Co
m
pu
ter Sci
ence a
nd Inf
or
mat
i
on
Tec
h
no
lo
gies
Vo
l.
1
, No
.
2
,
J
ul
y
2020
,
pp.
39
~
46
IS
S
N:
27
22
-
3221
,
DOI: 10
.11
591
/
csi
t.v
1i
2
.p3
9
-
46
39
Journ
al h
om
e
page
:
http:
//
ia
esprime
.com/i
ndex.
php/csit
Effici
ent
and
s
ca
l
able mul
titenant
p
lacemen
t a
pp
ro
ac
h
for i
n
-
memo
ry d
ata
b
ase over su
pp
le architectu
re
Arpit
a
S
hah
1
,
Na
re
ndr
a
P
atel
2
1
Facul
t
y
of Tec
h
nolog
y
and
Enginee
ring
,
Ch
aro
tar
Univer
sit
y
of
S
ci
en
ce
a
nd
Tech
nolog
y
(CHA
RUS
AT),
C
hanga
-
388421,
Gujar
a
t, India
2
Depa
rtment of
Com
pute
r
Engi
n
ee
ring
,
B
irl
a
Vishvaka
rm
a
Mah
a
vid
y
ala
y
a Engi
n
ee
ring
Coll
ege
-
GTU,
Vallabh
Vid
y
an
aga
r
-
388
120,
Gujar
at,
In
dia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
hist
or
y:
Re
cei
ved
Ja
n
23
, 2
0
2
0
Re
vised
Ma
y
1
, 20
2
0
Accepte
d
Ma
y
2
0
, 20
2
0
Of
la
t
e
Multi
te
n
ant
m
odel
wi
th
In
-
Mem
or
y
da
tabase
has
b
ec
om
e
prom
ine
n
t
are
a
for
rese
arch.
Th
e
pap
er
ha
s
used
adva
nt
ag
es
of
m
ult
itenan
c
y
to
r
educe
the
cost
for
har
dware
,
l
abor
an
d
m
ake
av
ai
l
ability
of
storag
e
b
y
sharin
g
dat
ab
ase
m
emory
and
fi
le
exec
uti
on.
The
purp
ose
of
thi
s
p
aper
is
to
g
ive
over
vie
w
of
pr
oposed
Supple
arc
hi
te
c
ture
for
implement
ing
in
-
m
emor
y
dat
ab
ase
b
ac
ken
d
and
m
ultite
n
a
nc
y
,
appl
i
ca
b
le
in
public
and
pr
iva
t
e
c
loud
se
tt
ings.
Ba
cke
n
d
in
-
m
emor
y
da
ta
base
uses
col
u
m
n
-
orie
nte
d
app
roa
ch
wit
h
dic
ti
on
ar
y
b
ase
d
compress
ion
te
chn
ique.
W
e
used
ded
icat
ed
sam
ple
benc
hm
ark
for
th
e
workload
proc
essing
and
al
so
a
dopt
the
SLA
pe
nal
t
y
m
ode
l
.
In
par
ticul
ar,
we
pre
sent
two
appr
oximati
o
n
a
lgorithm
s,
Multi
-
te
n
a
n
t
pla
c
ement
(MTP)
and
B
est
-
fi
t
Gree
d
y
to
show
the
qua
li
t
y
of
te
n
ant
pl
acem
ent.
The
expe
r
imental
resul
ts
show
th
at
Mul
ti
-
t
ena
nt
p
la
c
ement
(MTP)
al
gor
it
hm
is
sca
la
b
le
and
e
ffic
i
ent
in
compari
son
with
B
est
-
fit
Gre
ed
y
Alg
orit
hm
over
propo
sed
ar
chi
t
e
ct
ure
.
Ke
yw
or
d
s
:
Be
st
-
fit g
reed
y
algorit
hm
,
In
-
m
e
m
or
y database,
MTP (m
ulti
-
te
nan
t
placem
ent),
Mult
it
enan
cy
,
Supp
le
a
rc
hitec
ture
This
is an
open
acc
ess arti
cl
e
un
der
the
CC
BY
-
SA
l
ic
ense
.
Corres
pond
in
g
Aut
h
or
:
Arpit
a S
hah
,
Faculty
of Tec
hnology a
nd E
ng
i
neer
i
ng,
Char
otar U
nive
rsity
o
f
Scien
ce
a
nd Tec
hnol
og
y
(CH
ARU
S
AT),
Chan
ga
-
3884
21, G
uj
a
rat,
Indi
a
.
Em
a
il
:
arp
it
ashah
.ce
@ch
a
r
us
a
t.ac.i
n
1.
INTROD
U
CTION
Conve
ntion
al
l
y,
in
-
m
e
m
or
y
databases
hav
e
bee
n
i
n
us
e
f
or
ap
plica
ti
on
s
w
hich
we
re
pe
rfor
m
ance
sensiti
ve
su
c
h
as
fi
na
ncial
ser
vices
m
ark
et
s.
In
-
m
e
m
or
y
database
cl
aim
to
pro
vid
e
an
al
te
rn
at
ive
to
t
he
OLAP
.
In
ste
a
d
of
pull
ing
the
data
from
a
dis
k,
kee
ping
it
i
n
m
e
m
or
y
(R
AM)
s
pe
eds
up
the
processin
g
a
nd
r
esp
on
se
tim
e
of
data
by
orde
r
of
m
agni
tud
e.
This
is
t
he
reas
on
w
hy
in
-
m
e
m
or
y
Da
ta
base
is
boom
ing
i
n
i
ndus
try
these
days.
With
t
he
ex
ped
it
io
us
increase
of
S
oft
war
e
-
as
-
a
-
Se
rv
ic
e
(S
aaS
),
it
has
bec
om
e
i
m
po
rtant
t
o
operat
e
serv
ic
es
at
a
fa
ste
r
res
pons
e
ti
m
e
fo
r
SaaS
pr
ov
i
der
s
.
W
it
h
t
he
ai
m
of
to
re
du
ce
oper
at
iona
l
cost,
m
ulti
-
t
enan
cy
pro
vid
es
m
et
ho
ds
f
or
c
om
bi
ning
m
ulti
ple
te
nan
ts
of
hoste
d
a
pp
li
cat
io
n
i
nto
t
he
sam
e
syst
e
m
.
Mult
i
t
enan
c
y
can
be
em
plo
ye
d
in
the d
at
ab
a
se
la
ye
r
in
suc
h
a
way
that
a
sing
le
data
bas
e
can
be
us
e
d
by
m
ult
iple
custom
ers
i.e.
te
na
nts.
A
cl
oud
us
es
te
c
hnology
of
m
ulti
te
nan
cy
to
s
har
e
I
T
res
our
ces
am
on
g
m
ulti
ple
app
li
cat
ion
s
a
nd
te
nan
ts
sec
ur
el
y.
Virtuali
zat
io
n
-
base
d
arc
hitec
tures
is
us
ed
by
so
m
e
cl
ouds
to
is
olate
te
nan
ts
a
nd
so
m
e
us
es
custom
so
ftwa
re
arc
hitec
ture
s
to
get
the
job
done
.
I
n
t
his
pa
per
we
ha
ve
s
how
n
the
pro
pose
d
arc
hitec
tu
re
f
or
sta
nd
i
ng
te
nan
t
placem
ent
fo
r
qu
e
ry
re
qu
est
w
it
h
sam
ple
HR
ben
c
hm
ark
de
sign
c
om
bin
ed
both
a
ppr
oac
hes
in
me
m
or
y
an
d
m
ulti
te
nan
cy
.
