A T ace-Scaling Agen o Pa allel Applica ion T acingJ
Felix F ei ag. Jo di Caube , Jesus Laba a
Compu e A chi ec u e Depa men (DAC)
Eu opean Cen e o Pa allelism o Ba celona (CEPBA)
Uni e .'Ji a Poli ecnica de Ca alunya (UPC)
{ elix,jo dics,jesus}@ac. upc.es
Abs ac equi emen o ace iles. We show ha he agen can
ob ain such an unde s anding au oma ically a un ime
wi hou p og amme in e en ion o suppo .
The emainde o he pape is s uc u ed as ollows: In
sec ion 2 we desc ibe scalabili y p oblems o acing
mechanisms. Sec ion 3 shows he implemen a ion o he
ace-scaling agen . Sec ion 4 desc ibes some applica ions
and esul s o scaled acing. Sec ion 5 con ains u he
disussion o ou app oach. In sec ion 6 we conclude he
pape .
T acing and pe o mance analysis ools a e an
impo an componen in he de elopmen o high
pe o mance applica ions. T acing pa al'el p qg ams
wi h cu en acing ools. howe e . easily leads o la ge
ace iles wi h hund eds o lUegaby es. The s~o age.
isualiza ion, and analysis o such ace iles is o en
di icul .
We p opose a ace-scaling agen o acing p~ al'el
applica ions. which lea ns he applica ion behaVio in
un ime and achie es a small. easy o handle ade. The
agen dynamically iden i ies he amoun o in o b a ion
needed o cap u e he applica ion beha io . This
knowledge acqui ed a un ime allows eco ding o~ly he
non-i e a i e ace i i o ma ion, which d as ical'y !duces
he size o he ace ile.
2. Scalabili y o acing mechanisms
2.1. P oblems associa ed o la ge aces
The pe onnance analysis o pa allel p og ams easily
leads o a la ge numbe o ace iles, since o en se e al
execu ions o he ins umen ed applica ion a e ca ied ou
in o de o obse e he applica ion beha io unde sligh ly
changed condi ions. Ano he eason why se e al aces a e
needed is o s udy how he applica ion scales. All hese
aces o he di e en con igu a ions o he applica ion
and he en i onmen (numbe o p ocesso s, algo i hmic
changes, ha dwa e coun e s, ...) equi e s o age space.
Visualiza ion packages ha e di icul ies in showing
such la ge aces e ec i ely. La ge aces make he
na iga ion (zooming, o wa d/backwa d anima ion)
h ough hem e y slow and equi e he machine whe e he
isualiza ion package is un o ha e a la ge physical
memo y. O he n'ise, he esponse ime o he ool
inc eases signi ican ly, s ongly a ec ing he mo i a ion
o he p og amme o ca y ou he pe onnance analysis.
The high amoun o edundan ace in onna ion in
la ge ace iles hides he ele an de ails o he
applica ion beha io . When isualizing such la ge aces,
zooming down o iden i y he applica ion s uc u e
becomes an ine icien ask o he p og am analys . O en,
he analys needs o ha e a ce ain unde s anding o he
applica ion in o de o ca y ou an e icien pe onnance
analysis.
I. In oduc ion
Pe o mance analysis ools a e an impo an componen
o he pa allel p og am de elopmen and uning cyqle. To
ob ain he aw pe o mance da a o an applica i~n, an
ins omen ed e sion o he applica ion is un wi h p obes
ha ake measu es o speci ic e en s o pe o~ance
indica o s (i.e. ha dwa e coun e s, sub ou ines, pa allel
loops).
The ob ained ace da a can be summa ized on-line by
he acing ool. Mo e o en, howe e , i is s o ed ili1 ace
iles o o -line analysis. We ocus ou in e es in 1I acing
packages o pa allel p og ams, whe e all he ac~ui ed
da a is s o ed in ace iles o a de ailed analysis a !a la e
ime. T acing pa allel p og ams wi h such aci~ ools
easily leads o huge ace iles wi h hund ~s o
Megaby es, which has se e al p oblems conc!e ning
s o age, isualiza ion and analysis o such aces.
We p opose a ace-scaling agen o acing Qols o
pa allel applica ions. In un ime he agen lea$ he
pe iodic s uc u e in he applica ion beha io exhibi~ed by
many scien i ic p og ams. The cap ion o he appljca ion
beha io allows s o ing only he non-i e a i e ace
in o ma ion, which d as ically educes he s o age
I This wo k has been suppo ed by he Spanish Minis I)' o Science and Technology unde TIC2001 -0995-CO2-01 and by he Eu opean Union
(FEDER).
