On Reducing Regis e P essu e and Ene gy in
Mul iple-Banked Regis e Files
Jaume Abella*, An onio González*+
*Compu e A chi ec u e Depa men
Uni e si a Poli ècnica de Ca alunya
Ba celona (Spain)
[email p o ec ed]
+In el Ba celona Resea ch Cen e
In el Labs, Uni e si a Poli ècnica de Ca alunya
Ba celona (Spain)
an oni[email p o ec ed]pc.es
Abs ac
The s o age o specula i e alues in supe scala
p ocesso s is one o he main sou ces o complexi y and
powe dissipa ion. In his pape , we p esen a no el
echnique o educe egis e equi emen s as well as hei
dynamic and s a ic powe dissipa ion ha is based on
delaying he dispa ch o ins uc ions while minimizing i s
impac on pe o mance. The p oposed echnique ou pe o ms
p e ious schemes in bo h pe o mance and powe sa ings.
Wi h only 1.77% IPC loss, he mechanism achie es mo e
han 13% dynamic and 15% s a ic ex a powe sa ings in he
in ege ename bu e s and mo e han 9% dynamic and 10%
s a ic ex a powe sa ings in he FP ename bu e s.
Signi ican powe sa ings a e also achie ed i he p ocesso
uses a physical egis e ile o bo h commi ed and non-
commi ed alues ins ead o ename bu e s. Addi ionally he
egis e equi emen s a e educed by mo e han 18% and
13% o in ege and FP p og ams espec i ely.
1. In oduc ion
Powe dissipa ion has become a c i ical issue o bo h
high pe o mance and mobile p ocesso s. Dynamic powe
dissipa ion is he dominan ac o nowadays, bu s a ic powe
will become inc easingly signi ican in upcoming p ocesso s.
While dynamic powe is di ec ly ela ed o he ac i i y o he
ci cui s, s a ic powe depends on he amoun o powe ed-on
ansis o s and hei physical cha ac e is ics. The
managemen o specula i e egis e alues is one o he main
sou ces o ene gy dissipa ion in cu en supe scala
mic op ocesso s [6]. Addi ionally, his s uc u e is one o he
p ocesso ho spo s. Thus, educing powe consump ion in
his powe hung y s uc u e is c i ical no only om he
ene gy s andpoin bu om he empe a u e s andpoin .
Some banks o hese s uc u es can be u ned-o i hey
a e no used, and hus, some powe is sa ed. Some o hem
may be also u ned-o i hey may be used wi hou
con ibu ing signi ican ly o imp o e pe o mance.
Li e a u e on powe educ ion using adap i e schemes is
e y ex ensi e. Among hem, we could poin ou some
schemes o educe powe and complexi y [1][9][10][12][16].
Some o he au ho s ha e in es iga ed how o educe he
powe and complexi y o he egis e iles. C uz e . al. [5]
p oposed a mul ile el egis e ile o ganiza ion o low
complexi y and as access ime o he egis e s. Zyuban and
Kogge [17] s udied he complexi y o a cen alized egis e
ile and p oposed a scheme o dis ibu e i . Ou p oposal is
o hogonal o hese wo ks so hey can be easily combined.
Di e en app oaches ha e been ecen ly p oposed in
o de o educe he dynamic powe o he issue queue [7][4].
Folegnani and González [7] p oposed an issue queue design
whe e ene gy consump ion is e ec i ely educed using a
dynamic esizing mechanism o he issue queue.
In his wo k, we p opose an adap i e mic oa chi ec u e
ha achie es signi ican dynamic and s a ic powe sa ings in
he egis e ile, a he expense o a e y small pe o mance
loss. Ou p oposal is based on obse ing how much ime
ins uc ions spend in he eo de bu e and he issue queue,
and aking esizing decisions based on hese obse a ions.
E en i he ins uc ions a e eady o be dispa ched, i i is
expec ed ha hey would ha dly con ibu e o imp o e
pe o mance, hey a e no dispa ched. We compa e his
scheme wi h he app oach in [7], and show ha he p oposed
echnique p o ides signi ican ad an ages.
The es o he pape is o ganized as ollows. Sec ion 2
desc ibes he baseline o ganiza ion o he issue queue,
ename bu e s, egis e ile and eo de bu e . Sec ion 3
desc ibes he p oposed echnique and he mechanism used
o compa ison pu poses. Sec ion 4 e alua es he
pe o mance o he p oposed app oach. Finally, sec ion 5
summa izes he main conclusions o his wo k.
2. Baseline mic oa chi ec u e
In his sec ion we desc ibe he baseline mic oa chi ec u e,
wi h special emphasis on he s uc u es ha a e he a ge o
his wo k: ename bu e s, egis e ile, issue queue and
eo de bu e .