To
im
pr
ov
e
se
ver
util
iz
at
ion
an
d
res
ource
prof
i
t,
te
ste
d
tw
o
al
gorithm
(s)
(1)
Be
st
Fit
Gr
ee
dy
(
2)
MTP.
I
n
S
up
ple
arc
hitec
tur
e
it
con
sist
s
m
ai
nly
three
c
om
po
nen
ts
(
1)
Cl
us
te
r
hea
d:
m
ai
ntain
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2722
-
3221
Com
pu
t. Sci.
I
nf. Tec
hnol.
,
V
ol.
1
, N
o.
2
,
J
ul
y
20
20
:
39
–
46
40
placem
ent
inform
ation
ov
e
r
i
n
-
m
e
m
or
y
databas
e.
(
2)
R
oute
r:
base
d
on
cl
us
te
r
m
ap
it
forw
a
rd
s
que
ry
r
equ
e
st
to
the
s
uitable
instance
m
anag
er
an
d
(
3)
I
ns
ta
nce
Ma
na
ger
:
distrib
utio
n
of
re
quest
s
acro
s
s
the
te
na
nt
us
e
r
.
The
s
upple
a
rc
hitec
ture
a
dopt
Mi
cro
s
of
t
Az
ur
e
Plat
form
and
al
so
pro
vide
physi
cal
m
ac
h
ine
t
hat
tu
rn
e
d
int
o
virtu
al
disk p
ool.
2.
IN
-
ME
MO
R
Y DA
T
ABA
S
E
Earli
er
in
-
m
em
or
y
databases
hav
e
bee
n
use
d
in
the
perf
orm
ance
sensiti
ve
ap
plica
ti
on
s
wh
ic
h
we
re
perform
ance
sensiti
ve
li
ke
fin
ancial
serv
ic
es
m
ark
et
s.
H
owe
ver
now
a
da
ys
in
-
m
e
m
or
y
databases
hav
e
be
com
e
m
or
e
ge
ner
al
ly
ad
opte
d.
At
th
e
sam
e
tim
e,
t
he
Softwa
re
-
as
-
a
-
Se
r
vice
(
Sa
aS)
m
od
el
has
beco
m
e
popula
r
a
nd
custom
ers
are
gai
ning
i
nter
est
in
t
his
m
od
el
,
as
it
decre
ases
thei
r
th
e
bur
de
n
of
th
e
has
sle
of
op
erati
ng
the
syst
em
,
w
hich
re
qu
i
res
pro
vision
i
n
g
t
he
hard
war
e
a
s
well
as
m
ain
ta
inin
g,
ope
r
at
ing
a
nd
c
onf
igurin
g
app
li
cat
io
n
se
r
ver
s
a
nd
data
ba
ses.
T
he
key d
iffe
re
nce b
et
ween
in
-
m
e
m
o
ry
data
base (
I
MDB)
a
nd d
is
k
reside
nt
database
(D
R
DB)
is
that
da
ta
in
IMDB
re
sides
in
m
ai
n
m
e
m
or
y
per
m
anen
tl
y
an
d
in
DRDB,
data
reside
s
per
m
anen
tl
y o
n
dis
k.
Si
nce t
he
chi
p
de
ns
it
y i
s incr
easi
ng
da
y by day an
d sem
ic
on
duct
or
m
e
m
or
y i
s b
ecom
ing
cheap
e
r
s
o
it
’s
possible
t
o
st
ore
huge
am
ount
of
data
i
n
m
e
m
or
y
[1
]
.
IM
D
Bs
can
pro
vid
e
pe
rfor
m
ance
by
an
order
of
m
agn
i
tud
e
fa
ste
r
as
c
om
par
ed
to
D
RDBs.
In
-
m
em
or
y
wh
ic
h
is
on
e
of
t
he
m
e
m
or
y
residen
t
syst
e
m
s
and
c
om
par
es
it
s
pr
oc
essin
g
tim
e
with
a
typ
ic
al
disc
re
sident
database
.
Of
c
ourse
t
he
in
-
m
e
m
or
y
database
syst
e
m
(I
MDB
)
ga
ve
bette
r
pe
rfor
m
ance
an
d
respo
ns
e
ti
m
e
.
Com
plexity
in
in
-
m
e
m
or
y
databases
is
reduc
ed
as
the
num
ber
of
m
achine
instr
uc
ti
on
s
re
duced
,
buff
e
r
po
ol
m
a
nag
em
ent
is
not
required,
an
d
no
nee
d
of
ext
r
a
data
cop
ie
s
,
i
nd
e
x
pa
ges
dec
reases
,
an
d
their
sim
plifie
d
str
uctu
re
is
possi
ble.
As
a
co
ns
e
quence
the
desig
n
bec
om
es
si
m
ple
and
m
or
e
c
onci
se,
a
nd
re
qu
est
s
a
r
e
perform
ed
f
ast
er.
IM
DS
s
hav
e
al
so
lo
w
er
m
e
m
or
y
and
CP
U
requirem
ents.
Also
the
desig
n a
nd
a
rr
a
nge
m
ent
of
data
on
dis
k i
s
m
uch
m
or
e
unfa
vora
ble
tha
n
ar
ra
nge
m
ent
of
data
on
m
ai
n
m
e
m
or
y.
Re
al
tim
e
app
li
cat
ion
s
re
qu
ir
e
fast
respo
n
se.
S
o
I
MDBs
play
cr
ucial
r
ole
for
r
eal
tim
e
app
li
cat
io
ns
.
“
Ca
n
entire
dat
abase
fit
into
m
ai
n
m
e
m
or
y
?
”
The
a
nswer
dep
e
nds
on
t
he
app
li
cat
io
n.
If
siz
e
of
app
li
cat
io
n
is
li
m
it
ed
an
d
is
gr
ow
i
ng
at
slo
we
r
pace
the
n
it
is
possi
ble
to
ha
ve
e
ntire
data
ba
se
in
m
ai
n
m
e
m
or
y.
Fo
r
e
xam
ple
da
ta
base
em
ploy
ee
detai
ls
of
s
om
e
s
m
al
l
com
pan
y.
But
f
or
real
ti
m
e
appl
ic
at
ion
s
it
is
m
us
t
that
data
reside
i
n
m
ai
n
m
e
m
or
y.
If
siz
e
of
dat
abase
is
to
o
la
rg
e
to
fit
into
the
m
a
in
m
e
m
or
y
f
or
e
xam
ple
an
app
li
cat
io
n
with
sat
el
li
te
i
m
a
ge
d
at
a
the
da
ta
can
be
cat
egorized
as
ho
t
an
d
c
old
da
ta
.
T
he
da
ta
whic
h
i
s
fr
e
qu
e
ntly
re
quire
d
is
cat
eg
ori
zed
as
hot
a
nd
data
w
hich
is
rar
el
y
requir
ed
a
nd
is
volu
m
ino
us
is
col
d
data.
The h
ot d
at
a la
y i
n
in
-
m
ai
n
m
e
m
or
y an
d
c
ol
d
is st
or
e
d o
n d
isk.
2.1.
Chall
enges
w
ith in
-
mem
ory
sy
ste
ms
In
-
m
e
m
or
y
is
li
able
to
cha
nge
ra
pid
ly
wh
il
e
disk
sto
rag
e
is
non
vola
ti
le
.