P oceedings o he 14 h IEEE In e na ional Con e ence on Tools wi h A i icial In elligence (ICTAI’02)
1082-3409/02 $17.00 © 2002 IEEE
Au ho ized licensed use limi ed o: IEEE Xplo e Cus ome . Downloaded on Oc obe 13, 2008 a 08:54 om IEEE Xplo e. Res ic ions apply.
2.2. Rela ed wo k he ace ile. The analysis o such a educed ace allows
uning he main i e a i e body o he applica ion.
3. T ace-scaling agen
3.1 Recogni ion o i e a i e pa e ns
The mos equen app oach o es ic he size o he
ace in cu en p ac ice is o inse calls in o he sou ce
code o he applica ion o s a and s op he acing.
Sys ems such as Vampi T ace [6], VGV [4 , and
OMPI ace [I] p o ide his mechanism. This a oach
equi es he modi ica ion o he sou ce code, whi h may
no always be a ailable o he pe o mance analys .E en
i he sou ce code is a ailable, i is necessa y o e a
ce ain unde s anding o i be o e being able o p ope ly
inse he acing con ol calls.
The Pa adyn p ojec [5] de eloped an ins ume a ion
echnology (Dynins ) h ough which i is poss le o
dynamically inse and ake ou p obes in a ing
p og am. Al hough no e o is made o au om ically
de ec pe iods, he me hodology behind his app oa h also
elies on he i e a i e beha io o applica ions. The
au oma ic pe iodici y de ec ion idea we p esen his
pape could be use ul inside such a dynamic analy is ool
o p esen o he use he ac ual s uc u e he
applica ion.
In IBM UTE [7], an in e media e app oach is o lowed
o pa ially ackle he p oblem, which la ge aces ose o
he analysis ool. The acing acili y can gene a huge
aces o e en s, con aining in o ma ion wi h a lo o de ail
down o he le el o con ex swi ches and global ys em
ac i i ies. Then, il e s a e used o ex ac a a e ha
ocuses on a speci ic applica ion, summ izing
in o ma ion in eco d o ma s mo e amena le o
isualiza ion and be e desc ibing he appl ca ion
beha io . To p ope ly handle he as access o s eci ic
egions o a la ge ace ile he SLOG o ma (s alable
log ile o ma ) has been adop ed. Using a ame in x he
Jmnpsho isualiza ion ool [8] imp o es he acce s ime
o ace da a.
The ace o he applica ion is a da a s eam con aining
he alues o se e al pa ame e s. I he applica ion
con ains loops, hen i has segmen s wi h pe iodic
pa e ns. We apply a pe iodici y de ec ion algo i hm o he
da a s eam in o de o segmen he da a s eam in o
pe iodic pa e ns. The used algo i hm is ame based and
equi es a ini e leng h o pas da a alues o compu e he
pe iodici y.
We implemen he pe iodici y de ec o om [3] in he
ace-scaling agen in o de o pe o m he au oma ic
de ec ion o i e a i e s uc u es in he ace. The s e3ln o
pa allel unc ion iden i ie s om he ace is he inpu . The
ou pu o he agen is he indica ion whe he pe iodici y
exis s in he da a s eam and i s pe iod leng h.
The algo i hm used by he pe iodici y de ec o is based
on he dis ance me ic gi en by he equa ion
,V-I
d(m) = sign Ii x(i)- x(i -m)1
1=0 ( 1 ).
In equa ion (I) N is he size o he da a window, m is
he delay (O<m<M), M<=N, x[i] is he cu en alue o he
da a s eam, and d(m) is he alue compu ed o de ec he
pe iodici y. I can be seen ha equa ion (1) compa es he
da a sequence wi h he da a sequence shi ed m samples.
Equa ion (1) compu es he dis ance be ween wo ec o s
o size N by sumJning he magni udes o he L I -me ic
dis ance o N ec o elemen s. The sign unc ion is used o
se he alues d(m) o 1 i he dis ance is no ze o. The
alue d(m) becomes ze o i he da a window con ains an
iden ical pe iodic pa e n wi h pe iodici y m.
