e reditive ystem hutdown wetho d for inergy ving of
iventEhriven gomputtion
ghiErong rwng nd ellen gFErF u
heprtment of gomputer ieneD sing ru niversity
rsinhuD iwnD QHHD epuli of ghin
fhunghwdsFnthuFeduFtwg
estrt sn reent studyD rivstv et lF U onduted
n extensive nlysis on dierent systemEshutdown pE
his pper presents systemElevel power mngement
prohesF heir smple tres on n EserverD n eventE
tehnique for power sving of eventEdriven pplitionsF
driven pplitionD showed tht it sp ends over WH7 of its
e present new preditive systemEshutdown methodto
time in the lo ked stte nd the verge time visiting
exploit sleep mode opertions for power svingF e use
the lo ked stte is less thn one seondF es resultD
n exponentilEverge pproh to predit the upoming
the onventionl pproh will fil to eetively hieve
id le periodF e introdue two mehnismsD preditionE
power redution on this typ e of pplitionF hey hve
miss orretion nd preEwkeupD to improve the hit rtio
prop osed preditive systemEshutdown tehnique for enE
nd to redue the dely overhedF ixperiments on four
ergy sving of eventEdriven pplitionsF hey rst olE
dierent eventEdriven pplitions show tht our proE
leted smple tres of onEo tivity on n EserverF
posed method hieves high hit rtios in wide rnge of
henD they prop osed two preditive formuls sed on
dely overhedsD whih results in high degreeofpower
the nlysis of the smple tresF he rst formul ws
sving with low dely penltiesF
otined y using generl regressionEnlysis tehnique
I sntro dution
to orrelte the length of the up oming o p erio d to the
ith the dvent of p ortle omputing nd high denE
lengths of the previous on nd o p erio dsF he seond
sity vs iruitsD p ower dissiption hs emerged s
formul ws otined y the oservtion of the onEo E
priniple design onsidertion in vs designsF sn the
tivity ehviorF hey oserved tht the idle p erio d folE
pst few yersD hndful of p ower estimtion nd miniE
lowing long running p erio d tends to e short4F fsed
miztion metho ds hve een rep orted for hieving low
on this oservtionD they derived formul whih lters
power designs t iruitD lyoutD logiD ehviorlD nd
out the idle p erio d fullling the ove ondition nd the
rhiteturl levelsF everl exellent reviews of p ower
system enters the sleep stte otherwiseF fsed on the
estimtion nd minimiztion tehniques re given yeE
two formulsD they onduted series of exp eriments on
drm ID hevds nd wlik PD nd x jm QD RF
n lient pplitionF he results demonstrted tht
vowpower vs designs n e hieved t vriE
preditiveshutdown tehniques n redue the p ower
ous design levelsF sn this studyDwe fo us on utilizing
onsumption y lrge ftor ompred to the onvenE
preditive systemEshutdown tehnique for p ower svE
tionl metho dF roweverD one drwk of this pproh
ing of eventEdriven pplitionsF ower mngement
is tht the preditive formul is diretly derived from
tehniques hve een extensively pplied to g systemsF
the smple tres of sp ei pplitionF por dierent
por instneD the design of the owergTHQ S pplied
pplitionsD dierent preditive formuls re needed in
twotyp es of p ower mngementshemesX stti nd dyE
order to mke urte preditionsF
nmiF sn stti p ower mngementD the system denes
sn this pp erD we present new preditive systemE
severl sleep mo des with vrious levels of p ower sving
shutdown metho d for eventEdriven pplitionsF e use
nd delyoverhed whih n e ontrolled externlly
wellEknown exp onentilEverge pproh to predit
y softwreF sn dynmi p ower mngementD the system
the up oming idle p erio dF eintro due two mehE
will utomtilly detet the idle p erio ds nd disle the
nismsD preditionEmiss orretion nd preEwkeupD to imE
lo ks on p ortions of the g F epple9s w owerfo oks
prove the hit rtio nd to redue the delyoverhedF
T use dierent pproh whihenters rest mo de fter
ixp eriments on four dierenteventEdriven pplitions
P seonds of idle timeF huring rest mo deD the pro essor
re rep orted to demonstrte the eetiveness of our proE
