A Gene alized Model o In es iga ing
Scheduling Schemes in Compu a ional
Clus e s
Tien V. Do a,c,∗,Binh T. Vu b, Xuan T. T an a,c
, Anh P. Nguyen a
aDepa men o Ne wo ked Sys ems and Se ices,
Budapes Uni e si y o Technology and Economics,
H-1117, Magya ud´osok k¨o ´u ja 2., Budapes , Hunga y.
bIn e -Uni e si y Cen e o Telecommunica ions and In o ma ics, Budapes
Uni e si y o Technology and Economics, 4028 Deb ecen, Kassai ´u 26., Hunga y
cVie nam In e na ional Resea ch Ins i u e o Sciences,
Ho Hoan Kiem 1428, Hanoi, Vie nam
Abs ac
In his pape , we p esen a gene alized model o he pe o mance e alua-
ion o scheduling compu e-in ensi e jobs wi h unknown se ice imes in com-
pu a ional clus e s. We p opose he applica ion o pa ame e s de ined in he
SPECpowe ssj2008 benchma k o he S anda d Pe o mance E alua ion Co po-
a ion o cons uc a pe o mance e alua ion model. In addi ion, we also de ine
a me hod o ank physical se e s based on ei he he high pe o mance p io i y
o he ene gy e iciency p io i y, and measu es o cha ac e ize he pe o mance o
compu a ional clus e s.
We in es iga e h ee schemes (sepa a e queue, class queue and common queue)
o bu e ing jobs in a compu a ional clus e ha is buil om Comme cial O -
The-Shel (COTS) se e s. Nume ical esul s show ha he bu e ing schemes do
no ha e impac on pe o mance measu es ela ed o he ene gy consump ion o
he in es iga ed clus e . Howe e , he bu e ing schemes play an impo an ole
in he quali y o se ice pa ame e s such he wai ing ime and he esponse ime
expe ienced by a i ing jobs. Fu he mo e, Dynamic Vol age and F equency Scal-
ing should be ca e ully applied i one wan s o educe he ene gy consump ion o
compu a ional clus e s.
Keywo ds: he e ogeneous clus e model, bu e ing scheme, sepa a e queue, class
queue, common queue, anking o se e s
T. V. Do e al. A Gene alized Model o In es iga ing Scheduling
Schemes in Compu a ional Clus e s. Simula ion Modell. P ac ice and Theo y,
DOI:10.1016/j.simpa .2013.05.003, 2013
Accep ed o publ. in Simula ion Modell. P ac ice and Theo y
1 In oduc ion
The ad ance o high-speed ne wo king and powe ul compu e s oge he wi h
he apid decline o ha dwa e cos s led o he widesp ead applica ion o dis-
ibu ed sys ems o o e se ices [1,2]. In such compu a ional g id sys ems,
job scheduling is mul i-c i e ia in na u e and is a i al ask in s a e-o - he-a
esou ce alloca ion s udies. The apid inc ease in he complexi y o compu a-
ional clus e s and he numbe o use s has a signi ican impac on he ene gy
consump ion, which is o be aken in o accoun in he ope a ion o g id sys-
ems.
In he li e a u e job alloca ion algo i hms a e p oposed o schedule a i ing
jobs in compu a ional clus e s. These algo i hms a e applied dominan ly a
wo le els: g id- and clus e -le el. Scheduling policies co e ing and combining
bo h o hese le els a e lis ed in [3–5] and e e ences he ein. In addi ion, some
algo i hms a e implemen ed wi h ega d o he knowledge abou cha ac e is-
ics o jobs. These may belong o ei he clai oyan [6,7] o non-clai oyan
algo i hms [8,9].
Nowadays, op imizing ene gy consump ion has become as c ucial as imp o -
ing pe o mance [10,11]. Se e al powe managemen (PM) p ac ices can be
applied: he on-o echnique comple ely shu s down he idle componen s (e.g.,
disk, CPU), while Dynamic Vol age and F equency Scaling (DVFS) [12] lowe s
he ope a ing ol age/ equency o CPU o educe he ene gy consump ion.
