Runtime Software Power Estimation and Minimization Tao Li

Runtime Software Power Estimation and Minimization

Tao Li

Power-aware Computing

Power: Software Perspective & Impact

Power estimation: the first step to power management & optimization

Software contributes to & largely impacts power consumption

It is crucial to model power from the perspective of software

Evaluate software energy in early design stage

Understand impact of software optimizations on energy

Support run-time power management and optimizations

Power: Software Perspective & Impact (Contd.)

Instruction level modeling Computation intensive

High level macro-modeling Difficult to apply to general code

Event counting based modeling Impacted by the availability of performance counters

Architecture level simulation Large slowdown

Software Power Estimation: Current Techniques

Challenges in Run-time Power Estimation

High fidelity & fast speed

On-the-fly estimation capability, non-intrusive & low overhead

Simplicity, availability and generality

Experimental Methodology

SoftWatt: cycle-accurate & full-system power simulation framework

SimOS infrastructure, Wattch power model

Commercial OS & real applications

Out-of-order superscalar processor

Caches & memory hierarchy

Low-power disk

Experimental Methodology (Contd.)

Applications

E-mail and file management (sendmail, fileman)

Java (SPECjvm98: db, jess, javac, jack, mtrt, compress)

SPECInt95 (gcc, vortex)

Database (Postgres: select, update, join)

Miscellaneous (pmake, osboot)

OS Power Characterization OS power varies from one application to another

29 Watt (gcc) ~ 66 Watt (fileman)

Variance of power consumption in OS service routines & invocations

0102030405060

Av

g. P

ow

er

(W)

0246810121416

Std

. D

ev

. (%

)

Avg. Power (W) Std. Dev.(%)

OS Power Characterization (Contd.)

OS routine power correlates with its performance

Circuits used to exploit ILP burn significant portion of power

The number of in-flight instructions that flow through impacts circuit switching activity

For a given OS routine, similar IPC indicates similar circuit switching activity and therefore, similar power

OS Routine Power-Performance Correlation

SCSI Disk Interrupt Handler Read File System Call

Routine Level OS Power Model

Idea: use a linear regression model

Proutine=k1*IPCroutine+k0

to track the OS routine power showing different performance

Energy(OS)= Sum [ Energy(OS routines) ]= Sum [ Power(OS routines)*Time(OS routines) ]

Routine Level OS Power Model (Contd.)

Regression Model P = k1×I PC+k0 OS

Services k1 k0 ε

Comment

utlb 23.6 6.2 0.17% TLB miss handler

COW_fault 32.1 1.1 0.19% copy-on-write fault simscsi_ intr 33.9 1.3 1.94% SCSI disk I / O interrupt clock 36.4 0.6 2.68% clock interrupts read 29.6 4.7 4.53% read fi le write 34.3 1.5 1.27% write fi le open 34.3 1.2 0.41% open a fi le or serial port

close 30.4 3.9 2.61% close an open channel

: Model Fitting Error

Pre-characterization Low level energy simulation

Model fitting

Run-time estimation OS routine boundaries Evaluation using counter values

Routine Level OS Power Modeling

Routine based Regression ModelProutine=k1*IPCroutine+k0

-2%

-1%

0%

1%

2%

se

nd

ma

il

file

ma

n

db

jes

s

po

stg

res

.se

lec

t

po

stg

res

.up

da

te

os

bo

ot

Es

tim

ati

on

Err

or

(%)

Flat Regression ModelPOS=g1*IPCOS+g0

-15%

-10%

-5%

0%

5%

10%

15%

sen

dm

ail

file

man d

b

jess

po

stg

res.

sele

ct

po

stg

res.

up

dat

e

osb

oo

t

Est

imat

ion

Err

or

(%)

Cumulative Estimation Error

Flat Regression Model POS=g1*IPCOS+g0

Per-routine Estimation Error

Routine based Regression ModelProutine=k1*IPCroutine+k0

Per-routine Estimation Error (Contd.)

OS Energy Dissipation

0%10%20%30%40%50%60% % of OS Cycles

% of OS Energy

92% 89%

Phases in Programs(8-issue machine)

0

1

2

3

4

5

6

0.00 0.07 0.13 0.20 0.26Execution Time (seconds)

IPC

Benchmark: SPECjvm98 jessBenchmark: SPECjvm98 jess

Resources are utilized differently during different phases of program execution

Average IPC - User: 2.1, OS: 1.1

Power Minimization via Processor Resource Adaptations

Adapt processor resources to program needs

What can be adapted?

Bandwidth of fetch/decode/issue/retire…

Size of instruction window, re-order buffer,

load store queue…

Reduce power, retain performance

Effects of Tuning Processor Resource for the OS

8-issue -> 4-issue

OS Performance degradation: 4%

OS Power savings: 50%

1-issue 2-issue 4-issue 6-issue 8-issue

OS IPC 0.88 1.09 1.15 1.19 1.21OS Power(W) 6.4 12.2 21.7 31.1 42.8

Previous Approach for Adaptations

Sampling

Cycles

Sampling Window

IPC (Inst. Per Cycle)

Adaptation

A B C D E F

Problems with Sampling based Adaptations (Contd.)

OS executions Short-lived

AdaptationOverhead

User User UserOS OS User

sampling window

A B CTa

Ts

smallersamplingwindow

OS UserUser OS User

Th

OS-aware Routine based Adaptations

OS-aware: Identify OS executions via processor execution modes Just-in-time & full coverage of OS activities

Routine-based: Adapt processor resources at OS routine boundaries

Precise exceptions: drained pipeline Achieve minimum adaptation overhead

User UserOS OS User

OS Routine-basedOptimal Adaptations

Adaptations w ithMinimum Overhead

OS-aware Routine based Adaptations (Contd.)

Apply optimal adaptation for individual OS routine Exploit the routine level Energy-Delay Product

variance

0

0.2

0.4

0.6

0.8

1

clock COW_fault read

No

rma

lize

d E

ne

rgy

-De

lay

Pro

du

ct

1-issue2-issue4-issue6-issue8-issue

OS ServicesOS Services

Routine based Adaptations: OS Power

0

0.2

0.4

0.6

pmak

egcc

vorte

x

sendm

ail

filem

an dbje

ss

java

cja

ck

postgr

es.s

elec

t

postgr

es.u

pdat

e

osboot

AVG

No

rma

lize

d P

ow

er

Sampling based Adaptation (Window Size: 2048-cycle)Sampling based Adaptation (Window Size: 128-cycle)Routine based Adaptation

OS Performance

0.7

0.8

0.9

1

pmak

egcc

vorte

x

sendm

ail

filem

an dbje

ss

java

cja

ck

postgr

es.s

elec

t

postgr

es.u

pdat

e

osboot

AVG

No

rma

lize

d I

PC

Sampling based Adaptation (Window Size: 2048-cycle)Sampling based Adaptation (Window Size: 128-cycle)Routine based Adaptation

OS Power & Performance Tradeoff

0

0.2

0.4

0.6

0.8

No

rma

lize

d E

ne

rgy

.De

lay Sampling based Adaptation (Window Size: 2048-cycle)

Sampling based Adaptation (Window Size: 128-cycle)Routine based Adaptation