I he pe iodici y m in he da a s eam is se e al
magni udes less han he size N o he da a window, hen
he alue d(m) may become ze o o mul iples o m. On
he o he hand i he pe iodici y m in he da a s eam is
la ge han he da a window size N, hen he de ec o
canno cap u e he pe iodici y. The pe iodici y leng h we
ound in he used applica ions was usually small (be ween
5 -20) and less han 300. Fo an unknown da a s eam, he
window size N o he pe iodici y de ec o can be se
ini ially o a la ge alue, in o de o be able o cap u e
po en ially la ge pe iodici ies. Once a sa is ying
pe iodici y is de ec ed, he window size can be educed
dynamically.
2.3 Ou app oach
~Ou app oach o he scalabili y p oblem o aci is o
adap dynamically he aced ime. We p opose a ace-
scaling agen , which lea ns in un ime he s uc u e o he
applica ion. I au oma ically de e mines he le an
acing in e als, which a e su icien o cap e he
applica ion beha io . Wi h he ace-scaling age i is
possible o ace only one o se e al i e a ions he
dynamically de ec ed epe i i e pa e n in he appl ca ion
beha io . Ou app oach does no equi e limi " 9 he
g anula i y o acing, no he numbe o pa ame e s ead
a e e y acing poin , no he p oblem size. Due o he
dynamic in e cep ion o he calls o un ime lib a ie in he
acing ool, ou implemen a ion does no equi e he
sou ce code o he applica ion o achie e he scal ace.
In un ime he edundan ace in o ma ion is ide i ied
and only he non-i e a i e applica ion beha io is s ed in
P oceedings o he 14 h IEEE In e na ional Con e ence on Tools wi h A i icial In elligence (ICTAI’02)
1082-3409/02 $17.00 © 2002 IEEE
Au ho ized licensed use limi ed o: IEEE Xplo e Cus ome . Downloaded on Oc obe 13, 2008 a 08:54 om IEEE Xplo e. Res ic ions apply.
3.2. Implemen a ion The new sample o e w i es he column con aining he
oldes alues wi h he new dis ance. The implemen a ion
wi h ci cula lis s a oids mo ing he da a alues. The
numbe o ope a ions a ins an i a e educed, which leads
o a small o e head o his implemen a ion.
3.3. OpenMP and acing ool in eg a ion
The s uc u e o OpenMP based pa allel applica ions
usually i e a es o e se e al pa allel egions. Fo each
pa allel di ec i e he mas e h ead in okes a un ime
lib a y passing as a gumen he add ess o he ou lined
ou ine. The acing mechanism in e cep s he call and
ob ains a s eam o pa allel unc ion iden i ie s. This
s eam con ains all execu ed pa allel unc ions o he
applica ion, bo h in pe iodic and non-pe iodic pa allel
egions.
We ha e implemen ed he ace-scaling agen in he
OMPI ace acing ool [I]. The acing ool gene a es
ace iles, which consis o e en s (ha dwa e coun e
alues, pa allel egions en y/exi , use unc ions
en y/exi ) and h ead s a es (compu ing, idle, o k/join).
ull ace
da a s eam
i e a i e pa e ns
scaled ace iLo ali.o bcha.;o DO l aoc i. w iLIOD
11111111 IIIIIIIIIIIIIIIIII 111111111 ~
Figu e I. In e ac ion o he agen in he acing
ool.
In Figu e 1 he in e ac ion be ween he ins umen ed
applica ion, he acing ool, and he agen is illus a ed. I
can be seen ha he agen ecei es a da a s eam om he
acing ool. The da a s e~ con ains he alues o a
aced pa ame e such as he iden i ie s o he execu ed
unc ions in pa allel egions. The agen lea ns he
applica ion beha io . Ha ing his indica ion he acing
ool knows, which is he non-i e a i e in o ma ion o w i e
o he ace ile.
In ou implemen a ion o equa ion ( I ), we s o e ~ ini e
numbe o p e ious da a alues o he da a Is eam
including he mos ecen alue in a da a ec o . ~ n his
da a ec o he algo i hm pe o ms pe iodici y de c ion.
This da a ec o can be implemen ed as a FIFO b e o
leng h M+N. This ype o implemen a ion uses ~ leas
amoun o memo y, bu equi es a highe numbe o
ope a ions a e e y ins an i han o he implemen a ions.