S
o
regular
backup
s
m
us
t
be
ta
ken
(
on
dis
k)
and
at
the
sam
e
tim
e
it
is
to
be
ta
ken
care
t
ha
t
per
f
or
m
ance
of
IM
DB
is
no
t
aff
ect
ed.
If
di
sk
fail
,
data
on
oth
e
r
di
sk
s
can
be
sec
ur
e
d
a
nd
rec
overy
from
disk
is
easy
but
if
m
e
m
or
y
fail
s,
e
nt
ire
database
is
lost.
The
perform
ance
gain
of
IM
DB
can
be
li
m
it
ed
by
the
a
ppli
cat
ion
operati
ng
it
,
la
yo
ut
an
d
im
ple
m
entati
on
of
database
it
sel
f,
the
ha
rdwa
re
on
w
hich
t
he
da
ta
base
is
r
unni
ng
a
nd
the
ass
ociat
ion
with
e
xter
nal
de
vices
.
Lar
ge
vo
l
um
es
of
da
ta
with
lo
we
r
fr
e
qu
e
ncy
reads
are
not
m
uch
m
or
e
ef
fici
ent
with
IM
DBs.
Ma
ny
pa
pers
ha
ve
discusse
d
t
he
i
m
pact
of
m
e
mo
ry
reside
ncy
of
s
om
e
i
m
po
rtant
f
un
ct
io
nal
c
om
po
ne
nts
of
db
m
s
li
ke
co
nc
urren
cy
con
t
ro
l,
acce
ss
m
et
ho
ds
,
c
om
m
i
t
processi
ng,
query
pro
cessi
ng
a
nd
pe
rfor
m
ance
et
c.
Ma
ny
pa
pe
rs
ha
ve
discusse
d t
he
i
ssu
es
relat
ed
t
o IMDB
rec
ov
e
r
y and b
riefly
e
xam
ine so
m
e o
f
the
so
l
utions
[2
]
.
2.2.
In
-
mem
or
y ar
chitecture
To
un
der
sta
nd
the
a
rch
it
ect
ur
e
of
i
n
-
m
e
m
or
y,
arc
hitec
ture
of
OR
ACLE
da
ta
base
is
co
nsi
der
e
d
her
e
,
wh
ic
h
is
cat
eg
or
iz
e
d
unde
r
t
he
in
-
m
e
m
or
y
cache
arc
hitec
ture
.
The
el
em
e
nts
of
arc
hitec
ture
incl
ude
dat
abase
processes
,
m
e
m
or
y
-
residen
t
data
str
uctu
res
,
sh
a
re
d
l
ibra
ri
es
an
d
a
dm
ini
strat
ive
pro
gra
m
s.
In
de
xes,
s
yst
e
m
ta
bles,
ta
bles,
cur
s
ors,
loc
ks,
tem
po
rar
y
in
dex
e
s
an
d
c
om
pi
le
d
com
m
ands
to
gether
m
ake
up
t
he
m
e
m
or
y
reside
nt
data
s
tructu
res.
Th
r
ough
direct
li
nk
an
d
cl
ie
nt/ser
ver
co
nnect
ions,
the
ap
plica
ti
on
can
be
li
nk
ed
to
the
data
base
or
IMDB
cac
he
.
Inform
at
ion
is
receive
d
by
re
plica
ti
on
a
ge
nts
from
m
ast
er
databases
an
d
i
s
sen
t
to
s
ubscri
ber
databases
.
Asy
nchr
onous
dat
a
tra
ns
fe
rs
bet
ween
or
acl
e
da
ta
base
a
nd
ca
che
gro
ups
i
n
the
in
-
m
e
m
or
y
datab
ase
cac
he
a
re
pe
rfo
rm
ed
by
cache
ag
e
nt
.
Fi
gure
1
s
ho
ws
oracl
e’s
in
-
m
e
m
or
y
dat
abase
cache a
rch
it
ect
ur
e
[3].
E
xter
na
l
m
e
m
or
y
is acc
essed o
nly i
n t
hr
ee
cases:
To
l
oad co
py
of m
ai
n
m
e
m
or
y durin
g
syst
e
m
startup
.
Check
point
over
wr
it
ing,
recovery a
nd
on Logg
i
ng.
To per
sist
s d
at
a ab
o
ut
data a
nd c
onfig
ur
at
i
on ch
a
ng
e
s.
Evaluation Warning : The document was created with Spire.PDF for Python.
Com
pu
t. Sci.
I
nf. Tec
hnol.
Eff
ic
ie
nt a
nd s
cala
ble mult
it
enant
pla
ce
men
t
app
r
o
ac
h
for
in
-
memory
data
ba
s
e
over…
(
A
rp
it
a S
hah
)
41
Figure
1
.
In
-
m
e
m
or
y data
bas
e cache
arc
hitec
ture
So
for
t
hese
ty
pe
of
ta
s
ks
IM
DBs
rely
on
pa
ged
data
ha
nd
li
ng
w
hile
al
l
ot
her
operati
ons
run
purely
on
m
ai
n
m
e
m
or
y.
The
data
base
com
m
un
it
y
is
go
i
ng
to
ex
per
i
ence
a
great
s
hi
ft
in
m
ark
et
in
the
c
om
ing
ye
ars
a
s
in
-
m
e
m
or
y
dat
abases
are
bec
om
ing
m
or
e
e
ff
ect
ive
a
nd
af
ford
a
ble.
Alt
houg
h
i
n
-
m
e
m
or
y
database
m
ai
nly
consi
sts
two
s
tora
ge
a
ppro
ac
h
nam
ely
ro
w
-
ori
ented
(
data
base
re
-
str
uct
uri
ng)
a
nd
c
olum
n
-
or
ie
nted
.
Ma
ny
databases
can
u
se
both a
ppr
oa
ches
row
-
ori
ented
a
s
well
as
colum
n
-
ori
ente
d.
2.3.
Multite
nancy
W
it
h
t
he
ai
m
of
re
duci
ng
op
erati
on
al
co
st,
m
ul
ti
-
te
nan
cy
pro
vid
es
m
et
ho
ds
for
com
bin
in
g
m
ulti
ple
custom
ers
(i.e
.
te
na
nts)
of
de
plo
ye
d
a
pp
li
ca
ti
on
w
hich
r
un
on
t
he
sam
e
i
nfrastr
uctu
re
.
Database
la
ye
r
can
be
e
m
plo
ye
d
in
the
Mult
i
te
na
nc
y
arch
it
ect
ure
can
be
util
iz
ed
in
the
databa
se
la
ye
r
in
s
uc
h
a
way
that
a
sing
l
e
database
ca
n
be
s
har
e
d
by
m
ulti
ple
c
us
tom
er.
A
cl
ou
d
use
s
te
ch
nolo
gy
of
m
ulti
te
nan
cy
to
sh
a
re
IT
r
eso
ur
ce
s
a
m
on
g
m
ulti
pl
e
a
ppli
cat
ion
s
and
te
na
nts
se
c
ur
el
y.
Virt
ualiz
at
ion
-
base
d
a
rc
hitec
tures
is
use
d
by
s
om
e
cl
ou
ds
to
isolat
e
te
nan
ts
and
so
m
e
us
es
custom
so
ftwa
r
e
arch
it
ect
ures
to
get
the
job
done
.
De
pe
nd
i
ng
on
requirem
ents
of
custom
er
su
c
h
as
secu
rity
,
hi
gh
a
vaila
bili
ty
,
custo
m
iz
abili
ty
,
the
cho
ic
e
of
ap
pro
pr
ia
te
te
nan
cy
m
od
el
is
decide
d.