2.1. P ocesso
Two di e en o ganiza ions o he s o age o specula i e
alues ha e been s udied. The i s one is simila o ha o
he Alpha 21264 [6] and Pen ium IV [14]. In his case,
specula i e and commi ed alues a e s o ed in a cen alized
egis e ile. The second one is simila o ha o he HP
PA8700 [8]. In his case, commi ed alues a e s o ed in an
P oceedings o he 21s In e na ional Con e ence on Compu e Design (ICCD’03)
1063-6404/03 $ 17.00 © 2003 IEEE
a chi ec u al egis e ile, whe eas specula i e alues a e
s o ed in ename bu e s un il commi . In ege and FP alues
a e kep in sepa a ed iles o bo h cases. The e a e wo
egis e iles o he i s o ganiza ion and wo se s o ename
bu e s o he second one. In he es o he pape he
o ganiza ion based on a cen alized egis e ile will be
e e ed o as RegF whe eas he one based on ename bu e s
will be e e ed o as RenB.
2.2. Regis e iles and ename bu e s
This sec ion desc ibes he implemen a ion assumed o
he egis e ile, bu a simila implemen a ion has been
assumed o he ename bu e s. In ege and FP egis e iles
a e iden ical. A egis e ile is spli in o banks (8 en ies pe
bank in ou expe imen s). In o de o educe he bank access
ime, he bank selec ion logic and he decoding o he en y
o be accessed a e done in pa allel.
Figu e 1 shows he scheme o a ead ope a ion. One
en y o each bank is ead, and he ou pu logic selec s he
eques ed egis e among hose. I can be obse ed ha his
scheme o e laps he bank selec ion wi h he decoding and
eading o each bank. Figu e 2 illus a es a w i e ope a ion.
The wo dlines ha selec he eques ed egis e o e e y
bank a e ga ed by he bank selec ion logic. In his case, he
bank selec ion is o e lapped wi h jus he wo dline decoding
because he w i e mus be pe o med only in he p ope bank.
Figu e 1. Scheme o a ead ope a ion
Figu e 2. Scheme o a w i e ope a ion
This implemen a ion o he egis e ile educes hei
access ime a he expense o inc easing no ably hei
dynamic ene gy consump ion. I he access ime o his
s uc u e is no c i ical, a sequen ial decoding scheme could
be conside ed.
In his wo k, he pa allel implemen a ion o he mul iple-
banked egis e ile has been assumed o all he compa ed
mechanisms, including he baseline. This decision is jus i ied
by he es ima ed access ime o bo h schemes. Fo his
pu pose, we used CACTI 3.0 [13], wi h a con igu a ion o 16
banks, 8 egis e s pe bank, 64-bi da a wid h, 0.10 Pm
echnology, 16 ead and 8 w i e po s. Table 1 shows he
delays ob ained o each componen o he egis e ile.
Table 1. Delay and ene gy o he di e en
componen s o a mul iple-banked egis e ile design
Componen Abb e . delay (ps) ene gy (pJ)
Add ess ou ing add 84 1.3
Decode (4 o 16) 4 o16 232 3.8
Decode (3 o 8) 3 o8 203 1.5 pe deco
Wo dline + bi line wlbl 134 5.1 pe bank
Da a o/ om bank da a 104 10.8 pe bank
Ou d i e ou 106 27.6
Table 2 shows he delay o he c i ical pa h o bo h a ead
and a w i e ope a ion in bo h schemes. The able shows ha
he pa allel scheme educes he access ime by 27% o ead
ope a ions and 29% o w i e ope a ions wi h espec o he
sequen ial scheme. #BanksOn co esponds o he numbe o
u ned on banks a he ope a ion ime.
Table 2. Delays and ene gy o ead/w i e ope a ions
Seq. scheme C i ical pa h Delay (ps) Ene gy (pJ)
Read add+4 o16+3 o8+wlbl
+da a+ou
863 50.1
W i e add+4 o16+3 o8+wlbl 653 22.5
Pa allel scheme C i ical pa h Delay (ps) Ene gy (pJ)
Read add+3 o8+wlbl+da a+
ou
631 32.7+17.4 x
#BanksOn
W i e 4 o16+c l_wo dlines
(=add)+wlbl
450 10.2+12.3 x
#BanksOn
Tu ning o unused banks can sa e s a ic powe o bo h
schemes and dynamic powe o he pa allel one. A gi en
bank is u ned on as soon as a leas one o i s egis e s is
assigned o an ins uc ion as i s des ina ion ope and. A gi en
bank is u ned o when none o i s egis e s is being used.
This scheme can be easily implemen ed adding a bi
(BusyBi ) o e e y egis e . This bi is se when a egis e is
assigned o an ins uc ion and is ese when he ins uc ion
commi s and ees he p e ious mapping o i s des ina ion
egis e . The bank enable/disable signal is a NOR unc ion
o i s egis e s’ BusyBi s.
In o de o maximize he numbe o banks ha a e u ned
o , when a ee egis e is eques ed, he one wi h he lowes
bank iden i ie is chosen so ha he ac i i y in he egis e
ile is concen a ed on he banks wi h lowe iden i ie s.
2.3. Issue queue and eo de bu e
The assumed eo de bu e and issue queue ha e jus one
di e ence wi h espec o con en ional ones: hei occupancy
can be limi ed dynamically. This ea u e is used o con ol
he numbe o in- ligh ins uc ions and hus, o con ol he
p essu e on he egis e iles. As p e ious wo k [7][4], no
compac ion mechanism o he issue queue has been
assumed since compac ion esul s in a signi ican amoun o
ex a ene gy consump ion e en i i con ibu es o
pe o mance.