I is wo h men ioning ha PM is suppo ed in mode n COTS se e s. Fu -
he mo e, he ene gy consump ion o compu a ional clus e s can be educed
i he ope a o o compu a ional clus e s sys ema ically exploi s he esou ce
he e ogenei y o dis ibu ed sys ems wi h many di e en ypes o p ocesso s
o di e en powe cha ac e is ics and pe o mance capaci ies. This was he
main mo i a ion o he in es iga ions pe o med by Zikos and Ka a za [13],
whe e h ee policies applicable o clus e -le el scheduling we e compa ed:
SQEE (Sho es Queue based policy wi h Ene gy E iciency p io i y), SQHP
(Sho es Queue based policy wi h High Pe o mance p io i y), and PBP-SQ
(Pe o mance-Based P obabilis ic - Sho es Queue). Thei simula ion esul s
indica ed ha SQEE is he bes om he aspec o ene gy consump ion, SQHP
ou pe o ms he o he wo schemes a he p ice o highe ene gy consump ion,
and PBP-SQ p o ed o be he wo s among he h ee schemes. No e ha we
∗Co esponding au ho . Tel.: +36 14632070.
Email add ess: do@hi .bme.hu (Tien V. Do).
2
ha e also examined he con igu a ion s udied by Zikos and Ka a za [13] and
can con i m he conclusions o [13]. In addi ion, Te zopoulos and Ka a za [3],
Gkou ioudi and Ka a za [14] also in es iga ed he scheduling in eal- ime g id
sys ems.
In his pape , we ollow he same app oach applied in he s udy by Zikos and
Ka a za [13], whe e compu e-in ensi e jobs wi h unknown se ice imes a e
o be scheduled in a clus e wi h he e ogeneous se e s. Jobs a e execu ed
by se e s in a clus e . Compa ed o p e ious wo ks [13,3,14], his pape
p o ides a gene alized model based on pa ame e s de ined by S anda d Pe -
o mance E alua ion Co po a ion. We also de ine a me hod o ank physical
se e s based on ei he high pe o mance p io i y o ene gy e iciency p io i y,
and measu es o cha ac e ize he pe o mance o compu a ional clus e s. We
in es iga e h ee schemes (Sepa a e Queue, Class Queue and Common Queue)
o bu e ing jobs in a compu a ional clus e ha is buil om Comme cial
O -The-Shel (COTS) se e s. These p oposals allow a sys ema ic way o in-
es iga e he pe o mance o compu a ional clus e s.
Simula ion esul s show ha he bu e ing schemes do no ha e impac on
pe o mance measu es ela ed o he ene gy consump ion o he in es iga ed
clus e , bu a e signi ican ac o s ega ding he wai ing ime and he esponse
ime expe ienced by a i ing jobs. In addi ion, we also in es iga e a sa ing on
he ene gy consump ion when a speci ic se e is swi ched o i he e is no job
alloca ed o he speci ic se e and when a speci ic se e applies DVFS i a job
is alloca ed o he se e . Resul s show ha DVFS should be ca e ully applied
i one wan s o educe he ene gy consump ion o compu a ional clus e s.
The es o he pape is o ganized as ollows. In Sec ion 2, a gene alized model
along wi h me hods o ank physical se e s is p oposed. Simula ion esul s
a e p esen ed in Sec ion 3. Finally, Sec ion 4 concludes he pape .
2 A Gene alized Model and Scheduling Algo i hms
We conside a compu a ional clus e , which se es compu e-in ensi e jobs.
Following [13], we assume ha jobs
•can be execu ed on any se e ,
•a e a ended o by he Fi s Come Fi s Se ed (FCFS) se ice policy,
•a e non-p eemp ible, which means hey canno be suspended un il comple-
ion,
•ha e se ice imes unknown o he local schedule .