The
re
are
se
ve
ral
po
ssible
m
ulti
-
ten
ant
schem
es
li
ke
s
har
e
d
desi
gn,
VM
based
desig
n
et
c.
F
or
te
nan
t
app
li
cat
io
ns
ar
e
a
well
-
kn
own
exam
ple
of
a
ty
pe
of
a
ppli
cat
ion
w
hose
da
ta
and
wor
klo
a
ds
ca
n
be
par
ti
ti
on
e
d
easi
ly
.
For
ins
ta
nce,
with
te
na
nt
a
pp
li
cat
ions,
data
an
d
w
orkl
oad
can
ty
pi
cal
ly
be
pa
rtit
ion
e
d
al
ong
te
nan
t
bounda
ries
si
nc
e
m
os
t
reques
ts
are
withi
n
th
e
co
nf
i
nes
of
a
te
nan
t.
S
o,
by
c
on
si
der
i
ng
a
fra
m
ewo
r
k
w
hich
ta
ke
s
the
te
na
nt
w
or
klo
a
ds
a
s
in
put,
their
perf
orm
ance
SL
Os
(
Ser
vice
Level
Objecti
ves
),
a
nd
the
ser
ve
r
ha
rdware
wh
ic
h
is
obta
in
able to t
he Saa
S provi
der
.
3.
A
S
UPPLE I
NFRAST
R
U
CTU
RE F
OR MU
LT
ITE
N
ANC
Y
O
VER
IN
-
ME
MO
R
Y DA
T
ABA
S
E
Pr
op
os
e
d
arc
hi
te
ct
ur
e
f
or
m
ulit
it
enan
cy
us
in
g
in
-
m
e
m
or
y
is
show
n
in
Fi
gure
2
w
hich
c
onsist
s
m
ajo
r
three c
om
po
ne
nts.
3.1.
Co
m
ponen
t(
s
)
(1)
Cl
ust
er
hea
d:
T
her
e
is
si
ngle
cl
us
te
r
le
a
de
r
e
xists
in
a
S
upple
Infr
a
struc
ture
a
nd
it
as
s
ign
s
one
or
m
or
e
te
nan
t
to
serv
e
r.
Each
te
nan
t'
s
re
plica
is
assig
ned
to
t
he
ind
i
vidual
in
sta
nce
handler,
so
t
hat
each
in
sta
nce
handler
m
us
t
process
diff
e
r
ent
data
from
m
ulti
ple
te
nan
ts.
The
cl
ust
er
hea
d
m
ai
ntains
t
he
place
m
ent
inf
or
m
at
ion
ov
er
in
m
e
m
or
y
database
pe
rfo
r
m
ance
with
hard
dis
k
base
d
show
n
in
Fig
ur
e
2,
wh
ic
h
it
pro
p
agate
s
to
the
r
ou
te
.
I
n
ad
diti
on
t
he
cl
us
te
r
hea
d
al
so
obser
ve
a
ct
ive
nodes
,
s
ta
rts
and
st
ops
severs,
place
m
ent
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2722
-
3221
Com
pu
t. Sci.
I
nf. Tec
hnol.
,
V
ol.
1
, N
o.
2
,
J
ul
y
20
20
:
39
–
46
42
assessm
ent
and
m
igrati
on
of
te
nan
t
bet
ween
s
erv
e
rs.
Re
quest
proce
ssin
g
ca
nnot
assesse
d
di
rectl
y
by
the
cl
us
te
r
head..
T
he
cl
ust
er
hea
d
proce
s
s
h
as
to
run
on
al
l
act
ive
sever
s
to
handle
sin
gle
point
of
fai
lure
to
m
ake
hi
gh
ly
avail
abili
ty
that
can
be
possib
le
by
run
ning
a
cl
us
te
r he
ad
process
on
al
l
act
ive
ser
ver
s
.
(
2)
Rou
te
r:
It
f
orwards
qu
e
ry
re
quest
to
the
su
it
able
i
ns
ta
nce
m
anager
ba
sed
on
cl
us
t
er
m
ap
inf
orm
ation
a
nd
al
so
from
ou
tsi
de
the
cl
us
te
r.
Als
o,
it
pr
ov
i
des
l
ocati
on
tra
nspare
nc
y
of
a
te
na
nt'
s
database
.
The
job
of
r
oute
r
ne
ed
t
o
be
balanc
ed
t
he
load
ac
r
os
s t
he
tenan
t
’s replic
as in
rou
nd
-
r
ob
in p
at
te
r
n.
Figure
2
.
A
supp
le
i
nfrastr
uct
ur
e
f
or m
ulti
-
ten
ancy
over
in
-
m
e
m
or
y datab
ase
If
a
re
plica
of
a
te
nan
t
bec
om
e
s
una
vaila
ble
in
c
onseq
ue
nce
on
fail
ure
of
a
serv
e
r
[4
,
5
]
re
qu
e
sts
ne
e
d
to
be
b
al
a
nced
a
m
on
gst
the
re
m
ai
nin
g l
ive
re
plica
s. The
qu
e
ry r
e
su
lt
s
send
back t
o
t
he
app
li
cat
ion
th
r
ough
t
he
router.
(3)
I
ns
t
ance
Ma
na
ger
:
T
he
jo
b
of
i
nst
ance
m
ang
e
r
i
s
to
m
anag
e
th
e
dist
rib
ution
of
requests
ac
ross
the
te
nan
t
us
e
r.
Wh
en
insta
nce
m
anag
e
r
receive
a
w
rite
re
quest
,
it
wr
it
e
to
cac
hing
databa
se
[
6]
(if
no
s
pace
ov
e
r
m
ai
n
m
e
m
or
y)
an
d
al
s
o
f
orw
ard
t
he
re
quest
to
it
s
su
cce
ss
or
no
de
i
n
a
cl
us
te
r.
Wh
il
e,
ha
nd
li
ng
the
dat
abas
e
ov
e
r
no.
of
se
r
ver
s
c
on
c
urre
nt
requests
m
ay
cause
perform
ance
iss
ues
an
d
it
e
ff
ect
on
a
pp
li
cat
io
n
e
xec
ution.
To
res
olv
e
thi
s
prob
le
m
we
hav
e
e
xperim
e
nted
on
Or
acl
e
S
hardin
g
A
r
chite
ct
ur
e
to
s
upport
Ela
sti
c
scal
e.
Re
qu
est
f
or
t
he
w
rite
operati
on
need
to
be
ob
sti
natel
y
wr
it
te
n
on
i
no
de
out
of
n
no
des
a
nd
then
a
sync
hronou
sly
rep
li
cat
ed
t
o
th
e
nodes
i
+
1
t
o
n.
E
ach
te
nan
t
co
ns
ist
s
two
c
on
sist
e
nt
re
plica
s
an
d
I
ndivid
ual
instance
m
anag
e
r
is
co
up
le
d
up
with
a
l
ocal
Oracl
e
instance,
wh
ic
h
is
s
ha
re
d
am
on
gst
m
ul
ti
ple
te
nan
ts.
O
racle
insta
nce
su
pp
or
t
the
ap
proac
h
of
m
ulti
te
nan
cy
[7
]
an
d
in
-
m
e
m
or
y.
A
local
O
racle
instance
nee
ds
to
be
paire
d
with
instance m
anger.
3.2.