3. Adap i e schemes
This sec ion desc ibes he p oposed mechanism and he
mechanism used o compa ison pu poses[7].
a0
a0
a1
a1
a4 a3 a2
8 banks wi h 4 en ies pe bank. Reading egis e #30 (11110b)
8 banks wi h 4 en ies pe bank. W i ing egis e #30 (11110b)
a1
a0
a1
a0
a4 a3 a2
P oceedings o he 21s In e na ional Con e ence on Compu e Design (ICCD’03)
1063-6404/03 $ 17.00 © 2003 IEEE
3.1. P oposed mechanism
3.1.1. Unde lying concep s
Supe scala p ocesso s y o keep ull bo h he eo de
bu e and he issue queue. In gene al, dispa ching
ins uc ions as soon as possible is bene icial o pe o mance,
bu no o powe . In many cases ins uc ions a e held in he
issue queue and ha e a des ina ion egis e assigned o some
cycles be o e hey a e inally issued. F om he pe o mance
s andpoin , i is desi able no o delay he issuing o any
ins uc ion. F om he powe s andpoin , i is desi able ha
ins uc ions emain in he issue queue o he minimum
numbe o cycles, since educing i s occupancy allows
educing he numbe o egis e s equi ed. Ou p oposal ies
o achie e hese objec i es by means o a ious heu is ics:
x The i s heu is ic ies o educe he ime ha ins uc ions
spend wai ing o being issued in he issue queue. I i is
obse ed ha ins uc ions wai oo long, he ins uc ion
window size ( eo de bu e size) is educed and hus, he
dispa ch o ins uc ions is delayed. Reducing he numbe
o en ies in he eo de bu e educes he numbe o
egis e s in use.
x The second heu is ic ies o p e en si ua ions in which
he limi ed ins uc ion window size is ha ming
pe o mance. E en i ins uc ions spend oo much ime in
he issue queue, i is desi able o be less agg essi e when
he e a e ew ins uc ions in he eo de bu e .
x Finally, he e a e some e en s ha equi e an immedia e
ac ion. In pa icula , L2 da a cache misses, which ha e a
e y long la ency, s all he commi o ins uc ions o
many cycles. Thus, in case o an L2 miss i is in e es ing
o inc ease he ins uc ion window size o allow he
p ocesso o p ocess mo e ins uc ions while he miss is
being se iced.
Deciding when ins uc ions spend oo long in he issue
queue is one o he icky pa s o he mechanism. We a e
in e es ed in inding ou he minimum numbe o cycles ha
he ins uc ions equi e o spend in he issue queue wi hou
losing signi ican IPC. In o de o gain some insigh , we ha e
expe imen ally obse ed he beha io o di e en p og ams
(7 benchma ks om SPEC2000) o sho in e als o ime.
I we jus conside he in e als o ime wi h simila IPC, we
can obse e some ends: a) he minimum ime ha
ins uc ions spend in he issue queue and he ime ha hey
spend in he eo de bu e a e co ela ed, b) his co ela ion
is no linea : he longe he ime in he eo de bu e , he
longe he ime in he issue queue bu he a io be ween he
la e and he o me dec eases as he ime spen in he
eo de bu e inc eases.
3.1.2. Implemen a ion o he mechanism
The i s and second heu is ics ou lined abo e a e based
on measu ing he numbe o cycles ha ins uc ions spend in
he issue queue and in he eo de bu e , as discussed in he
p e ious sec ion. Howe e , an exac compu a ion o hese
pa ame e s may be qui e expensi e in ha dwa e (e.g. ime
s amps o each en y) and consume a non-negligible amoun
o ene gy. Acco ding o Li le’s law [15] o queuing sys ems
in which a s eady-s a e dis ibu ion exis s, he ollowing
ela ion holds:
Lq =
O
Wq
whe e Lq,
O
and Wq s and o he a e age queue size, he
a e age numbe o a i als pe ime uni and he a e age
ime ha a cus ome spends in he queue. In he issue queue
and he eo de bu e , he a i al a es (
O
) a e exac ly he
same. Since we a e in e es ed in he a io be ween ime in he
issue queue and ime in he eo de bu e , ins ead o
coun ing how many cycles (Wq) e e y commi ed ins uc ion
spends in he issue queue and he eo de bu e , we will
coun how many ins uc ions (Lq) a e in hese s uc u es
e e y cycle. This app oxima ion implies ha all ins uc ions
ha a i e o he queues bu do no commi a e also coun ed.
We ha e obse ed ha he e ec o conside ing o no hese
ins uc ions does no p o ide signi ican di e ences.
We ha e expe imen ally con i med ha his ela ion
be ween queue size and wai ing ime holds o he 7
benchma ks men ioned in he p e ious sec ion. We ha e
obse ed ha he a e age numbe o cycles spen in he issue
queue and he a e age issue queue occupancy ollows a nea -
linea ela ion, and he same holds o he eo de bu e .