3
Jobs a e o be execu ed by physical se e s acco ding o a speci ic scheduling
algo i hm. In wha ollows, we desc ibe scheduling algo i hms ha alloca e
a i ing jobs o a compu a ional clus e . To make he p esen a ion comp e-
hensible, he classi ica ion o se e s is p o ided in Sec ion 2.1. Then, he
desc ip ion o scheduling algo i hms is gi en in Sec ion 2.2.
2.1 Ranking o Se e s
In a compu a ional clus e , each physical se e belongs o a speci ic se e
ype. Le Sdeno e he se o se e ypes and K=|S|be he numbe o se e
ypes. Se e ype s,s∈S, is cha ac e ized by pa ame e s Cs,Pac,s and Pid,s,
whe e Csis he ssj ops alue ( he numbe o ope a ions inished du ing he
measu emen in e al di ided by he numbe o seconds de ined o his in e -
al, showing he h oughpu –wo kload ope a ions pe second– o his pe iod
a 100% a ge load) and Pac,s deno es he a e age ac i e powe measu ed ac-
co ding o he SPECpowe ssj2008 benchma k o he S anda d Pe o mance
E alua ion Co po a ion (SPEC) a 100% a ge load (99.7% ac ual load).
We assume ha when a se e is busy, i unc ions a he ull load wi h he
powe consump ion o Pac,s. When a se e is idle, he se e s ops CPU main
in e nal clocks ia so wa e and Pid,s is he powe consump ion o a se e in
idle s a e.
We in oduce wo unc ions as ollows.
p(s) = Cs
max
i∈SCi
, s ∈S, (1)
e(s) =
Cs
Pac,s
max
i∈S
Ci
Pac,i
s∈S. (2)
I is wo h emphasizing ha Cs/Pac,s,s∈S, is he pe o mance o powe
a io o class s.
Le Spand Sedeno e he o de ed se s o se e ypes ha a e anked using
unc ion (1) and (2), espec i ely. In anking Sp( he high pe o mance p io i y
anking) and Se( he ene gy e iciency p io i y anking),
•se e ypes a e anked om 1 o |S|=Kbased on he compu ed alues
p(s) and e(s), s∈S, espec i ely;
•numbe one is assigned o se e ype a g max
s p(s) and a g max
s e(s),
espec i ely;
4
• ank K, assigned o se e ype a g min
s p(s) and a g min
s e(s), espec-
i ely.
Fo anking Sp, i he e a e wo se e ypes wi h he same ssj ops alue, he
se e ype o highe a e age ac i e powe ge s a highe index.
We o ganize physical se e s based on hei ype. Physical se e s o he same
ype o m a class. The classes a e o de ed acco ding o ei he anking Sp
(when he high pe o mance p io i y is chosen) o anking Se(when he ene gy
e iciency p io i y is p e e ed). Physical se e s a e indexed by pai (i, j),
(i= 1 ...,K;j= 1,...,M(i)). No e ha se e (1, j), j= 1,...,M(1),
has he highes p io i y and se e (K, j), j= 1,...,M(K), has he lowes
p io i y.
2.2 Scheduling
The ask o scheduling algo i hms is o alloca e a i ing jobs o physical
se e s. Scheduling algo i hms can ake in o accoun se e al ac o s such as
he pe o mance, he powe consump ion, he numbe o wai ing jobs. Fu -
he mo e, he o ganiza ion o wai ing space o jobs ha a e no immedia ely
se ed upon hei a i al is an impo an ques ion. In [13], he au ho s assumed
ha an a i ing job will wai in a speci ic physical se e a e he scheduling
decision, which is qui e s aigh o wa d om he aspec o implemen a ion. In
his pape , we call his queueing solu ion as a sepa a e queue scheme which is
illus a ed in Figu e 1. Fu he mo e, we also in es iga e wo u he schemes
o bu e ing jobs.