Be
nchma
r
k de
sign
We
ha
ve
use
d
ded
ic
at
e
d
sam
ple
be
nch
m
arks
f
or
m
ixed
w
orkloa
d
pr
oces
sing
that
be
nc
hm
ark
cal
le
d
HRSB
-
MT
s
ho
wn
i
n
Table
1.
Our
te
sti
ng
e
xp
erim
ent
is
on
O
racle
in
-
m
e
m
or
y
colum
n
datab
ase,
w
hich
runs
for
m
ixed
w
orkloa
d
proce
ssin
g
a
pp
li
cat
io
n.
I
n
colum
n
-
ori
ente
d
a
ppr
oac
h,
da
ta
base
kee
p
ev
ery
at
tribu
te
f
or
in
a
separ
at
e
c
olu
m
n
str
uctur
e
a
nd
is
i
deal
f
or
a
na
ly
ti
cs,
since
it
al
lo
ws
f
or
s
pe
edy
data
retrie
val
w
hen
on
ly
a
fe
w
colum
ns
are
sel
ect
ed
bu
t
t
he
query
acce
sses
a
huge
portion
of
the
data
set
.
Wh
e
n
DML
op
erati
on
(
i
ns
ert,
update
or
delet
e)
occurs
on
bo
t
h
m
eth
ods
the
n
r
ow
-
or
ie
nted
f
or
m
at
is
ext
rem
ely
eff
ect
ive
f
or
processin
g
DML
as
it
m
anipu
la
te
s
an
entire
r
ow
or
record
in
on
e
go.
A
col
um
n
appr
oach
is
no
t
so
pr
of
ic
ie
nt
as
com
par
ed
t
o
r
ow
form
at
a
t
pr
oc
essing
r
ow
-
wis
e
DML
but
f
or
OLAP,
la
r
ge
r
chun
k
of
data
Colum
-
or
ie
nte
d
gi
ves
sim
ultan
eo
us
data exe
cutio
n. HRSB
-
MT
m
od
el
als
o o
pted
d
ic
ti
on
a
ry
-
bas
ed
c
om
pr
essio
n
te
ch
niques
.
Table
1.
HR
sc
hem
a b
ench
m
ark
-
MT
Table_
n
a
m
e
Tables
ize(
MB)
Reg
io
n
9
5
1
9
5
5
Co
u
n
tries
2
2
0
2
2
4
Locatio
n
s
332
Dep
art
m
en
ts
5171
Jo
b
s
1
0
3
2
5
6
7
E
m
p
lo
y
e
es
1
9
5
6
8
7
9
Evaluation Warning : The document was created with Spire.PDF for Python.
Com
pu
t. Sci.
I
nf. Tec
hnol.
Eff
ic
ie
nt a
nd s
cala
ble mult
it
enant
pla
ce
men
t
app
r
o
ac
h
for
in
-
memory
data
ba
s
e
over…
(
A
rp
it
a S
hah
)
43
3.3.
Resource c
on
s
umpt
i
on
of m
ultipl
e h
omog
eneous te
nant
So
,
by
co
ns
ide
r
ing
a
f
ram
ewo
r
k
w
hich
ta
ke
s
work
l
oa
ds
of
the
te
na
nt
as
input,
their
perfor
m
ance
SLO
s
(S
er
vice
Le
vel
Objecti
ves
)
[
8,9]
and
t
he
ser
ve
r
ha
rdwar
e
w
hich
is
obta
ina
ble
to
the
D
aa
S
pro
vid
e
r,
a
nd
res
ult
into
a
cost
-
ef
fe
ct
ive
recipe
whic
h
s
pecifies
u
ti
li
zat
ion
of
hardw
a
re
t
o
delive
r
a
nd
ho
w
the
t
enan
ts
are
sche
du
le
d
on
a
vaila
ble
ha
rdwar
e
resou
r
ce.
Eac
h
te
na
nt
s
co
ntain
the
sam
e
siz
e
and
request
rate
on
a
seve
r.
T
otal
no.
of
us
ers
is d
ist
ri
bute
d un
i
form
ly a
m
on
g al
l t
enan
ts an
d
the
se
rv
e
r
is fil
le
d u
p
on
ly
15
-
20%
o
f
it
s m
ai
n
m
e
m
or
y is
us
e
d.
T
he
te
na
nt
siz
e
ts
is
di
vid
e
d
by
the
r
esulta
nt
am
ount
of
m
e
m
or
y.
As
a
res
ult
de
pendin
g
on
the
c
ho
se
n
value
for
ts
se
rv
e
r
m
ay
con
ta
in
fe
w
or
m
or
e
te
na
nts,
s
o
we
va
ry
ts.
From
20
to
19
8
MB
(f
r
om
40
0,0
00
t
o
4,000,0
00
rows
).
Re
qu
est
p
er
te
nan
t
is
de
note
d
by
TR
a
nd
i
t
m
ay
increase u
ntil
SL
O
vi
olate
d
[
10
,
11
].
Figur
e
3 Sh
ows
that
w
hen th
e
num
ber
of te
nan
ts
is i
ncr
ease
d
by a
fa
ct
or of
10, t
hr
oughput
dec
rea
ses
by
10%.
A
SLO
per
s
pecti
ve
[
12
]
vio
la
te
s
w
het
her
the
se
rv
e
r
s
can
f
un
c
ti
on
is
util
iz
ed
by
sm
a
ll
nu
m
ber
of
la
r
ge
te
nan
ts
or
se
ver
al
sm
a
ll
tenan
ts.
Figure
3
.
Ma
xi
m
u
m
Th
ro
ug
hput
with
ou
t
viol
at
ion
the
p
e
rfo
rm
ance SLO
4.
STANDI
NG
TE
NANT PL
AC
E
MENT
Fo
r
stan
ding te
nan
t
placem
ent, w
e
h
a
ve
c
ons
idere
d
f
ollo
wing
gen
e
ral
data
as
in
pu
t.
A vali
d
te
na
nt
assignm
ent is
perform
ed
us
in
g bina
ry d
eci
si
on
x
∈
{0
,1
}
S × T
× R
, w
he
re
()
,
y
ti
x
= { 1, if
tena
nt copy
y
is
on se
ver
i
, othe
r
wise
0
}
(1)
s
∈
{
0,1}
N
(2)
w
he
re
s
i
=
1
i
nd
i
cat
es that sp
eci
fic ser
ver
i
is a
ct
ive, o
t
herwis
e ser
ver is i
nac
ti
ve.
Now,
as
pe
rform
ing
te
na
nt
pl
ace
m
ent
al
so
ne
eds
to b
e
c
hec
ked
t
hat
re
plica
of
te
nan
t
is
al
locat
ed
to
a
serv
e
r
once
or
no
t
an
d
no
tw
o
co
pies
of
the
s
a
m
e
te
nan
t.
S
o,
to
c
hec
k
t
he
s
pecific
c
onditi
on
we
ha
ve
ap
plied
so
m
e con
strai
nt
s.
,
1
y
ti
iN
x
∀
t
∈
T
,
∀
y
∈
R
(3)
,
1
y
ti
yR
x
∀
t
∈
T
,
∀
i
∈
N
(4)
The
se
r
ver
i
n
-
m
e
m
or
y ca
pacit
y needs
to fit
on full
am
ou
nt
siz
e o
f
all
tena
nt on
t
hat ser
ve
r.
()
,
(
)
(
)
y
t
i
i
t
T
y
R
t
x
c
i
s
∀
i
∈
N
(5)
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2722
-
3221
Com
pu
t. Sci.
