Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
10
,
No.
2
,
A
pr
il
2020
, p
p.
15
07
~
1514
IS
S
N:
20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v
10
i
2
.
pp1507
-
15
14
1507
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om/i
nd
ex
.ph
p/IJ
ECE
Perform
ance ana
lysis of
contain
er
-
based n
etworki
ng
Solutions for hi
gh
-
perform
ance c
ompu
ting
cloud
Sa
n
g
B
oem Li
m
1
, Joon
Woo
2
, G
u
oh
u
a
Li
3
1
Depa
rt
m
ent
of S
m
art
ICT
Conv
erg
ence, Konkuk
Univer
sit
y
,
Ko
rea
2
,3
Nati
ona
l
Insti
t
ute
of
Superc
om
puti
ng
and
Ne
tworking,
Kor
ea
In
stit
ute of
Sc
ie
n
c
e
and
T
ec
hnolog
y
Inform
at
ion
,
K
ore
a
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
J
ul
4
,
2019
Re
vised
Oct
11
,
20
19
Accepte
d
Oct
20
, 20
19
Rec
en
tly
,
cl
oud
servic
e
provide
r
s
have
be
en
gr
a
dual
l
y
cha
ng
ing
from
virt
ual
m
ac
hine
-
base
d
cl
oud
infra
st
ruc
ture
s
to
c
onta
in
er
-
base
d
cl
oud
-
n
at
i
v
e
infra
struc
ture
s
t
hat
consid
er
per
form
anc
e
and
workload
-
m
ana
ge
m
ent
issues
.
Sever
al
da
ta
netw
ork
per
form
an
ce
issue
s
for
v
irtual
insta
nc
es
h
av
e
ar
isen,
and
var
ious
ne
twor
king
solut
ions
have
b
ee
n
ne
wl
y
develope
d
or
u
ti
l
iz
ed
.
In
thi
s
pap
er,
we
propose
a
solut
io
n
suita
bl
e
for
a
hi
gh
-
per
form
anc
e
computing
(HP
C)
cl
oud
thr
ough
a
per
form
anc
e
compari
son
ana
l
y
sis
of
containe
r
-
b
ase
d
net
working
solu
t
ions.
W
e
constru
ct
ed
a
supe
rco
m
pute
r
-
base
d
te
st
-
bed
c
luste
r
to
ev
al
ua
te t
he
s
erv
icea
bi
li
t
y
b
y
exe
cu
ti
ng
HP
C
j
obs.
Ke
yw
or
d
s
:
Cl
oud
c
om
pu
ti
ng
Con
ta
ine
r
-
base
d netw
orkin
g
HP
C
Netw
ork
Perfo
rm
ance an
al
ysi
s
Copyright
©
202
0
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Guo
hu
a
Li
,
Nati
on
al
I
ns
ti
tute o
f Su
per
c
om
pu
ti
ng
and
N
et
work
i
ng
,
Korea
Insti
tute o
f
Scie
nce a
nd Tech
nolo
gy In
form
ation
,
245 Dae
hak
-
r
o, Y
us
e
ong
-
gu,
Daej
e
on, 3
4141, K
or
ea
.
Em
a
il
:
gh
le
e@
kisti
.r
e.
kr
1.
INTROD
U
CTION
Trad
it
io
nally
,
high
-
perf
or
m
a
nce
c
om
pu
ti
ng
(
HP
C)
was
use
d
m
ai
nly
by
nat
ur
al
sci
enc
e
areas
li
ke
weathe
r
f
or
eca
sti
ng
,
m
olecular
bio
l
og
y,
a
nd
s
pace
ex
plo
it
at
ion.
HP
C
becom
es
ve
ry
dem
a
nd
i
ng
te
c
hnology
du
e
to
the
im
po
rtance
of
bi
g
data
analy
sis,
that
cannot
be
proces
se
d
with
the
t
rad
it
io
nal
co
m
pu
ti
ng
env
i
ronm
ent
[
1].
A
wi
de
ra
nge
of
c
om
pu
te
r
arc
hitec
ture
is
,
al
s
o,
re
quire
d
to
process
big
data.
O
ne
s
olu
t
ion
to
these
requirem
ents
is
to
a
dd
c
loud
ca
pa
bili
ty
to
the
HP
C
e
nvir
on
m
ent.
Cl
oud
c
om
pu
ti
ng
can
e
asi
ly
ada
r
t
ra
pid
changes
of
hardw
a
re
an
d
s
ofr
tware
te
ch
olog
ie
s.
By
us
in
g
c
loud
com
pu
ti
ng,
m
os
t
of
t
he
us
ers
al
s
o
can
reduce
analy
sis t
i
m
e and cost
of
hard
war
e
and s
oft
w
are [2].
The
desi
gn
of
netw
orks
in
a
cl
oud
inf
ras
tru
ct
ure
c
onfi
gurati
on
is
usual
ly
div
i
ded
into
public,
m
anag
em
ent,
and
guest
net
works
[
3].