Thus, we can conclude ha using occupancy a ios ins ead o
ime a ios does no esul in signi ican di e ences.
In o de o le e age he ela ion be ween he ime spen in
he issue queue and he ime spen in he eo de bu e , he
p oposed mechanism uses he a io be ween bo h
occupancies (IQ occupancy / ROB occupancy) o ake
esizing decisions. I his alue is highe han a gi en
h eshold, he window size is dec eased by N ins uc ions (8
ins uc ions in ou expe imen s), and i i is lowe han
ano he h eshold, he window size is inc eased by N
ins uc ions.
Figu e 3. Heu is ic o esize he eo de bu e and he
issue queue
These h esholds a e dynamically adap ed acco ding o
he obse a ions made in he abo e sec ion, ha is, hey
depend on he eo de bu e size. Figu e 3 de ails he
app oach o esizing he eo de bu e and he issue queue.
ROB_size s ands o he physical size o he eo de bu e
(128 ins uc ions in ou e alua ion), and ROB_dyn_size
s ands o he maximum numbe o allowed ins uc ions in
he eo de bu e a a gi en ime (simila de ini ion applies
(1) THRESHOLD_LOW = 1 – ROB_dynamic_size / ROB_size
(1) THRESHOLD_HIGH = THRESHOLD_LOW + 1/8
(2) FRACTION = #ins _in_IQ / #ins _in_ROB
(3) i (FRACTION > THRESHOLD_HIGH)
(3) ROB_dyn_size = max(ROB_dyn_size–8, 32)
(3) else i (FRACTION < THRESHOLD_LOW)
(3) ROB_dyn_size = min(ROB_dyn_size+8, ROB_size)
(4) i (L2 miss du ing he pe iod)
(4) ROB_dyn_size = min(ROB_dyn_size+8, ROB_size)
(5) i (#cycles_disp_s all > IQ_THRESHOLD_HIGH)
(5) IQ_dyn_size = min(IQ_dyn_size+8, IQ_size)
(5) else i (#cycles_disp_s all < IQ_THRESHOLD_LOW)
(5) IQ_dyn_size = max(IQ_dyn_size–8, 8)
P oceedings o he 21s In e na ional Con e ence on Compu e Design (ICCD’03)
1063-6404/03 $ 17.00 © 2003 IEEE
o IQ_size and IQ_dyn_size). In o de o a oid an ex emely
small eo de bu e , he ollowing cons ain is applied:
ROB_size/4 dROB_dyn_size dROB_size. The h esholds a e
se acco ding o (1). The ac ion o ime ha ins uc ions
spend in he issue queue e sus he ime ha hey spend in
he eo de bu e is app oxima ed as (2). This pa ame e is
a e aged o each in e al o ime. A he end o each
in e al, esizing decisions a e aken acco ding o he c i e ia
desc ibed in (3): he eo de bu e dynamic size is inc eased
by 8 ins uc ions, dec eased by 8 ins uc ions, o le
unchanged depending on he alue o he FRACTION
pa ame e and he h esholds.
Finally he hi d heu is ic in he abo e sec ion is
implemen ed as ollows. Whene e he e is an L2 cache
miss, he eo de bu e size is inc eased, as (4) in igu e 3
shows. In heo y only da a misses should be conside ed bu
o he sake o simplici y, we do no dis inguish be ween
ins uc ion and da a misses since he majo i y o L2 misses
co espond o da a.
Regis e ile banks (o ename bu e s) a e u ned o
when hey a e no busy as explained in sec ion 2.2. Issue
queue occupancy is u he con olled by a mechanism ha
moni o s how many cycles he dispa ch is s alled due o
una ailable en ies in he issue queue. As de ailed in sec ion
(5) o igu e 3, i s alls a e oo equen , he issue queue size
is augmen ed (#cycles_disp_s all s ands o he numbe o
cycles ha he dispa ch is s alled because ins uc ions canno
be placed in he issue queue). I s alls a e e y a e, he issue
queue size is dec eased. This simple mechanism along wi h
he adap i e mechanism o limi he eo de bu e
occupancy achie es a signi ican egis e ile p essu e
educ ion wi h e y small pe o mance loss. In ou
expe imen s, o an in e al size o 128 cycles, di e en
alues o he issue queue h esholds we e e alua ed (2, 4, 8,
16, 32, 64) ob aining signi ican powe sa ings and small
pe o mance deg ada ion o hese pai s o alues: <16,32>
and <16,64>. To simpli y he implemen a ion and a oid
doing some di isions and mul iplica ions, in ege a i hme ic
is used ins ead o FP one. In pa icula , he h esholds a e
scaled as ollows:
THRESHOLD_LOW = ROB_size – ROB_dyn_size
THRESHOLD_HIGH = THRESHOLD_LOW + ROB_size/8
In ou expe imen s we use a 128 en y eo de bu e , so
THRESHOLD_HIGH co esponds o THRESHOLD_LOW +
16. Th esholds a e compa ed wi h FRACTION so his
pa ame e is also scaled as ollows:
FRACTION = ROB_size x #ins _in_IQ / #ins _in_ROB
The mul iplica ion in he abo e exp ession is i ial o
implemen since he eo de bu e size is a powe o 2. Fo
he di ision, he di idend has 11 bi s and he di iso has 7
bi s, assuming an in e al o 128 cycles. This equi es a
a he small ha dwa e. In ac , an i e a i e di ide can be
used ins ead o a pa allel one, since delaying he esizing
decisions by a ew cycles does no ha e any p ac ical impac .