I
nf. Tec
hnol.
,
V
ol.
1
, N
o.
2
,
J
ul
y
20
20
:
39
–
46
44
On
e
of
the
im
p
or
ta
nt
key
featur
es
of
gr
ee
dy
heurist
ic
s
is
th
at
they
are
c
om
pu
ta
ti
on
al
ly
l
ess
rig
oro
us
than
m
et
a
-
heu
r
ist
ic
s.
Gr
ee
dy
a
lgorit
hm
s
are
loo
s
el
y
base
d
on
t
he
well
-
kn
own
be
st
-
fit
al
gorithm
.
It
al
s
o
deliver
s
good
resu
lt
s
for
the
relat
ed
bi
n
-
p
ac
king
probl
e
m
.
Althou
gh
the
pr
ob
le
m
in
a
Be
st
Fit
al
gor
it
h
m
con
sist
c
onsta
nt
appr
ox
im
at
ion
rati
o
over
bi
n
pack
i
ng.
So,
w
hen
the
te
na
nts
are
sm
al
l
it
incli
ned
t
o
bundl
e
lots
of
te
nan
t
s
on
a
serv
e
r.
Be
st
-
fit
al
gorithm
is
use
d
for
te
nan
t
pl
ace
m
ent,
w
hic
h
fin
ds
t
he
se
rver
with
th
e
le
as
t
rem
ai
nin
g
re
s
ource
that
can
acc
omm
od
at
e
each
te
nan
t,
a
nd
in
ca
se
su
c
h
a
se
r
ve
r
ca
nnot
be
f
ound,
ei
ther
re
la
xes
the
c
on
st
raints
gr
a
dual
ly
or
use
a
n
e
w
server
.
Starts
with a
r
andom
p
la
cem
ent.
It a
ssu
m
es a
f
ixe
d n
um
ber
of te
nan
t
t
ype
s a
nd,
on eac
h
se
rv
e
r, i
t assi
gns te
na
nts
of the
sam
e
ty
pe
to
the
sa
m
e d
at
abase
se
rv
e
r
i
ns
ta
nce
wi
th a
fixe
d
am
ou
nt
of
m
ai
n
m
e
m
or
y al
locat
ed
[
13]
.
4.1.
Prop
os
ed
mul
titen
ant pl
ace
ment
(
MTP
)
al
go
ri
t
hm
Wh
e
n
te
na
nts
qu
a
ntit
y
is
increasin
g
the
n
w
e
need
t
o
ap
pl
y
te
nan
t
place
m
ent
al
go
rith
m
to
al
locat
e
serv
e
r
res
ource
s
[
14,
15
]
.
Pa
ra
m
et
ers
f
or
Mul
ti
-
te
nan
t
place
m
ent
al
gorith
m
as
sh
ow
n
i
n
Ta
ble
2
a
re
il
lustrate
the tena
nt,
se
r
ve
r
ca
pacit
y, re
pl
ic
a o
f
te
nan
t a
nd o
t
her res
our
ce pa
ram
et
ers.
Table
2
.
MT
P
par
am
et
ers
Sy
m
b
o
l
Meanin
g
T
N
T
Sh
o
ws n
o
.
o
f
tenants,
Ti
represen
t ith
ten
an
t
S
I
S
No
.
o
f
servers (
to
s
h
o
w r
eso
u
rce
allo
catio
n
)
σ :
T
→
NT
+
Fu
n
ctio
n
r
etu
rns
M
ain
m
e
m
o
r
y
requ
ire
m
en
t of
a
g
i
v
en
tenan
t
c
:
S
→
IS
+
Fu
n
ctio
n
r
etu
rns
M
ain
m
e
m
o
r
y
capacity
o
f
ith serv
er
The
pur
pose
of
m
entione
d
al
gorithm
wh
ic
h
im
ple
m
ented
on
propose
d
a
rch
it
ect
ure
is
to
im
pr
ove
util
iz
at
ion
of
s
erv
e
r
t
o
the
set
of
te
na
nts.
All
ocati
ng
prop
e
r
resou
rces
over
the
te
na
nts,
we
ha
ve
de
plo
ye
d
th
e
te
nan
t
placem
e
nt
al
gorithm
in
Cl
us
te
r
head
t
o
im
pr
ov
e
ser
ve
r
util
iz
at
ion
.
To
f
ulfil
te
nant
'
s
serv
ic
e
qual
it
y
as
its
basis
re
quir
e
m
ent,
the
perf
or
m
ance
te
sti
ng
us
e
basic
f
un
ct
ion
s
wappi
ng
as
a
re
su
lt
it
is
consi
der
i
ng
m
i
nim
u
m
nu
m
ber
of
ser
ve
rs
as
an
d
w
he
n
it
is
re
quire
d.
O
n
the
ba
se
of
m
ee
ti
ng
the
re
qu
i
rem
en
ts
of
serv
ic
e
qual
it
y
of
the
te
nan
t,
e
xperi
m
ental
te
st
us
e
Be
st
fit
heu
ris
ti
c
al
go
rithm
s
wh
ic
h
is
com
par
ed
with
the
pro
posed
M
ulti
-
te
nan
t
placem
ent algorit
hm
. I
n
a
res
ul
ti
ng
outcom
es MTP r
eq
uires
le
ss num
ber
of
serv
e
rs.
Algorithm
: MTP
Inpu
t:
T
Tenan
ts,
Server in
-
m
e
m
o
r
y
capacity
c
,
Rep
lica
y
Fin
d
bin
ary
valu
e t
h
ru eq
.
(1)
(0,1
)
T={
t
1
,t
2
..
.
..
.t
n
}=
so
rted list o
f
tenan
t i
n
decreasin
g
ord
er
with
r
esp
ect of
siz
e
for serv
er
i
in
S
=
{ 1
,2,3
}
do
for
each
t
i
∈
T
(
2
)
,
ti
x
if
(
2
)
,
ti
x
then
lo
cate oth
er
serv
e
r
j
su
ch
that
(
1
)
,
tj
x
call swap(
t
,
i,
j)
else
p
lace
t
i
o
n
server
end i
f
end
f
o
r
call σ(
T[
i
]
)
if
(
T[
i
]
<
c
)
then
S
[
j
]
=
S
[j
]
-
T
[
i
]
end if
end f
o
r
Evaluation Warning : The document was created with Spire.PDF for Python.
Com
pu
t. Sci.
I
nf. Tec
hnol.
Eff
ic
ie
nt a
nd s
cala
ble mult
it
enant
pla
ce
men
t
app
r
o
ac
h
for
in
-
memory
data
ba
s
e
over…
(
A
rp
it
a S
hah
)
45
5.
E
X
PERI
MEN
TAL SET
UP
AND
RES
UL
TS A
NA
L
YS
I
S
The
ex
per
im
ents
we
re
ca
rr
ie
d
ou
t
on
a
cl
us
te
r
of
t
hr
ee
se
rv
e
r
s
w
hich
is
sho
wn
in
Fi
gure
2,
each
hav
i
ng
16GB
m
e
m
or
y,
VM
(s),
run
nin
g
Ce
nt
OS
a
nd
O
racle
in
-
m
e
m
or
y
database.
W
e
pro
duced
up
to
60
te
nan
t
s.
Eac
h
te
nan
t
runs
wi
th
m
ix
work
l
oa
ds
i
n
pro
portion
to
the
ge
ne
rated
query
pa
tt
ern
.