is p owered down ut the sGy devies remin onF ypiE
p osed metho dF
llyD onventionl shutdown pprohes re rried out
sed on the rule of go to sleep fter the system hs een
P he prop osed metho d
id le for predenedperiod of timeFroweverD this pE
PFI inergy sving using systemEshutdown
proh p oses n ovious drwk E the system ontinues
tehnique
to onsume p ower in tht intervl of idle timeF
pigure I depits simple shutdown pproh for
y
eventEdriven pplitionsF hen the system detets n
his work ws supp orted in prt y the xtionl iene
idle p erio dD it will determine whether it should styin
gounil of FyFgF under qrnt xg VSEPPPIEiEHHUEHQR nd xg
the running stte or enter the sleep stteF sf the system VTEPPPIEiEHHUEHPIF
0-89791-993-9/97 $10.00 1997 IEEE
! iY @IA Wake−up s
Running Sleep
iq a s @i C A Y @PA
i
a s @i C A @s i AY @QA
i
Go to sleep
a s @ A i @ A Y @RA
i i
pigure IX e simple shutdown pprohF
b HY @SA
i @ AC
i
RI i
ab s b @TA
R R I R Y
@ A
delay delay
i @ AC
i i
X @UA ab a
th
@ A
RESWR R EW R
ve derivtions indite tht iqnsF I nd U re
(a) (b) he o
the neessry onditions under whih the idle p erio d will
pigure PXwo p ossile senrios when pplying simple
hieve energy svingsF e dene s the threshold
th
systemEshutdown shemeX@A s ! i D@A s i F
idle p erio d tht results in energy sving of the systemF
deides to enter the sleep stteD the system rst p erforms
PFP redition of idle p erio ds
num er of housekeeping pro eduresD suh s kingE
he nlysis in the previous setion indites tht
up dt nd system sttusF he system then stys in
prediting the idle p erio d s is vitl for n eetivepower
the energyEsving sleep stte until n externl wkeup
mngement mehnismF roweverD the distriution nd
signl o ursF hen wkeup signl o ursD the system
vrition of s my e strongly dep endent on the user eE
will p erform reovery pro eduresD suh s dt restoringD
hviorD trget pplitionsD working frequenyD nd opE
nd then resume the running stteF sf the deision is not
erting systemF
to enter the sleep stteD the system stys in the running
sn our pprohD we dpt the exponentilEverge
stte s usy witingF yne of the ruil issues of this
pprohV used in the g sheduling prolem for
pprohiswhether or not to shutdown the systemD nd
the predition of the idle p erio dF sn the g sheduling
if so whenc4 so tht p ower dissiption of the system n
prolemD op erting systems need to predit the length of
e reduedF his issue will e disussed s followsF
the next g urst in order to mke pproprite pro ess
vet s e the idle time p erio dD i the delyoverhed
shedulingF sn generlD the next g urst is predited
of entering the sleep stte from the running stteD the
s n umultiveverge of the mesured lengths of
sleeping time p erio dD nd the delyoverhed for reE
previous g urstsF imilrlyDwe n predit the next
suming the running stte from the sleep stte @the wke
idle p erio d y the umultiveverge of the previous
up pro essAF nd re the p ower onsumption vlE
idle p erio dsF he reursive predition formul is shown
ues of the system in the running nd sleep sttesD resp eE
elowF
tivelyF is the verge p owerEdissiption overhed
i
s a i C@I A s Y @VA
nCI n n
of entering the sleep stte from the running stte nd
resuming the running stte from the sleep stteF ypiE
where s is the new predited vlueD s is the lst
nCI n
llyD ! ! F pinllyD iq denotes the energy
i
predited vlueD i is the ltest idle p erio dD nd is
n
gin of the systemF
onstnt ttenution ftor in the rnge etween H to
IF sn this formulD s is the inerti nd i is the fore
n n
essume tht the system will enter the sleep stte
to push the predited idle p erio d towrd the tul idle
whenever it detets n idle p erio dF pigure P@A shows
p erio dF e n use iqnF V to predit the up oming idle
the rst senrio in whih the idle p erio d is longer thn
p erio dD whih is funtion of the ltest idle p erio d nd
the delyoverhed of entering the sleep stte from the
the previous predited vlueF he prmeter ontrols
running stte @iFeFD s ! i AF ine s ! i D the system will