The ene gy consump ion o he addi ional ha dwa e is
negligible because only h ee small coun e s a e upda ed
e e y cycle and he es o he s uc u es wo k only once
e e y in e al (128 cycles in ou expe imen s). Assuming
ha he di ide is implemen ed as a adix 4 di ide (2 bi s o
he quo ien a e compu ed each cycle), he o al ha dwa e
equi ed is one mul iplexo and less han 20 uni s (adde s,
inc emen e s and compa a o s) whose inpu s always ha e 11
bi s o less. We ha e expe imen ally e i ied ha delaying
he esizing o he eo de bu e by 2 o 3 cycles o allow o
an i e a i e di iso has negligible impac on pe o mance.
3.2. Mechanism used o compa ison
The p oposed mechanism has been compa ed wi h he
mechanism p oposed in [7], which will be e e ed o as
FoGo in he es o he pape . The issue queue has he same
s uc u e o bo h he p oposed mechanism and he
mechanism used o compa ison, bu he esizing schemes
a e di e en .
FoGo educes powe consump ion in he egis e iles and
ename bu e s by dynamically esizing he issue queue. The
mechanism moni o s he pe o mance con ibu ion o he
younges bank o he issue queue (8 ins uc ions in hei
expe imen s) and measu es how much hese en ies
con ibu e o he IPC. I he con ibu ion is below a
h eshold, he issue queue size is educed by one bank. On
he o he hand, he size o he queue is inc eased
pe iodically. In pa icula , his mechanism coun s he numbe
o commi ed ins uc ions ha we e issued om he 8
younges en ies in he issue queue. I he e a e less han N
ins uc ions issued om he younges pa du ing an in e al
o ime, he issue queue size is educed. Thei expe imen s
showed ha using an in e al o 1000 cycles and a h eshold
o 25 ins uc ions educes signi ican ly he issue queue
occupancy wi h a e y small pe o mance loss. E e y 5
in e als, he issue queue size is inc eased by one bank.
Fo he compa ison p esen ed below, we ha e chosen he
con igu a ion wi h he pa ame e s ha hey epo as he
mo e app op ia e ones (FoGo1000) and he same pa ame e s
bu wi h an in e al o 128 cycles – he same as he one used
by he p oposed mechanism – wi h a co esponding
h eshold o 3 ins uc ions issued om he younges pa
(FoGo128).
4. Pe o mance e alua ion
In his sec ion we p esen pe o mance and powe esul s
o he p oposed mechanism, and compa e i wi h he
echnique p oposed in [7].
4.1. Expe imen al amewo k
Powe and pe o mance esul s a e ob ained h ough
Wa ch [2], which is an a chi ec u e-le el powe and
pe o mance simula o based on SimpleScala [3]. Some
enhancemen s a e he sepa a ion o he eo de bu e and he
issue queue, and he ex ension o model he po s o he
egis e iles and ename bu e s. The model equi ed o
mul iple-banked s uc u es has been ob ained om CACTI
3.0 [13], which is a iming, powe and a ea model o banked
cache memo ies. The ollowing able desc ibes he assumed
p ocesso con igu a ion.
P oceedings o he 21s In e na ional Con e ence on Compu e Design (ICCD’03)
1063-6404/03 $ 17.00 © 2003 IEEE
Table 3. P ocesso con igu a ion
Fe ch, decode, issue, commi wid h: 8 ins uc ions
B anch p ed.: Hyb id 2K Gsha e, 2K bimodal, 1K selec o
BTB: 2048 en ies, 4-way
L1 Icache: 64KB, 2-way, 32 by e line (1 cycle)
L1 Dcache: 64KB, 4-way, 32 by e line, 4 R/W po s (2 cycles)
L2 uni ied cache: 512KB, 4-way, 64 by e line (10 cycles hi ,
50 cycles miss, 2 cycles in e chunk)
Fe ch queue: 64 en ies IQ: 80 en ies ROB: 128 en ies
RegF mic oa chi ec u e:
INT egis e s: 112 (14 banks x 8), 16R+8W po s
FP egis e s: same as INT egis e s
RenB mic oa chi ec u e:
INT ename bu e s: 80 (10 banks x 8), 16R+8W po s
FP ename bu e s: same as INT ename bu e s
INT unc ional uni s: 6 ALU, 3 mul /di
FP unc ional uni s: 4 ALU, 2 mul /di
Technology: 0.10
P
m
Fo his s udy we ha e selec ed he whole Spec2000
benchma k sui e [18] wi h he e inpu da a se . We ha e
simula ed 100 million o ins uc ions o each benchma k
a e skipping he ini ializa ion pa . The benchma ks we e
compiled wi h he Compaq/Alpha compile wi h –O4 –
non_sha ed lags.