Applying equa ion (I) on he da a ec o equi ,s M x
N ope a ions o compu e he alues o d(m) a he ins an i
o he da a s eam. I can be obse ed, howe e , ha some
ope a ions a e done wi h he same da a alues ~e e al imes a di e en ins an s i. The p e iously co pu ed
dis ances be ween ec o elemen s could be s o d o
educe he numbe o compu a ions made by he algo i hm
a ins an i. The e is a ade-o be ween he nwnbe o
compu a ions made a ins an i. and he amoun o memo y
needed by he algo i hm. In o de o educe he amiun o
compu a ion we implemen a FIFO o ganized ma ix o
size MxN whe e p e iously compu ed dis anc s a e
s o ed. Using his dis ance ma ix we compu e ~ each
ins an i he alue o di(m) o all alues o m, whd; e x(i)
is he alue o he da a alue a he cu en samplel i, and
x(i-m) is he da a alue ob ained m samples be o b. The
compu ed alues o d;(m) a e w i en in he 9olumn
co esponding o he ins an i in he ma ix. i
In case o using he dis ance ma ix o s o e p eJiously
compu ed dis ances, hen only M ins ead o j x N
ope a ions need o be made a ins an i o ob ain e new
dis ances di(m). The alue o d(m) o all alue m is
ob ained by summing he elemen s o each aw 1o he
dis ance ma ix, i.e. he p e iously compu ed di ances
and each mos ecen dis ance d;(m). This means ha a
e e y ins an i new alues a e w i en in a columnlo he
dis ance ma ix, and d(m) is compu ed as he sum lo he
alues in each aw using he p e iously compu ed ~alues
o he o he columns. Using he dis ance ma ix he l ize o
he da a ec o can be educed o M, since a ins an i only
he dis ances be ween he ne~l da a alue and he M I pas
da a alues need o be compu ed. I can be s en in
equa ion ( I) ha i he da a window size N and h delay
M is inc eased, la ge i e a i e s uc u es can be de ec ed.
Then, he numbe o compu a ions o ob ain d( ) also
inc eases. Howe e , when inc easing N and M a d he
p e iously compu ed dis ances a e e-used, h n he
inc ease o ope a ions is only linea . i
In o de o educe he numbe o shi s o hel FIFO
ope a ions, he da a s uc u es o he pe iodici y de ec o
concep ually wo king as FIFO o ganized ma ix and FIFO
o ganized da a ec o a e p og ammed as ci cula lil s. A
each ins an i he poin e o he cu en lis elemen shi s
by one such ha i poin s o he oldes alues in e lis .
P oceedings o he 14 h IEEE In e na ional Con e ence on Tools wi h A i icial In elligence (ICTAI’02)
1082-3409/02 $17.00 © 2002 IEEE
Au ho ized licensed use limi ed o: IEEE Xplo e Cus ome . Downloaded on Oc obe 13, 2008 a 08:54 om IEEE Xplo e. Res ic ions apply.
4. Applica ions o scaled acing 4.3. Imp o emen o he ease o isualiza ion
4.1. Expe imen al se up
We ace he applica ions gi en in Table 1 i Fou
applica ions om he NAS benchma k sui e: E (cl~ss A),
I
Lu (class A), Cg (class A) and Sp (class A); aI1d i e
applica ions o he SPEC95 sui e: Swim, Hyd o2dj Apsi,
Tomca , and Tu b3d, all wi h e da a se . !
All expe imen s a e ca ied ou on a Silicon G~aphics
O igin 2000. The OpenMP applica ions a e execu ~d in a
dedica ed en i onmen wi h 8 CPUs. We con igq e he
ace-scaling agen such ha a e ha ing de ec,ed 10
i e a i e pa allel egions i s ops w i ing ace da al o he
ile un il i obse es a new p og am beha io~. The
pa ame e s con ained in he ace ile a e he h eaq s a es
and OpenMP e en s, which include wo ha dwa e
coun e s.
Conside ing he ull ace in Figu e 4 (see mal page) a
i s isual pe cep ion o he p og am beha io can be
qui e misleading. Fo example, i seems ha he e is a lo
o o ~join ac i i y in he i s h ead (whi e colo ) while
his is only an e ec o he display p ecision. The eason
is ha a he scale ha had o be used o display he whole
ace, each pixel ep esen s a la ge ime in e al (152 ms)
wi hin which one h ead can pe o m many changes o
ac i i y.
In Figu e 5 (see inal page) we can easily iden i y ha
he e is a pe iodic pa e n (pe iod bounda ies agged wi h
lags). I can be obse ed ha a e a ce ain numbe o
epe i ions his pa e n changes and ha a new pe iodic
pa e n is hen epea ed. The di ec look a he ull ace o
Figu e 4 ha dly e eals ha he e is a special beha io in
he middle pa . The lags in Figu e 5 iden i y he pe iod.