•Sepa a e Queue Scheme: an a i ing job will wai in a speci ic physical se e
a e he scheduling decision p esen ed in Algo i hm 1. As one obse es, he
scheduling algo i hm chooses he se e wi h he sho es queue. I he e a e
mo e idle se e s o se e s o he same queue leng h, a se e is selec ed
based on he p io i y o i s se e ype.
•Class Queue Scheme: he e is a common bu e associa ed o each class
(Figu e 2). Jobs scheduled o a speci ic class wai in he bu e o a speci ic
class when all se e s in he speci ic class a e busy. When a job depa s om
any physical se e in he speci ic class, he i s wai ing job in he bu e o
he speci ic class will be immedia ely ou ed o ha se e . This p ocedu e
is also pe o med when an a i ing job ha is ou ed o he bu e inds an
emp y se e in he speci ic class. Algo i hm 2 ou es an a i ing job based
on c i e ia: idle se e s, he p io i y and he sho es queue leng h o classes.
•Common Queue Scheme: he e is one bu e o s o ing jobs. On a job a i al,
i he LS inds all se e s busy, he job will be s o ed in he common queue
and will wai o an idle se e . A job is immedia ely se ed i Algo i hm 3
5

inds an idle se e upon i s a i al.
LS
. . . . . .
. . . . . .
. . . . . .
. . .
}
Class 1
. . .
. . . . . .
. . . . . .
. . . . . .
. . .
}
Class 2
. . . . . .
. . . . . .
. . . . . .
. . .
}
Class K
Se e queues
Fig. 1. Sepa a e Queue
LS
. . . . . .
. . .
}
Class 1
. . .
. . . . . .
. . .
}
Class 2
. . . . . .
. . .
}
Class K
Class queues
Fig. 2. Class Queue
LS
. . .
}
Class 1
. . . . . .
. . .
}
Class 2
. . .
}
Class K
Common queue
Fig. 3. Common Queue
6
Algo i hm 1 The scheduling algo i hm o Sepa a e Queue
bes se e ←(1,1)
o i= 1 →Kdo
o j= 1 →M(i)do
i se e (i, j) is FREE hen ⊲ ee se e ound in class i
bes se e ←(i, j)
GOTO SCHEDULE
else
i queue leng h o se e (i, j)<queue leng h o se e
bes se e hen
bes se e ←(i, j)
end i
end i
end o
end o
SCHEDULE: ROUTE job o queue o se e bes se e
Algo i hm 2 The ou ing algo i hm o Class Queue
bes class ←1
o i= 1 →Kdo
o j= 1 →M(i)do
i se e (i, j) is FREE hen ⊲ ee se e ound in class i
bes class ←i
GOTO SCHEDULE
else
i queue leng h o class i < queue leng h o class bes class hen
bes class ←i
end i
end i
end o
end o
SCHEDULE: ROUTE job o queue o class bes class
7
Algo i hm 3 The ou ing algo i hm o Common Queue
o i= 1 →Kdo
o j= 1 →M(i)do
i se e (i, j) is FREE hen ⊲ ee se e ound in class i
ee se e ←(i, j)
GOTO SCHEDULE
end i
end o
end o
SCHEDULE:
i ound ee se e hen
ROUTE job o queue o class ee se e
else
ROUTE job o Common Queue
end i
8
I is wo h emphasizing ha he p ac ical implemen a ion o he Sepa a e
Queue Scheme is he easies . Tha is, wai ing jobs can be placed inside each
physical se e . Fo example, jobs and pa ame e s can be alloca ed in he local
disk o each physical se e .
To implemen he Sepa a e Queue Scheme, he Class Queue Scheme and he
Common Queue Scheme we p opose a p ac ical me hod as ollows.
•A ile se e is ope a ed in a clus e . Files a e accessed using Se e Mes-
sage Block/ he Common In e ne File Sys em (SMB/CIFS) p o ocol[15] o
Ne wo k File Sys em (NFS) p o ocol [16].