4.2. In e al leng h
In o de o choose a sui able in e al o esize he
s uc u es, we ha e done some expe imen s. Figu e 4 shows
he IPC wi h espec o he baseline o di e en in e al
leng hs using 3 benchma ks om SpecINT2000 (gap, gzip,
wol ) and 3 om SpecFP2000 (ammp, applu, a ). I can be
seen in igu e 4 ha in gene al, longe in e als imp o e
pe o mance. Figu e 5 shows he eo de bu e occupancy
educ ion o di e en in e al leng hs. I can be obse ed
ha sho e in e als achie e highe occupancy educ ion.
Highe occupancy educ ion will ansla e in o be e
oppo uni ies o sa e powe and educe egis e p essu e.
Figu e 4. IPC o di e en in e al leng hs
Figu e 5. Reo de bu e occupancy educ ion o
di e en in e al leng hs
Figu es 4 and 5 show ha a 32-cycle in e al ha dly
educes he eo de bu e occupancy wi h espec o a 128-
cycle in e al whe eas i esul s in sligh ly highe
pe o mance deg ada ion. In addi ion, he sho e he
in e al, he highe he ene gy o e head o esizing he
s uc u es. The 512-cycle in e al is sligh ly be e in e ms
o pe o mance bu i is no so e ec i e o educe he eo de
bu e occupancy. We can conclude ha he 128-cycle
in e al achie es he bes adeo be ween powe and
pe o mance.
4.3. Pe o mance and powe esul s
The pe o mance e alua ion has been done compa ing
wo e sions o he p oposed echnique, wo e sions o
FoGo,FoGo128 and FoGo1000 as desc ibed abo e, and a
baseline wi h no adap i e s uc u es. The wo e sions o ou
echnique co espond o di e en h eshold alues o he
IQ_THRESHOLD_HIGH (32 o 64). We will e e o hem
as IqRob32 and IqRob64 espec i ely in he es o he pape .
The baseline a chi ec u e does no esize he issue queue no
he eo de bu e bu includes he mechanisms ha we ha e
assumed o IqRob and FoGo o u n o unused egis e ile
o ename bu e banks.
4.3.1. Pe o mance
Figu e 6 shows he IPC loss o he di e en mechanisms.
IqRob32 and IqRob64 ha e be e pe o mance han
FoGo1000 and FoGo128 espec i ely o he SpecINT2000
and he whole Spec2000, and achie e simila esul s o he
SpecFP2000. On a e age, IqRob32 loses less han 2% in IPC
and IqRob64 loses less han 3.5%. FoGo educes he size o
he issue queue when he IPC con ibu ion o he younges
bank is below a ixed h eshold. This h eshold basically
de e mines he loss o IPC ha he mechanism may cause
and hus, i has a bigge impac o p og ams wi h lowe IPC,
such as some o he SpecINT2000.
Figu e 6. IPC loss o di e en echniques
4.3.2. Reo de bu e and issue queue
IqRob achie es lowe eo de bu e and issue queue
occupancies han FoGo. Ha ing ewe ins uc ions in hese
s uc u es implies ha ewe egis e s a e used so mo e
powe is sa ed. IqRob is signi ican ly mo e e ec i e han
FoGo, especially o in ege applica ions.
Table 4 shows he e ec i eness o IqRob o educing he
issue queue and eo de bu e sizes. On a e age, he
IPC wi h espec o baseline
90%
92%
94%
96%
98%
100%
32 128 512
in e al leng h (cycles)
ammp
applu
a
gap
gzip
wol
Reo de bu e occupancy educ ion
0%
5%
10%
15%
20%
25%
30%
32 128 512
in e al leng h (cycles)
ammp
applu
a
gap
gzip
wol
IPC loss s. baseline
0,0%
0,5%
1,0%
1,5%
2,0%
2,5%
3,0%
3,5%
4,0%
SPECINT SPECFP SPEC
FoGo1000
FoGo128
IqRob32
IqRob64
P oceedings o he 21s In e na ional Con e ence on Compu e Design (ICCD’03)
1063-6404/03 $ 17.00 © 2003 IEEE
maximum eo de bu e size is se o abou 70% o i s o al
capaci y. Abou 45% o he en ies a e occupied and 25% o
he en ies a e enabled bu emp y. This is mainly due o
sec ions o code whe e ins uc ions spend ew cycles in he
issue queue. The IqRob mechanism ends o inc ease he
eo de bu e size in hese si ua ions because hese
ins uc ions a e qui e powe e icien ( hey do no e ain
egis e s oo many cycles).
Table 4. Size educ ion
Reo de Bu e SpecINT SpecFP Spec
IqRob32 35.4% 23.5% 29.0%
IqRob64 34.2% 22.4% 27.9%
Issue Queue SpecINT SpecFP Spec
FoGo1000 18.7% 11.8% 15.0%
FoGo128 31.2% 20.4% 25.4%
IqRob32 28.7% 20.4% 24.2%
IqRob64 34.3% 24.1% 28.8%
4.3.3. In ege egis e ile and ename bu e s
As discussed abo e, educing he numbe o in- ligh
ins uc ions esul s in a lowe numbe o egis e s in use.