Wi h he scaled ace i is immedia e o zoom o an
adequa e le el o see he ac ual pa e n o beha io . In he
isualiza ion o he scaled ace, he i e a i e ace
in o ma ion is no shown (Figu e 5 black a ea), since he
acing mechanism did no w i e i o he ace ile.
Table I. E alua ed benchma ks.
Benchma ks ---
Applica ion
NAS E
NAS Cg
NAS Lu
NAS Sp
Apsi
Hyd o2D
Swim
Tomca
Tu b3d
NAS
benchma ks 4.4. Reduc ion o he ace ile size
We examine how much he mce ile size educes when
using he ace-scaling agen . Figu e 2 shows he size o
he ace iles o he NAS and SPEC95 benchma ks
ob ained wi h and wi hou using he agen . I can be seen
ha wi h scalable acing he ace iles a e educed
signi ican ly. The NAS Lu ace ile, o ins ance, educes
om I73 Mb o 8 Mb, which is a educ ion o 95%. Had
we aced less han IO i e a ions, he ace size would
educe mo e.
SPEC95 p
benchma ks
4.2. Applica ion s uc u e iden i ica ion
-
.0
~
-
. .
-0)
c:
~
Q)
u
~
...
The ace-scaling agen allows inse ing in o$a ion
abou he de ec ed applica ion s uc u e in hei ace
eco ds, which indica es he s a /end o an i ~ a i e
pa e n. This in onna ion is highly use ul o he I/nalys
because one o he i s ac i i ies when acing a la g ace
is o zoom down, ying o iden i y an a ea o !a ew
pe iods ha can be aken as e e ence o loo~ng a
de ails. The acing ool w i es hese e en s indica ing
pe iodic pa e ns o he ace e en i he w i ing ! o all
o he ace in o ma ion is suspended.
In Figu e 3 (see mal page) wo i e a i e egionsi o he
NAS E benchma k wi h hei h ead s a es a e shown.
The bounda ies o he i e a i e egions a e ep esen ed as
lags, which e eal he applica ion s uc u e. The numbe
o pe iodic pa e ns and hei du a ion can easJly be
comDu ed om he De iodici e en in he ace ilei
P oceedings o he 14 h IEEE In e na ional Con e ence on Tools wi h A i icial In elligence (ICTAI’02)
1082-3409/02 $17.00 © 2002 IEEE
Au ho ized licensed use limi ed o: IEEE Xplo e Cus ome . Downloaded on Oc obe 13, 2008 a 08:54 om IEEE Xplo e. Res ic ions apply.
e icien analysis. We ha e p oposed a ace-scaling agen ,
which allows s o ing da a o a comple e analysis while
achie ing a small ace ile. We ha e implemen ed he
agen , which lea ns he applica ion beha io in un ime
and allows s o ing only he non-i e a i e ace da a. We
ha e shown ha he size o such a scaled ace ile
becomes signi ican ly educed, while in he aced in e al
he ele an applica ion beha io is cap u ed. We
obse ed ha he scaled aces a e easy o handle by
isualiza ion ools and he scaled ace le s he analys
us e obse e ele an applica ion beha io such as he
applica ion s uc u e. Ou implemen a ion o he ace-
scaling agen has a small o e head and i is used in
un ime. The scaled ace can subs i u e he ull ace in
se e al pe o mance analysis asks, since i allows he
pe o mance analys o each he same conclusions on he
applica ion pe o mance as when using he ull ace.
-
~.Full ace
.10l e a io s
--
.c
~
.c
-C)
c
~
Qj
~
..
...
NAS lu Hyd o2D
Figu e 2. Compa ison o he ace ile size wi~h ull
and scalable acing.
5. Discussion
7. Re e ences
The o e head o an implemen a ion is an impo an
pe o mance ac o o eal- ime ools. In [2] w~ ha e
e alua ed he o e head p oduced by he ace-$caling
agen . I was obse ed ha he o e head in odu4ed by
acing is small in e ms o he execu ion im The o iginal acing ool adds 1% -3% o he execu io ime.
Wi h he ace-scaling agen , he o e head is 3% -6 0.
