: shorter then 5 seconds

Linux File System Behavior(2/2)

MP3 playback (300 sec) Disk Idle Time: 292 seconds 66%: shorter than 8 seconds 6% only: longer than 16 second

breakeven time

Disk Energy Savings vs. Memory Size

Example: MPEG playback 64~496MB memory increase 3.4% disk energy savings

Design Guidelines(1/3) Maximize idle phases

Aggressive, Speculative Prefetching With hints to improve accuracy

Bursty Periodic Update

Coordinating across applications Arrange for all applications to run out of data at the

same time Prefetch daemon

Design Guidelines(2/3)

User Space

Kernel Space

Application layer

Memory Management

Coordi natedActi vi ty

Hard Disk

Read Requests Wri te Requests

Hi nts

Prefetchd Kupdated

kswapd

Pages

Design Guidelines(3/3) Maintaining responsiveness

Responsiveness may be decreased because of Increased disk congestion due to burstiness Latency penalty for disk power-up

Solution: Preactivate disk Monitor application progress and file system cache state

Data consumption and production rates Initiate prefetch cycle before applications run out of data

Interactive responsiveness Prioritized disk queues

Required to service quickly unpredicted demand misses during periods of high disk congestion

Prototype(1/5) Epoch-based algorithm in the basic memory

management mechanisms of the Linux kernel

Each epoch consists of two phases: Request generation phase Idle phase

Prototype(2/5) Request generation phase

Flush all dirty buffers Predict future data accesses Compute the amount of memory that can b

e used Prefetching Storing new data

Free the required amount of memory Prefetch

Prototype(3/5) Idle phase

Estimate time to next request Disk spin-down

Predicted idle time is higher than the disk’s break-even time

New epoch triggered by A new prefetching cycle has to be initiated A demand miss took place One or more dirty buffers have expired and it is tim

e for them to be flushed Low system memory

Prototype(4/5) The Prefetch Cache

Prefetch cache size Large enough to contain all predicted data accesses Without causing eviction of useful data

Type of first miss determines prefetch cache size for next epoch Compulsory Miss Prefetch Miss Eviction Miss

Prototype(5/5)

Update Policy Update daemon

Flushes all dirty buffers once per minute. Modified openopen system call

Postpone write-behind until the file is closed Application without strict reliability constraints

Examples: Copying of a file or MP3 encoding of a CD track

Experimental Evaluation(1/7) Dell Inspiron 4100 Laptop

512MB of memory Hitachi DK23DA hard disk

Workload scenarios MPEG Playback (two 76MB files) Concurrent MPEG Playback and MP3 encoding

MPEG Player Input: two 76MB files MP3 Encoder Input: 10 WAV files (626MB) MP3 Encoder Output: 10 MP3 files (42.9MB)

Power management policy: Linux: 10 second fixed threshold Bursty: Predictive algorithm that monitors application progre

ss and file system cache state

Experimental Evaluation(2/7)

Cumulative distribution of disk idle time intervals during MPEG playback

Experimental Evaluation(3/7)

Distribution of disk idle time intervals during concurrent MPEG playback and MP3 encoding

Experimental Evaluation(4/7)

Experimental Evaluation(5/7) execution of the MPEG player on Linux with 492MB

Experimental Evaluation(6/7) execution of the MPEG player on Bursty with 64MB and 128MB

Experimental Evaluation(7/7)

Conclusion

Works well with predictable applications Energy savings scale with memory size Up to 78.5% disk energy savings Less than 5% performance penalty acr

oss all workloads and memory sizes