IqRob achie es highe educ ions han FoGo due o i s
highe e ec i eness a educing he eo de bu e size.
FoGo1000 educes he egis e p essu e by 7%, and
FoGo128 does i by 15%, whe eas IqRob32 and IqRob64
achie e educ ions o 18% and 20% espec i ely. These
egis e p essu e educ ions a e exac ly he same o bo h
a chi ec u es (RenB and RegF) since hey ha e been
con igu ed wi h exac ly he same numbe o egis e s (80
ename bu e s + 32 logical egis e s o RenB, and 112
egis e s o RegF). Figu es 7 and 8 show ha IqRob and
FoGo achie e highe dynamic powe sa ings in he egis e
ile and ename bu e s han he baseline.
I can be seen ha highe powe sa ings a e achie ed o
he RenB a chi ec u e. The main eason is ha ename
bu e s wi h high index a e eed as soon as he ins uc ion
commi s and hus, used egis e s co espond almos always
o low-index egis e s. In his way, high-index banks can be
u ned-o in mos o he cases when he numbe o unused
egis e s is highe han he size o a bank. Fo he RegF
a chi ec u e i may happen ha a egis e wi h high index is
alloca ed o an ins uc ion and emains alloca ed o a e y
long pe iod o ime a e he ins uc ion commi s, p e en ing
he co esponding bank o be u ned o .
Figu e 7. Dynamic powe sa ings o he in ege
egis e ile and ename bu e s w. . . he baseline
Figu e 8. S a ic powe sa ings o he in ege egis e
ile and ename bu e s wi h espec o he baseline
4.3.4. Floa ing poin ename bu e s
FP ename bu e s a e ha dly used by in ege p og ams so
we epo powe s a is ics only o FP p og ams. Figu es 9
and 10 show dynamic and s a ic powe sa ings. I can be
seen ha IqRob ou pe o m FoGo in bo h dynamic and s a ic
powe educ ion o bo h a chi ec u es. Addi ionally he FP
egis e equi emen s a e educed by mo e han 13% o bo h
IqRob echniques and less han 10% o FoGo echniques.
Figu e 9. Dynamic powe sa ings o he FP egis e
ile and ename bu e s wi h espec o he baseline
Figu e 10. S a ic powe sa ings o he FP egis e ile
and ename bu e s wi h espec o he baseline
4.3.5. Summa y
Table 5 summa izes he main pe o mance me ics o he
p oposed IqRob mechanism wi h espec o he baseline
con igu a ion. No e ha he baseline al eady includes a
mechanism o u n o unused egis e ile and ename bu e
banks. The FoGo mechanism is also shown o compa ison
pu poses.
Dynamic powe sa ings s baseline
0%
5%
10%
15%
20%
SPECINT
SPECFP
SPEC
SPECINT
SPECFP
SPEC
RenB RegF
FoGo1000
FoGo128
IqRob32
IqRob64
Dynamic powe sa ings
0%
2%
4%
6%
8%
10%
12%
14%
SPECFP SPECFP
RenB RegF
FoGo1000
FoGo128
IqRob32
IqRob64
S a ic powe sa ings
0%
2%
4%
6%
8%
10%
12%
14%
SPECFP SPECFP
RenB RegF
FoGo1000
FoGo128
IqRob32
IqRob64
S a ic powe sa ings s baseline
0%
5%
10%
15%
20%
SPECINT
SPECFP
SPEC
SPECINT
SPECFP
SPEC
RenB RegF
FoGo1000
FoGo128
IqRob32
IqRob64
P oceedings o he 21s In e na ional Con e ence on Compu e Design (ICCD’03)
1063-6404/03 $ 17.00 © 2003 IEEE
DP,SP,IQ,RenB and RegF s and o dynamic powe
sa ings, s a ic powe sa ings, issue queue, ename bu e s
and egis e ile espec i ely.
Table 5. Summa y o esul s
FoGo
1000
FoGo
128
IqRob
32
IqRob
64
IPC Loss 2.1% 3.6% 1.8% 3.3%
INT Reg P essu e 7.2% 15.3% 18.1% 19.8%
FP Reg P essu e 5.8% 9.9% 13.0% 15.5%
INT RegF DP 4.1% 6.2% 8.1% 10.9%
INT RegF SP 3.2% 5.4% 9.8% 13.0%
FP RegF DP 0.8% 2.0% 6.0% 7.4%
FP RegF SP 1.0% 2.9% 7.3% 8.0%
INT RenB DP 6.0% 8.0% 13.0% 15.7%
INT RenB SP 3.9% 7.0% 14.2% 18.2%
FP RenB DP 2.0% 4.5% 9.3% 10.6%
FP RenB SP 2.7% 6.2% 10.2% 12.1%
Table 5 shows ha IqRob32 and IqRob64 ou pe o m
FoGo1000 and FoGo128 in all me ics espec i ely.