C
oncer
nin
g
the
siz
es
of
the
te
nan
ts,
we
m
easur
e
d
the
in
-
m
e
m
or
y
datab
ase
with
m
ulti
t
enan
cy
set
ti
ngs
existi
ng
in
Mi
cro
s
of
t
SQL
Az
ure
[16
-
17]
.
Fig
ur
e
4
sho
ws
that
nu
m
ber
of
te
nant
s
placed
on
ea
ch
ser
ver
(i.e.
no
of
ser
ve
r
=
2,
3
an
d
4)
f
or
res
ourc
e
al
locat
ion
[
16
]
.
I
f
t
her
e
a
re
8
te
nan
ts
la
bele
d
as
A,
B,C
,
D,
E
,
F,G,H
a
nd
for
three
ser
ver
nu
m
ber
ed
as
1,
2
and
3.
If
te
na
nt
B
,
C,
D
placed
on
s
erv
e
r
1
the
n
1
=
{
B
,
C,
D
}
li
kew
i
se
for
ot
he
rs
te
na
nts
too.
W
he
n
the
num
ber
of
serv
e
rs
is no
t
e
nough, both
al
gorithm
best
fi
t
an
d
MTP
u
se
s
di
ff
e
ren
t
strat
egies. Figure
4
sho
ws
no.
of
t
enan
ts
placed
on
eac
h
serv
e
r
num
bered
as
1
,
2
a
nd
3.
I
n
this
pa
per
we
ad
opt
the
SLA
penal
ty
m
od
el
[
17
]
,
i
f
qu
e
ries
arr
ive
d
duri
ng
serv
e
r
overloa
d
will
fail
to
noti
ce
their
SL
A
dead
li
ne
s
a
nd
the
pen
al
ty
nee
d
to
be
pa
id
by
serv
ic
e
pro
vid
er;
an
d
ot
her
queries
wil
l
m
ee
t
their
SL
A.
Re
aso
n
is
us
ing
the
l
oa
d
of
a
te
na
nt
will
ch
ang
e
ver
y
fr
e
quently
for
opti
ng
SL
A
m
od
el
.
SLA
pen
al
ti
es
occ
ur
m
os
tl
y
du
e
to
prolo
nged
sys
tem
ov
erloa
d
i
ns
te
ad
of
a
te
m
po
rar
y
burst in
ar
rival
of query
f
or
a
s
hort
per
io
d (e
xam
ple arr
i
val dur
i
ng 10 m
il
l
ise
cond
).
Figure
4
.
N
o. of tena
nt
on seve
r
As
it
sh
ow
n
i
n
Figure
5
s
ho
ws
that
ho
w
que
ry
pr
ocesse
d
us
in
g
tw
o
dif
fer
e
nt
a
ppr
oac
hes:
Be
st
fit
Gr
ee
dy
an
d
M
TP,
in
w
hich
B
est
fit
work
e
d
on
ha
rd
disk
ba
sed
qu
e
ry
pro
cessi
ng
w
hile
MTP
proces
se
d
que
ry
ov
e
r
in
-
m
e
m
or
y
database
cl
ust
er.
Cost
of
query
processi
ng
thr
ough
te
na
nt
is
com
par
at
ively
low
in
MT
P
tha
n
Be
st
fit
Gr
ee
dy
an
d
al
s
o
w
orks
ef
fecti
vely
tha
n
best
fit.
We
ha
ve
al
s
o
exp
e
rim
ented
scal
abili
ty
of
t
enan
t
placem
ent.
Fig
ur
e
6
s
how
s
runn
i
ng
ti
m
e
of
MTP
wh
ic
h
is
for
50
te
na
nts.
Runnin
g
tim
e
f
or
Be
st
fit
gr
ee
dy
i
s
insig
nificant
be
low
0.3
sec
ond
e
ve
n
for
30
t
enan
ts
,
wh
e
rea
s
MTP
ta
kes
bi
t
longer
ti
m
e
t
han
Be
st
fit
gr
eedy
appr
oach.
Figure
5
.
Ef
fec
ti
ven
ess
of
te
na
nt p
la
cem
ent
Figure
6
.
Scal
a
bili
ty
o
f
te
na
nt
placem
ent
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2722
-
3221
Com
pu
t. Sci.
I
nf. Tec
hnol.
,
V
ol.
1
, N
o.
2
,
J
ul
y
20
20
:
39
–
46
46
6.
CONCL
US
I
O
N
Mult
it
enan
cy
with
in
-
m
e
m
or
y
database
(
opte
d
in
O
racle)
s
peeds
up
the
processin
g
a
nd
r
esp
on
se
ti
m
e
of
data
re
quest
.
For
in
-
m
e
m
o
ry
databa
se
we
ha
ve
te
ste
d
over
di
ff
e
ren
t
platfor
m
li
ke
SAP
H
ANA,
G
rid
Gr
a
i
n
and
Or
acl
e.
I
n
this
pa
pe
r
we
ha
ve
pro
posed
m
ul
ti
te
nan
c
y
arch
it
ect
ure
(supp
le
)
us
in
g
i
n
-
m
e
m
or
y
datab
ase
wi
t
h
pro
po
se
d
MTP
al
gorithm
.
F
r
om
the
per
s
pe
ct
ive
of
ef
fici
ency
th
e
pa
per
sh
ows
pro
po
se
d
MT
P
al
gorithm
in
com
par
ison
with
Be
st
fit
G
ree
dy
a
ppro
ac
h
wi
th
database
be
nc
hm
ark
(
HRS
B
-
MT)
over
fe
w
te
nan
ts
to
im
pro
v
e
the
qu
al
it
y
of
te
nan
t
placem
ent.
Wh
il
e
this
pap
e
r
f
ocuses
on
s
upplem
ent
arc
hitec
ture
f
or
m
ulti
te
nan
t
with
i
n
-
m
e
m
or
y datab
ase, in
futu
re
will
w
or
k on dy
nam
ic
p
la
cem
ent app
ro
ac
h.
REFERE
NCE
S
[1]
Hung,
Tra
n
&
Huang,
Chuanh
e.
“
An
Eff
e
ct
iv
e
Data
Placem
ent
Strateg
y
in
Main
-
Mem
or
y
Data
base
C
luster.
”
Proce
edi
ngs
-
2n
d
Inte
rnat
ional
Confe
renc
e
on
N
et
working
and
D
istribut
ed
Comp
uti
ng
,
ICND
C
,
p
p.
93
-
98
.
2011
.
[2]
Soon
M.
Chung.
“
Para
ll
el
m
ai
n
m
emor
y
d
ataba
se
s
y
stem.
”
In
Proc
ee
dings
o
f
the
19
92
ACM/
SIGA
P
P
Symposium
on
Appl
ie
d
comput
i
ng:
technologic
al
cha
ll
eng
es
of
the
1990's
(SA
C
'
92)
,
Ha
l
B
er
ghel
,
Ed
Dea
ton
,
Georg
e
Hedri
c
k,
David
Roa
ch, a
n
d
Roger
W
a
inwright
(
Eds.
). ACM
,
New York, NY
,
US
A,
pp.
273
-
282,
1992
.
[3]
Orac
le
Dat
ab
ase
In
-
Mem
or
y
with
O
rac
l
e
Da
ta
bas
e
12c
Re
l
ea
se
Mar
ch
2017
fro
m
htt
ps://
do
cs.
ora
c
le
.
com/cd/
E1666
2_01/doc
/tim
este
n.