A
gu
e
st
netw
ork
is
capab
le
of
data
com
m
un
ic
at
ion
betwee
n
vi
rtual
instances
wh
ic
h
are
run
ning
on
one
host,
m
ulti
ple
hosts,
or
acr
os
s
the
dif
fer
e
nt
sub
nets
[4
-
6
]
.
T
he
Do
c
ker
[
7
]
con
ta
ine
r
platf
or
m
is
an
open
-
s
ource
co
nt
ai
ner
m
anag
e
m
ent
pro
j
ect
la
un
c
hed
by
Do
c
ke
r
I
nc.
This
is
a
li
gh
twei
ght
c
on
ta
ine
r
te
ch
nolo
gy
that
bundle
s
an
d
r
uns
the
ser
vice
ope
r
at
ing
en
vir
onm
ent.
Wh
e
n
co
nf
igurin
g
a
cl
oud
e
nvir
onm
ent
based
on
the
D
oc
ker
c
on
ta
ine
r,
an
or
chestrati
on
s
oft
war
e
s
uc
h
a
s
Kube
rn
et
es
or
Do
c
ke
r
Sw
arm
is nee
de
d
to
ef
fecti
vel
y
m
anag
e a
nd
eff
ic
ie
ntly
all
oc
at
e the r
e
sour
ces re
qu
i
red f
or c
on
ta
ine
rs [
8
-
10
]
.
The
go
al
of
thi
s
pa
per
is
t
o
e
valuate
a
s
uitable
co
ntaine
r
-
ba
sed
netw
ork
s
olu
ti
on
f
or
H
P
C
cl
oud
by
perform
ing
be
nch
m
ark
te
sts
us
in
g
Me
ssa
ge
Passin
g
I
nterfac
e
(MP
I)
be
nc
hm
ark
su
it
e.
We
hav
e
te
ste
d
se
ver
a
l
netw
orkin
g
so
l
utions
us
in
g
K
ub
e
r
netes
wh
ic
h
has
an
e
xcell
ent
a
uto
-
rec
overy
ca
pab
il
it
y.
This
st
ud
y
incl
ud
e
s
a
res
ult
of
the
perf
or
m
an
ce
te
sts
on
netw
ork
band
width
of
va
rio
us
cl
us
t
er
net
wor
k
c
onfi
gurati
ons
a
nd
an
evaluati
on
of
t
he
HP
C
se
rv
ic
e
abili
ty
of
the
ba
re
-
m
et
al
and
c
on
ta
ine
r.
A
dd
it
ion
al
ly
,
a
s
umm
ary
of
the
eva
luati
on
resu
lt
a
nd r
ec
om
m
end
at
ion
on c
on
ta
ine
r
-
bas
ed netw
ork
s
ol
ution i
s
prov
i
de
d.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
2
,
A
pr
i
l
2020
:
1505
-
1514
1508
2.
R
EL
ATED
W
ORKS
2.1.
C
onta
in
e
r netw
orkin
g on a sin
gle
node
Ther
e
a
re
tw
o
ty
pes
of
c
on
t
ai
ner
-
base
d
ne
tworki
ng
s
olu
t
ion
s
that
a
re
r
el
at
ed
to
a
sin
g
no
de
an
d
m
ul
ti
ple
nodes
.
Re
prese
ntati
ve
netw
orki
ng
s
olu
ti
ons
for
c
onta
iners
on
a
s
ing
le
node
are
div
i
de
d
int
o
t
hr
e
e
cat
agories:
bri
dg
e
netw
orkin
g,
host
netw
orkin
g,
an
d
m
acvlan
net
work
i
ng
[
11
]
.
In
the
bri
dge
netw
orki
ng
[
12
]
,
a
docker
0
or
use
r
-
def
i
ned
br
i
dge
is
cr
eat
ed
th
rou
gh
the
ph
ysi
cal
netw
ork
int
erf
ace
to
c
on
tr
ol
the
tra
ff
ic
be
tween
con
ta
ine
rs
in
t
wo
nam
espaces
on
a
si
ng
le
ho
st
Fig
ure
1
(
a
)
.
A
host
pro
cess
[
1
3
]
,
s
uc
h
as
t
he
ss
hd
da
e
m
on
,
is
create
d
wit
h
a
s
pecific
port
w
hile
an
othe
r
m
ai
n
ser
vice
da
e
m
on
i
n
a
co
ntainer
is
c
reated
with
a
nothe
r
port
hav
i
ng
th
e
sam
e
vi
rtual
IP
a
ddress
i
n
t
he
sa
m
e
nam
espace
f
or
th
e
host
ne
tworki
ng
Fi
gure
1
(
b
)
.