Addi ionally IqRob echniques achie e highe powe sa ings
in egis e iles and ename bu e s, and educe egis e
p essu e mo e e ec i ely han FoGo echniques.
5. Conclusions
We ha e p esen ed a no el scheme ha dynamically
limi s he numbe o in- ligh ins uc ions in o de o sa e
dynamic and s a ic powe in egis e iles o ename bu e s,
and educes egis e p essu e. The p oposed mechanism is
based on moni o ing how much ime ins uc ions spend in
bo h he issue queue and he eo de bu e and limi hei
occupancy based on hese s a is ics.
The p oposed mechanism has been e alua ed in e ms o
pe o mance, and dynamic and s a ic powe sa ings o he
whole Spec2000.
Resul s ha e been compa ed wi h a s a e-o - he-a issue
queue esizing echnique, and i has been shown ha he
p oposed echnique ou pe o ms p e ious wo k in e ms o
pe o mance and powe sa ings. The p oposed echnique
achie es mo e han 13% dynamic and 15% s a ic ex a
powe sa ings in he in ege ename bu e s (45% dynamic
and 54% s a ic o al powe sa ings wi h espec o no
u ning o banks) and mo e han 9% dynamic and 10% s a ic
ex a powe sa ings in he FP ename bu e s (23% dynamic
and 30% s a ic o al powe sa ings wi h espec o no
u ning o banks). Signi ican powe sa ings a e also
achie ed o he egis e iles i hey a e used ins ead o
ename bu e s.
Addi ionally he egis e equi emen s a e educed by
mo e han 18% o he in ege egis e s and mo e han 13%
o he FP ones.
Acknowledgemen s
This wo k has been suppo ed by CICYT p ojec
TIC2001-0995-C02-01, he Minis y o Educa ion, Cul u e
and Spo s o Spain, and In el Co po a ion. We would like o
hank he anonymous e iewe s by hei commen s.
Re e ences
[1] R.I. Baha , S. Manne. Powe and Ene gy Reduc ion Via Pipelining
Balancing. In ISCA 2001.
[2] D. B ooks, V. Tiwa i, M. Ma onosi. Wa ch: a F amewo k o
A chi ec u al-Le el Powe Analysis and Op imiza ions. In ISCA 2000.
[3] D. Bu ge and T. Aus in. The SimpleScala Tool Se , Ve sion 3.0.
Technical epo , Compu e Sciences Depa men , Uni e si y o
Wisconsin-Madison, 1999.
[4] A. Buyuk osunoglu, D. Albonesi, S. Schus e , D. B ooks, P. Bose
and P. Cook. A Ci cui Le el Implemen a ion o an Adap i e Issue
Queue o Powe -Awa e Mic op ocesso s. In GLSVLSI 2001.
[5] J.L. C uz, A. González, M. Vale o, N. Topham. Mul iple-Banked
Regis e File A chi ec u es. In ISCA 2000.
[6] J. Eme . EV8: The pos -ul ima e alpha. Keyno e a PACT 2001.
[7] D. Folegnani and A. González. Ene gy-E ec i e Issue Logic. In
ISCA 2001.
[8] D. Halpe in. PA-RISC 8x00 Family o Mic op ocesso s wi h
Focus on PA-8700. Ho Chip Con e ence, 2000.
[9] T. Ka khanis, J.E. Smi h, P. Bose. Sa ing Ene gy wi h Jus in
Time Ins uc ion Deli e y. In ISLPED 2002.
[10] S. Manne, A. Klause , D. G unwald. Pipeline Ga ing: Specula ion
Con ol o Ene gy Reduc ion. In ISCA 1998.
[11] S. Palacha la, N.P. Jouppi, J.E. Smi h. Complexi y-E ec i e
Supe scala P ocesso s. In ISCA 1997.
[12] D. Ponoma e , G. Kucuk, K. Ghose. Reducing Powe
Requi emen s o Ins uc ion Scheduling Th ough Dynamic Alloca ion o
Mul iple Da apa h Resou ces. In MICRO 2001
[13] P. Shi akuma and N.P. Jouppi. CACTI 3.0: An In eg a ed Cache
Timing, Powe and A ea Model. Resea ch epo 2001/2, WRL, Palo
Al o, CA (USA), 2001.
[14] E. Sp angle, D. Ca mean. Inc easing P ocesso Pe o mance by
Implemen ing Deepe Pipelines. In ISCA 2002.
[15] W.L. Wins on. Ope a ions Resea ch Applica ions and Algo i hms.
Ed. Duxbu y P ess. Second edi ion, 1991.
[16] S.H. Yang, M.D. Powell, B. Falsa i, T.N. Vijaykuma . Exploi ing
Choice in Resizable Cache Design o Op imize Deep-Submic on
P ocesso Ene gy-Delay. In HPCA 2002.
[17] V. Zyuban, P. Kogge. The Ene gy Complexi y o Regis e Files. In
ISLPED 1998.
[18] SPEC2000. www.specbench.o g/osg/cpu2000/
P oceedings o he 21s In e na ional Con e ence on Compu e Design (ICCD’03)
1063-6404/03 $ 17.00 © 2003 IEEE