In applica ions wi h a pe iodic pa e n ",'e ex ec o
each he same concluysions on pe o mance I when
analysing a subse o he i e a ions, i.e. he scaled hce. In
[2] we ha e compa ed he pe o mance indices coTpu ed
om he scaled and ull aces. Ou esul s show ~ he
same pe o mance conclusions can be ob ained I when
analysing he scaled ace o he applica ions. i
The agen lea ns he applica ion beha io he s eam o unc ion iden i ie s. I could be possible he
agen de ec s i e a i e beha io in he execu ed un ions,
bu a he same ime he pe o mance o he o he i dices
(cache misses, TLB misses, ...) could di e signi can ly
om one i e a ion o ano he . I such a case occu ~ in an
isola ed pa allel egion, he agen would no de e~ his
si ua ion. I
Ou ool elies on he i e a i e beha io O apPli ions, whe e loops a e execu ed many imes. Many sc. n i ic
applica ions ha e such a s uc u e. The s udied N S and
SPEC95 benchma ks, which mos ly pe o m n e ic
compu a ions, exhibi i e a i e applica ion beha ~o . In
case o ha ing he ace-scaling agen ac i a e4 wi h
ano he class o applica ions, which a e non-i e a i e,
simply no pe iodic beha io would be de ec ed and he
whole ace would be w i en o he ile.
[1] J. Caube , J. Gimenez, J. Laba a, L. DeRose, J.
Ve e . "A Dynamic T acing Mechanism o Pe o mance
Analysis o OpenMP Applica ions." In In e na iona/
Wo kshop on Open, 1P App/ica ions and Too/s
(WO, 1PAT 2001), July 2001, pp. 53-67.
[2] J. Caube , F. F ei ag, J. Laba a. "Compa ison o
scaled and ull aces o OpenMP applica ions." Tech.
Repo UPC-DAC-200 1-31.
[3] F. F ei ag, J. Co balan, J. Laba a. "A dynamic
pe iodici y de ec o : Applica ion o speedup
compu a ion." In In e ma iona/ Pa a//e/ and Dis ibu ed
P ocessing Symposium (IPDPS2001), Ap i12001.
[4] J. Hoe linge , B, Kuhn, W. Nagel, P. Pe e sen, H.
Rajic, S. Shah, J. Ve e , M. Voss, and R. Woo. An
In eg a ed Pe o mance Visualize o MPl/OpenMP
P og ams. In In e na iona/ Wo kshop on Open, 1P
Applica iom and Tools (WO, 1PAT2001), July 2001, pp.
40-52.
[5] B. P, Mille , M. D. Callaghan. The Pa adyn Pa allel
Pe o mance Measu emen Tools. IEEE Compu e
28(11): 37-46, No embe 1995.
[6] Pallas: Vampi ace. Ins al/a ion and Use 's Guide,
h m:llwww,12allas.de
[7] C. E. Wu, A, Bolma cich, M. Sni , D. Woo on, F.
Pa pia, A. Chen, E. Lusk and w, G opp. F om T ace
Gene a ion o Visualiza ion: A Pe o mance F amewo k
o Dis ibu ed Pa allel Sys ems. In P oceedings o
Supe Compu ing (SC 2000), No embe 2000.
[8] 0. Zaki, Eo Lusk, W, G opp, and D. Swide , Towa d
scalable pe o mance isualiza ion wi h Jwnpsho . In
In e na iona/ Jou na/ o High Pe o mance Compu ing
App/ica iom. 13(2): pages 277-288, 1999.
6. Conclusions
We ha e desc ibed he some scalabili y p obl~ms o
acing in cun-en pe o mance analysis ools an~ why
hese a e a p oblem o s o age, isualiza io~, and
P oceedings o he 14 h IEEE In e na ional Con e ence on Tools wi h A i icial In elligence (ICTAI’02)
1082-3409/02 $17.00 © 2002 IEEE
Au ho ized licensed use limi ed o: IEEE Xplo e Cus ome . Downloaded on Oc obe 13, 2008 a 08:54 om IEEE Xplo e. Res ic ions apply.

Figu e 3. Visualiza ion o he h ead s a es in he NAS B applica ion
in 2 i e a i e pa allel egions. Ligh colo =idle, da k colo =compu ing.
Figu e 4. Visualiza ion o he whole Hyd o2D lexecu ion ace ( ull ace).
Figu e 5. Visualiza ion o he Hyd o2D execu ion wi h scaled acing (scaled ace).
P oceedings o he 14 h IEEE In e na ional Con e ence on Tools wi h A i icial In elligence (ICTAI’02)
1082-3409/02 $17.00 © 2002 IEEE
Au ho ized licensed use limi ed o: IEEE Xplo e Cus ome . Downloaded on Oc obe 13, 2008 a 08:54 om IEEE Xplo e. Res ic ions apply.