the reltiveweight of reent nd pst history in the preE
suessfully enter the sleep stte nd resume the running
ditionF sf a HD then s a s F sn other wordsD the
nCI n
stte when wkeEup signl o ursF nder this ondiE
reent history hs no eetF yn the other hndD if aID
tionD the energy gin iq is s @i C A
i
then s a i F sn this seD the predition only tkes
nCI n
while the dely p enltyis F pigure P@A shows the seE
into ount the most reent idle p erio d ut ignores the
ond senrio in whih the idle p erio d is shorter thn the
previous preditionsF sn our implementtionD is set to
delyoverhed of entering the sleep stte from the runE
e IaP so tht the reent history nd pst history re
ning stte @iFeFD s i AF sn this seD the system will
eqully weightedF e n expnd iqnF V s elowF
never enter the sleep stte nd will suer long dely
p enlty@del y a C@i sAA nd negtive energy
n nCI
gin @iq a s @i C A AF reneD in order
i
s a i C @I Ai C C@I A i C@I A s X
nCI n n I H H
to hieve energy svingD the idle p erio d must e longer
@WA
thn the delyoverhed for entering the sleep stte from
iqnF W indites tht the predited idle p erio d is the
the running stte @iFeFD s ! i AF
weighted verge of previous idle p erio dsF irly idle
vet us further nlyze the minimum required idle p eE p erio ds hve less weight s sp eied y the exp onentil
rio d for hieving p ositive energy gin s followsF ttenution ftorsF I1R1 I2 R2 I3 R3 I4 R4 RIR RIR
I3’ I4’ delay
(a) RESWR REWS I R I1R1 I2 R2 I3 R3 I’ I’ Busy S waiting (a) (b)
I3’ H
pigure RX ixmples of preEwkeup mo delX@A s b
Sth H
D @A s `sF I3’’ s nd h
2 Sth
o resolve the seond prolemD we dd sturtion I3’’’
(b)
ondition to iqnF V s elowF
pigure QX gorretion of predition missX @A n impulseEli ke
idle p erio dD @A orretion using wthdog shemeF
if @i C@I As bs A
n n n
s a s Y @IHA
nCI n
PFQ gorretion of predition misses
where is onstntF nder the sturtion onditionD
he prop osed predition formul @iqnF VA n eeE
the growing rte of s is limited to times p er up dteF
tively predit idle p erio ds in most ses exept the o urE
rene of impulseElike idle p erio ds E suddenD very long
PFR reEwkeup
idle p erio d s Q o urs fter ontinuousD nerly uniform
es desri ed in the previous setionD when the sysE
idle p erio ds E during the predition pro essD s shown in
tem resumes the running stte from the sleep stte @iFeFD
pigure Q@AF por exmpleD the user works on the system
system wkeEupAD the system needs to p erform some reE
for while nd then go es to nswer telephone resulting
overy pro eduresF sn other wordsD the system suers
in the system idling for long p erio d of timeF hen suh
dely p enltyof y restoring the system sttusF
n impulseElike idle p erio d o ursD the predition of the
his dely p enltymyhve gret impt on system
up oming idle p erio d s Q nd the following one s R will not
resp onsiveness in some sesD esp eilly when is to o
e urteF he resons re disussed s followsF ell
long to e negletedF yne wy to resolve this prolem is
tht our prop osed predition formul @iqnF VA predits
to preEwkeup the system efore the rrivl of the next
the up oming idle p erio d y the umultiveverge of
wkeup signlF his n e omplished y prediting
the previous idle p erio dsF hen very long idle p erio d
the o urrene of the next wkeup signlF
H
o urs fter ontinuousD nerly uniform idle p erio dsD the
vet s e the tul idle p erio dD s the predited idle
H
predited vlue of this long idle p erio d is often muh
p erio dD nd h a js s j the error of the preditionF pigE
H
lower thn the tul idle p erio d @s Q bsQAF his unE
ure R depits two p ossile senrios when pplying the
derestimtion is undesirle for energy svingD esp eilly
preEwkeup shemeF sn the rst senrioD weoverestiE
H
when the predited vlue is lower thn F sn this seD
th
mte the predited idle p erio d y h @iFeFD s bsA nd
the system will sty in the running stte insted of enE
h F sn this seD the system will wkeup h
tering the sleep stte whih results in lrge mount