•Each queue is alloca ed a sepa a e spool a ea/di ec o y in he ile se e . The
spool a eas a e accessible by he Local Schedule (LS) and he espec i e
se e s.
•The LS main ains he communica ion wi h he physical se e s in he clus e
wi h he help o a G id Compu ing F amewo k ha allows loading and
execu ing asks. The LS also has he knowledge on he in o ma ion o he
occupancy o each queue and he s a es o physical se e s in he clus e .
Based on he knowledge, he communica ion mechanism and he common
spool a eas, wai ing jobs can be easily alloca ed o se e s depending on he
applied bu e ing app oaches. The alloca ion can be pe o med quickly in he
as local ne wo k o he clus e , which minimally a ec s he pe o mance
o he clus e .
2.3 Pe o mance Measu es and Ene gy Me ics
Le wlbe he wai ing ime in queue o job lbe o e se ice and slbe he se ice
ime ha akes he se e o p ocess job l. The mean wai ing ime WT(n)
and mean se ice ime o ncomple ed jobs a e calcula ed as ollows:
WT(n) = 1
n
n
X
l=1
wl
and
ST(n) = 1
n
n
X
l=1
sl.
Response ime lo job lis he ime pe iod be ween he a i al ins an and
he depa u e ins an o job l. We ha e l=wl+sl.
The a e age esponse ime RT(n) o njobs ha inished se ice is
RT(n) = 1
n
n
X
l=1
l.
9
1300
1400
1500
1600
1700
50% 60% 70% 80% 90%
AAE(W.s/job)
U
Sepa a e-Queue
Class-Queue
Common-Queue
(a) EE policy applied
1300
1400
1500
1600
1700
50% 60% 70% 80% 90%
AAE(W.s/job)
U
Sepa a e-Queue
Class-Queue
Common-Queue
(b) HP policy applied
Fig. 10. AEswi ch−o s. sys em load
In Figu es 9 and 10, he ene gy consump ion pe job is compa ed when idle
se e s a e no swi ched o and a e swi ched o , espec i ely. I can be ob-
se ed ha he ene gy consump ion is independen o he scheduling schemes.
When idle se e s a e no swi ched o (Figu e 9), he a e age ope a ing ene -
gies o bo h policies ha e he g ea es di e ence no iceable a medium sys em
load, whe e EE consumes 1769.27 W.s/job and HP 1989.82 W.s/job. The en-
e gy consump ion pe job con e ges o 1550 W.s/job as he load inc eases.
When idle se e s a e swi ched o (Figu e 10), he HP and EE policies show
opposing endencies: su p isingly, wi h HP p io i y, he ene gy consump ion
pe job dec eases as sys em load g ows, while AEswi ched−o inc eases wi h
EE p io i y. The phenomenon can be explained by he ade-o be ween pe -
o mance and ene gy sa ing. A medium sys em load, mos job eques s a e
execu ed on se e s ope a ing wi h lowe ene gy cos , hus he ene gy dedi-
ca ed o he execu ion o jobs is signi ican ly smalle (1399 W.s/job s. 1649
W.s/job). As demand o jobs g ows, he use o high pe o mance se e s
becomes ine i able, since he policies a e based on sho es queue. In con-
sequence, he di e ence in AE be ween SQEE and SQHP dec eases a high
sys em load (1510 W.s/job s 1535 W.s/job a 90% o u iliza ion).
The impac o sa ing ene gy by swi ching o se e s is illus a ed in Figu e 11.
I is obse ed ha he sa ing dec eased as he load is inc eased.