1121/b56058
/
arc
h.
h
tm
[4]
Jan
Schaf
fne
r
,
B
enj
amin
Ec
kar
t,
Dea
n
Jac
obs,
Ch
risti
an
Schw
arz
,
Hass
o
Plat
tne
r
,
a
nd
Alexa
nd
er
Z
e
ie
r.
“
Predic
t
ing
in
-
m
emor
y
da
tabase
per
form
an
c
e
for
aut
om
at
ing
cl
uste
r
m
ana
ge
m
ent
ta
sks
.
”
In
Proce
ed
ings
of
t
he
2011
I
EE
E
2
7th
Inte
rnational
Co
nfe
renc
e
on
Dat
a
Engi
n
ee
ring
(I
CDE
'
11)
.
I
EEE
Com
pute
r
Soci
ety
,
W
ashingt
on
,
DC,
US
A,
pp
.
1264
-
1275
,
201
1
.
[5]
Huang,
Yunkui
&
Zha
ng
,
Yansong
&
Ji,
Xiaodo
ng
&
W
ang,
Zh
a
nwei
&
W
ang,
S
han.
“
A
Data
Di
stribut
ion
Stra
teg
y
for S
ca
l
abl
e
Mai
n
-
Mem
ory
Dat
a
base
.
”
Advanc
es
in
We
b
and
Ne
t
work
Technol
ogi
es,
and
Informat
ion Managemen
t
,
AP
We
b
/WA
IM
2
009
Int
ernati
ona
l
Workshops
:
W
CMT,
RTBI
,
DBI
R
-
ENQOIR
,
P
AI
S
,
Suzhou
,
Chin
a
,
pp
.
13
-
24
,
Apr
il
2009
.
[6]
Jac
obs,
Dea
n
,
an
d
Stefa
n
Aulbach.
“
Rum
ina
ti
ons
on
Multi
-
T
ena
n
t
Data
base
s.
”
In
B
TW
,
vol.
103,
pp
.
514
-
521.
2007.
[7]
Ning
Zh
ang,
Junichi
T
at
emura
,
Ji
gnesh
Pat
el
,
and
Haka
n
Ha
ci
gum
us.
“
Re
-
ev
al
ua
ting
designs
for
m
ult
i
-
t
ena
n
t
OL
T
P
workloads
on
S
SD
-
base
dI/O
subs
y
stems
.
”
In
Pr
oce
ed
ings
of
the
2014
ACM
SIGMO
D
Inte
rnatio
nal
Confe
r
enc
e
on
Manage
ment
o
f D
ata
(SIG
MO
D
'
14)
.
ACM
,
New
York,
NY
,
US
A,
pp
.
1383
-
13
94
,
2014
.
[8]
F.
Farbe
r,
C
.
Mathi
s,
D.
D
.
Cul
p,
W
.
Klei
s
and
J.
Schaf
fne
r
.
“
An
in
-
m
emory
dat
ab
ase
s
y
st
em
for
m
ult
i
-
te
n
an
t
appl
i
ca
t
ions
.”
Pr
oc.
BTW
,
pp
.
65
0
-
666,
2015
[9]
Zhe
n
L
iu,
Mark
S.
Squill
an
te,
an
d
Joel
L
.
W
olf.
“
On
m
axi
m
iz
ing
service
-
le
v
el
-
ag
ree
m
ent
prof
it
s.
”
In
Proceedi
ngs
of
the
3rd
ACM
conf
ere
n
ce
on
Elec
troni
c
Comm
e
rce
(EC
'
01)
.
AC
M,
New
York,
NY
,
US
A,
vol.
29
,
no.
3
,
pp
.
213
-
223
,
De
ce
m
ber
2001
.
[10]
Y.
Chi,
H
.
J.
Mo
on,
and
H
.
Hac
ıg
üm
ü
.
“
s.
iCBS:
Inc
rement
a
l
Cost
-
base
d
Schedu
ling
under
Pie
ce
w
ise
Li
n
ea
r
SLAs
.
”
PV
LD
B
,
vo
l. 4,
no.
9
,
pp
.
563
–
5
74,
2011
.
[11]
Y.
Chi
,
H
.
J.
M
oon,
H.
Ha
cı
gü
m
ü¸s,
and
J.
T
atem
ura
.
“
SLA
-
Tr
ee
:
A
Fram
ewor
k
for
E
ffi
ci
en
tly
Supporting
SL
A
-
base
d
De
ci
sions
in
Cloud
Com
puti
ng.
”
In
EDBT
,
pp.
129
–
140
,
20
11
.
[12]
J.
Csirik
,
D.S.
Johns
on
.
“
Bounded
spa
ce
on
-
l
ine
bin
pa
cki
ng:
b
e
st
is
better
th
an
first.
”
In
Proc
eedings
of
th
e199
1
sec
ond
annual
A
CM/SIGACT
-
SIAM
,
San
Franc
isco,
vo
l.
31,
no.
2
,
pp
.
115
-
138
,
M
arc
h
1991
.
[13]
F.
W
ang,
J.
Li,
J.
Zh
ang
and
Q.
Huang,
“
Resea
r
ch
on
the
Mult
i
-
Te
nan
t
Pl
ac
eme
nt
Gen
et
i
c
Algo
rit
hm
Based
o
n
Euc
a
l
y
ptus
Pla
tf
orm
,
”
2016
12th
Inte
rnational
Co
nfe
renc
e
on
Com
putat
ional
Int
el
ligence
and
Se
cur
it
y
(
CIS)
,
W
uxi,
pp.
382
-
385
,
20
16
.
[14]
H.
Koziolek,
“
T
he
SP
OS
AD
Ar
chi
t
ec
tur
al
St
y
le
for
Mult
i
-
t
ena
n
t
Software
Appl
ic
a
ti
ons,
”
2011
Nint
h
Work
ing
IEE
E
/IF
IP
Conf
ere
nce on
So
ft
w
are
Architecture
,
Bould
er,
CO,
p
p.
320
-
327
.
201
1
.
[15]
Y.
Zh
ang,
Z
.
W
ang,
B.
Gao
,
C
.
Guo,
W
.
Sun
a
nd
X.
Li,
“
An
Eff
ective
Heur
isti
c
for
On
-
li
n
e
Te
nan
t
Pl
acem
ent
Problem i
n
SaaS
,
”
2010
IE
EE Int
ernati
onal
Confer
enc
e
on
We
b
Se
rvic
es
,
Miami,
FL,
pp
.
425
-
432.
2
010
.
[16]
Dann
y
Stro
cki
s
“
Microsoft
Azure
”
from
https
:/
/a
zu
re.
m
i
cro
soft.
com/en
-
us/r
esourc
es/sample
s/ac
ti
v
e
-
dir
ec
to
ry
-
dotne
t
-
web
app
-
m
ult
it
ena
n
t
-
ope
nidc
onne
ct
[17]
Y.
Xiao
y
ong
,
T.
Hong
y
an,
L
.
Yi
ng,
J.
Tong
,
L.
Ti
an
che
ng
and
W
.
Zhonghai,
“
A
Com
pet
it
iv
e
Penal
t
y
Mod
el
f
or
Avail
ab
il
ity
B
ase
d
Cloud
SLA,
”
2015
IE
EE
8th
I
nte
rnational
Con
fe
renc
e
on
Cloud
Computing
,
Ne
w
York
Cit
y
,
N
Y
,
pp.
964
-
970
,
20
15
Evaluation Warning : The document was created with Spire.PDF for Python.