time hed of the next wkeup signlD s shown in pigE
of unneessry p ower onsumptionF yn the other hndD
ure R@AF husD the dely p enltyish whih is shorter
when prediting the idle p erio d s R whih is followed
thn the originl dely p enlty F sn dditionD the enE
long idle p erio dD our prop osed formul tends to overesE
ergy gin is iq a s @ A @i C A @
i
H
timte the durtion of the idle p erio d @s R `sRAF st is
H
A h F yn the other hndD if s bs nd h ! D
lso undesirle euse the system my flsely enter the
the system will e woken up y the originl wkeup sigE
sleep stte nd suers unneessry p ower onsumption
nlF sn this seD the preEwkeup hs no eet on the
nd delysF
redution of the dely p enltyF pigure R@A shows the
H
o llevite the rst prolemD we use wthdog
seond senrio in whih s `sF sn this seD the deE
sheme to p erio dilly monitor the urrent idle p erio dF
ly p enlty is zero ut the energy gin is redued to
hen n idle p erio d o ursD the system predits the duE
iq a@s hA@ A @iCA@ AF
i
rtion of the idle p erio dF sf the predited vlue is lower
sn other wordsD when the predited idle p erio d is less
thn D the system stys in the usy witing stte nd
th
thn the tul idle p erio dD there will e no resp onsiveE
strts up timer to tre the tul idle p erio dF he
ness delyFroweverD the energy sving will not e s
system then p erforms new predition every time
th
eetive s the originl oneF
to determine whether the system should enter the sleep
PFS gontrol mehnism of the prop osed
stteF por exmpleD in pigure Q@AD y the end of runE
metho d
ning stte of PD long idle p erio d s Q o urs t the end
H
of running stte F sf the predited idle p erio d s Q pigure S shows the nite stte mhine of the proE
is lower thn D the system stys in the usy witing p osed metho d whih onsists of three sttesX r unning D
th
stteF efter n time p erio dD the system p erforms the or r etD nd sl eepF snitillyD the system is in the running
th
HH
predition ginD resulting in s Q F sf the predited vlue stteF hen n idle p erio d o ursD the system predits
H
is lrger thn D then it enters the sleep stteF ytherE the durtion s of the up oming idle p erio dF sf the preE
th
wiseD it will p erform idle p erio d predition fter nother dited vlue is lower thn the threshold vlue D then
th
timeF the system resets the timer nd stys in the running th 1.0 1.0 exponential-average I’ /c=0 Hit Ratio exponential-average 0.6 with miss correction 0.6 exponential-average 0.4 exponential-average 0.4 On-Threshold Running 0.2 0.2 On-Threshold 0.0 0.0 c=Sth Wakeup|c=I’−W 010.001 0.01 0.1 010.001 0.01 0.1 I’>Sth/c=0 Tcost (sec) Tcost (sec) (a) X-server (c) Tin I’’ exponential-average Correct Sleep 0.8 0.8 with miss correction Hit Ratio Hit Ratio Hit Ratio I’’>Sth/c=0 0.6 exponential-average 0.6 exponential-average exponential-average with miss correction 0.4 0.4 X pinite stte mhine of the prop osed metho dF pigure S On-Threshold 0.2 On-Threshold 0.2 itingF ytherwiseD the system enters the stte s usy w 0.0 0.0 010.001 0.01 0.1 0 0.001 0.010.1 1 10 itingD the system moniE sleep stteF huring the usy w Tcost (sec) Tcost (sec) tors the timer nd enters the orret stte every of th (b) Netscape (d) Telnet timeF sn the orret stteD the system reEpredits the pigure TX gomprisons of hit rtiosF HH durtion of the idle p erio dF sf the predited vlue s is lower thn D then the system resumes the running th H t t stte s usy witingF ytherwiseD it enters the sleep hy a IHH7X @IRA t stteF sn the sleep stteD the system will return to the running stte when wkeup signl rrivesF sf the preE vet e the elpsed time of the sleep stteF he s wkeup sheme is pplied to the systemD then the system energy sving is the rtio etween the totl p ower disE will return to the running stte either when wkeup siption efore nd fter pplying the systemEshutdown signl rrives or the system hs een shutdown for sheme s H predited idle p erio d @ a s AF t Q ixp eriments i a Y @ISA H