3.2.3 DVFS
To in es iga e he impac o DVFS, we c ea e a scena io (applying Common
Queue bu e ing scheme) whe e p ocesso s educe hei equency and ol age
compa ed o he ull powe (a %100 a ge load). In his scena io, he ssj ops
alue (co esponds o 70% o he ull capaci y) and he powe consump ion o
he se e s applying DVFS when hey execu e jobs a e as ollows:
16

1000
1100
1300
1600
2000
50% 60% 70% 80% 90%
Mean sys em ene gy consump ion(W.s/job)
U
No swi ch-o
Swi ch-o
(a) EE policy applied
1000
1200
1400
1600
1800
2000
50% 60% 70% 80% 90%
Mean sys em ene gy consump ion(W.s/job)
U
No swi ch-o
Swi ch-o
(b) HP policy applied
Fig. 11. Mean ene gy consump ion s. sys em load
•Ace AW2000h-Aw170h F2 (In el Xeon E5-2670): ssj ops is 4517449 and
he powe is 1169W.
•Ace AW2000h-AW170h F2 (In el Xeon E5-2660): ssj ops is 3706521 and
he powe is 881W.
•Powe Edge R820 (In el Xeon E5-4650L): ssj ops is 1961157 and he powe
is 317W.
In Figu es 12 and 13, we plo he a e age esponse ime and he a e age ene gy
consump ion pe job s load o he DVFS scena io and he scena io (deno ed
as “no DVFS”) wi h he ull p ocessing capaci y. I is wo h emphasizing ha
a he same load alue he numbe o a i ing jobs is less in he DVFS scena io
han in he scena io a he ull p ocessing capaci y (no e ha DVFS and “no
DVFS” also swi ch o idle se e s). The impac o DVFS is qui e d as ic: he
inc ease in he a e age esponse ime is much highe han he educ ion in he
a e age ene gy consump ion pe job.
1
1.2
1.5
1.8
2.2
2.6
50% 60% 70% 80% 90%
Response Time (s)
U
DVFS+Swi ch-o
No DVFS
(a) EE policy and Common Queue
applied
1
1.2
1.5
1.8
2.2
2.6
50% 60% 70% 80% 90%
Response Time (s)
U
DVFS+Swi ch-o
No DVFS
(b) HP policy and Common Queue
applied
Fig. 12. A e age esponse ime s. sys em load
In Figu es 14, 15 and 16, we plo esul s when DVFS and “no DVFS” handle
he same numbe o jobs by keeping he same a i al a e. As an icipa ed,
he p ice o DVFS is he deg ada ion in he quali y o se ice compa ed o
a case when p ocesso s unc ion a hei ull p ocessing capaci y, which is
17
800
1000
1200
1400
1600
1800
50% 60% 70% 80% 90%
Mean Ene gy Consump ion(W.s/job)
U
DVFS+Swi ch-o
No DVFS
(a) EE policy and Common Queue
applied
800
1000
1200
1400
1600
1800
50% 60% 70% 80% 90%
Mean Ene gy Consump ion(W.s/job)
U
DVFS+Swi ch-o
No DVFS
(b) HP policy and Common Queue
applied
Fig. 13. Mean ene gy consump ion s. sys em load
clea ly obse able in Figu e 14. Howe e , he ene gy consump ion o DVFS is
highe han he non-DVFS when EE policy and Common Queue a e applied.
The impac o DVFS on he ene gy consump ion is only obse ed o he HP
policy and Common Queue bu e ing (see Figu e 16). The esul s illus a e
ha DVFS should be ca e ully uned i one wan s o apply DVFS o educe
he ene gy consump ion o compu a ional clus e s.
0.8
1.1
1.5
1.8
2.2
2.5
9.0 10.8
Response Time (s)
λ
DVFS+Swi ch-o
No DVFS
(a) EE policy and Common Queue
applied
0.8
1.1
1.5
1.8
2.2
2.5
9.0 10.8
Response Time (s)
λ
DVFS+Swi ch-o
No DVFS
(b) HP policy and Common Queue
applied
Fig. 14. A e age esponse ime s. λ
0
0.2
0.4
0.6
0.8
1
0 5 10 15 20
CDF o RT
x (seconds)
No DVFS,lambda = 9.0
DVFS+Swi ch-o ,lambda = 9.0
No DVFS,lambda = 10.8
DVFS+Swi ch-o ,lambda = 10.8
(a) EE policy and Common Queue
applied
0
0.2
0.4
0.6
0.8
1
0 5 10 15 20
CDF o RT
x (seconds)
No DVFS,lambda = 9.0
DVFS+Swi ch-o ,lambda = 9.0
No DVFS,lambda = 10.8
DVFS+Swi ch-o ,lambda = 10.8
(b) HP policy and Common Queue
applied
Fig. 15. CDF o esponse imes
18
1200
1400
1600
1800
2000
9.0 10.8
Mean Ene gy Consump ion(W.s/job)
λ
DVFS+Swi ch-o
No DVFS
(a) EE policy and Common Queue
applied
1200
1400
1600
1800
2000
9.0 10.8
Mean Ene gy Consump ion(W.s/job)
λ
DVFS+Swi ch-o
No DVFS
(b) HP policy and Common Queue
applied
Fig. 16. Mean ene gy consump ion s. λ
4 Conclusion
In his pape we p esen ed a gene alized model o he pe o mance e alua ion
o scheduling compu e-in ensi e jobs wi h unknown se ice imes in compu-
a ional clus e s. We ga e an implemen a ion o his model on h ee schemes
(Sepa a e Queue, Class Queue and Common Queue), di e ing in he way in-
coming jobs a e bu e ed, and on wo job-scheduling policies (SQEE, SQHP),
p io i izing ei he ene gy e iciency o high pe o mance. We de ined a anking
me hodology o physical se e s, which is used o schedule jobs.
Nume ical esul s show ha he bu e ing schemes do no a ec he ene gy con-
sump ion o he in es iga ed clus e s. Howe e , hey ha e a signi ican impac
on he mean esponse ime and mean wai ing ime o incoming jobs. Fu he -
mo e, he Common Queue and Class Queue schemes ma kedly ou pe o m he
Sepa a e Queue scheme. The e o e, a good bu e ing scheme can esul in im-
p o ed o e all clus e pe o mance wi hou inc eased powe consump ion and
ene gy cos . Fu he mo e, Dynamic Vol age and F equency Scaling should be
ca e ully applied o he pu pose o educing he ene gy consump ion o com-
pu a ional clus e s.
Acknowledgemen
The publica ion was suppo ed by he T´
AMOP-4.2.2.C-11/1/KONV-2012-
0001 p ojec . The p ojec has been suppo ed by he Eu opean Union, co-
inanced by he Eu opean Social Fund.
19
Re e ences
[1] I. Fos e , “The ana omy o he g id: enabling scalable i ual o ganiza ions,”
The In e na ional Jou nal o High Pe o mance Compu ing Applica ions,
ol. 15, no. 3, pp. 200–222, 2001.
[2] B. Yagoubi and Y. Slimani, “Dynamic load balancing s a egy o g id
compu ing,” T ansac ions on Enginee ing, Compu ing and Technology, ol. 13,
pp. 260–265, 2006.
[3] G. Te zopoulos and H. D. Ka a za, “Pe o mance e alua ion o a eal- ime g id
sys em using powe -sa ing capable p ocesso s,” The Jou nal o Supe compu ing,
ol. 61, no. 3, pp. 1135–1153, 2012.
[4] A. Tche nykh, J. M. Ram´ı ez, A. A e isyan, N. Kuzju in, D. G ushin, and
S. Zhuk, “Two le el job-scheduling s a egies o a compu a ional g id,” in
P oceedings o he 6 h in e na ional con e ence on Pa allel P ocessing and
Applied Ma hema ics, PPAM’05, (Be lin, Heidelbe g), pp. 774–781, Sp inge -
Ve lag, 2006.
[5] S. Zikos and H. D. Ka a za, “Resou ce alloca ion s a egies in a 2-le el
hie a chical g id sys em,” Simula ion Symposium, Annual, ol. 0, pp. 157–164,
2008.
[6] S. Zikos and H. D. Ka a za, “A clai oyan si e alloca ion policy based on se ice
demands o jobs in a compu a ional g id,” Simula ion Modelling P ac ice and
Theo y, ol. 19, no. 6, pp. 1465–1478, 2011.
[7] S. Zikos and H. D. Ka a za, “The impac o se ice demand a iabili y on
esou ce alloca ion s a egies in a g id sys em,” ACM T ans. Model. Compu .
Simul., ol. 20, pp. 19:1–19:29, No . 2010.
[8] Y. He, W. Hsu, and C. Leise son, “P o ably e icien online non-clai oyan
adap i e scheduling,” in Pa allel and Dis ibu ed P ocessing Symposium, 2007.
IPDPS 2007. IEEE In e na ional, pp. 1 –10, ma ch 2007.
[9] S. Zikos and H. D. Ka a za, “Communica ion cos e ec i e scheduling policies
o nonclai oyan jobs wi h load balancing in a g id,” Jou nal o Sys ems and
So wa e, ol. 82, no. 12, pp. 2103–2116, 2009.
[10] T. V. Do and C. Ro e , “Compa ison o scheduling schemes o on-demand
iaas eques s,” Jou nal o Sys ems and So wa e, ol. 85, no. 6, pp. 1400–1408,
2012.
[11] L. Benini, A. Bogliolo, and G. De Micheli, “A su ey o design echniques o
sys em-le el dynamic powe managemen ,” Ve y La ge Scale In eg a ion (VLSI)
Sys ems, IEEE T ansac ions on, ol. 8, pp. 299 –316, june 2000.
[12] M. Weise , B. Welch, A. Deme s, and S. Shenke , “Scheduling o educed
CPU ene gy,” in P oceedings o he 1s USENIX con e ence on Ope a ing
Sys ems Design and Implemen a ion, OSDI ’94, (Be keley, CA, USA), USENIX
Associa ion, 1994.
20
[13] S. Zikos and H. D. Ka a za, “Pe o mance and ene gy awa e clus e -le el
scheduling o compu e-in ensi e jobs wi h unknown se ice imes,” Simula ion
Modelling P ac ice and Theo y, ol. 19, no. 1, pp. 239–250, 2011.
[14] K. Gkou ioudi and H. D. Ka a za, “Mul i-c i e ia job scheduling in g id using
an accele a ed gene ic algo i hm,” J. G id Compu ., ol. 10, no. 2, pp. 311–323,
2012.
[15] C. He el, Implemen ing CIFS - The Common In e ne File Sys em. P en ice
Hall, 2003.
[16] B. Callaghan, B. Pawlowski, and P. S aubach, “NFS Ve sion 3 P o ocol
Speci ica ion.” RFC 1813 (In o ma ional), June 1995.
[17] P. Fishwick, “Simula ion oolki .” h p://www.cs.sunysb.edu/~algo i h/
implemen /simpack/implemen .sh ml.
[18] SPEC, “Ace AW2000h-Aw170h 2 (in el xeon e5-2670) machine.” h p://
www.spec.o g/powe _ssj2008/ esul s/ es2013q1/powe _
ssj2008-20121212-00590.h ml, Feb ua y 2013.
[19] SPEC, “Ace AW2000h-Aw170h 2(in el xeon e5-2660) machine.” h p://www.
spec.o g/powe _ssj2008/ esul s/ es2012q4/powe _
ssj2008-20120918-00546.h ml, Oc obe 2012.
[20] SPEC, “Powe Edge
820 (in el xeon e5-4650l) machine.” h p://www.spec.o g/powe _ssj2008/
esul s/ es2012q4/powe _ssj2008-20121113-00586.h ml, No embe 2012.
[21] S anda d Pe o mance E alua ion Co po a ion. h p://www.spec.o g/.
[22] SPEC, “ssj ops.” h p://www.spec.o g/powe /docs/SPECpowe _
ssj2008-Resul _File_Fields.h ml#Ops.
21