The
m
acvlan
netw
orkin
g
[
1
4
]
dev
i
des
the
physi
cal
netw
ork
inter
face
with
Macvl
an
tag
s
c
orres
pondin
g
to
co
ntainers
i
n
diff
e
re
nt n
am
espaces
Fig
ur
e
1
(
c
)
.
Figure
1.
Co
ntainer
netw
orki
ng so
l
utio
ns
on one
host [
15
]
2.2
.
C
onta
in
e
r netw
orkin
g on
m
ultipl
e n
od
es
A
n
overlay
ne
twork
[
1
6
]
i
s
the
m
os
t
com
m
on
ly
us
ed
c
on
ta
ine
r
ne
twork
on
m
ulti
ple
nodes
.
An
ove
rlay
net
work
c
onfig
urat
ion
is
nee
ded
to
al
lo
w
c
onta
iners
t
o
c
omm
un
ic
at
e
with
ea
ch
oth
e
r
on
m
ulti
ple
nodes
[
17
-
19
]
.
Re
pr
e
sentat
iv
e
overlay
netw
orkin
g
s
olu
ti
ons
for
c
onta
in
ers
a
re
div
ide
d
into
t
wo
cat
agories:
li
nu
x
bri
dge
overlay
netw
ork
an
d
Fla
nn
el
ne
tworki
ng.
D
oc
ker
O
ver
la
y
Netw
ork
a
nd
Weav
e
Net
ne
tworki
ng
so
luti
ons
us
e
the
Li
nux
B
rid
ge
dr
i
ver
as
a
tunnel
inter
fac
e
to
form
a
tun
nel
f
or
the
ne
twork
traf
fic
be
tween
con
ta
ine
rs
on
diff
e
re
nt
hosts,
as
s
how
n
in
F
igure
2
(
a
)
.
W
e
ave
Net
c
reates
it
s
ow
n
Wea
ve
B
rid
ge
as
a
virtu
a
l
router
cal
le
d
v
Ro
uter
.
Fla
nn
e
l
networ
k
is
de
velo
ped
sp
eci
fical
ly
fo
r
Kubernete
s,
a
nd
is
easy
to
co
nfi
gur
e
.
Figure
2
(
b
)
shows
t
hat
it
c
orrespo
nd
s
t
o
t
he
P
OD
str
uctu
r
e
of
K
uber
nete
s
a
nd
re
fer
e
nc
es
the
Ro
utin
g
Ta
ble
thr
ough the
do
cker0
i
nterf
ace
v
ia
fl
annel
d
.
Figure
2. Co
ntainer
netw
orki
n
g d
olu
ti
ons
on m
ulti
ple n
od
es [
15
]
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N:
20
88
-
8708
Perf
orma
nce
analysis
of c
on
t
ain
er
-
base
d ne
tw
or
ki
ng
Soluti
on
s
for
hi
gh
…
(
Sang B
oem L
im
)
1509
C
al
ic
o
netw
orkin
g
[
20
]
is
a
netw
orkin
g
s
olu
ti
on
t
hat
is
optim
iz
ed
fo
r
t
he
native
cl
ou
d
and
sim
ple,
scal
able
secu
r
it
y
te
chn
ology
.
U
nlike
oth
e
r
m
ulti
-
ho
st
ov
e
rlay
net
work
i
ng
so
l
utions
ex
plaine
d
a
bove
,
it
us
es
Ca
li
co’
s
own
dr
i
ver
in
s
te
ad o
f t
he Li
nux B
rid
ge ke
r
ne
l
dr
ive
r.
As s
ho
w
n i
n
Fig
ur
e
3,
tra
ff
ic
th
r
ough
al
l
the conta
ine
rs
i
s rou
te
d
acc
ord
ing
t
o
the
in
-
ke
rn
el
ru
le
by
util
iz
ing
the
f
ire
w
al
l functi
ons.
Fig
ure
3
.
Cal
ic
o netw
orkin
g
[
15
]
3.
PERFO
R
MANC
E
A
NA
L
Y
SIS
3.1.
B
an
dwi
d
th
test
Fo
r
the
ben
c
hm
ark
en
vir
on
m
ent,
we
are
usi
ng
a
cl
us
te
r
s
pecified
i
n
Ta
ble
1
with
Ce
nt
OS
ope
rati
ng
syst
e
m
.
On
e
c
on
t
ro
ll
er
no
de
and
f
our
w
ork
nodes
are
assig
ned.
Am
on
g
t
he
f
our
wor
k
nodes,
tw
o
nodes
ha
ve
7Gbyte
s
of
m
em
or
y
and
t
welve
c
or
es
a
nd
the
oth
e
r
tw
o
node
s
hav
e
15
Gb
yt
es
of
m
e
m
or
y
and
ei
ght
cor
e
s
.
As
s
pecified i
n Table
2, a n
et
work
ba
ndwidt
h
te
st i
s p
er
for
m
ed
wh
il
e cha
ng
i
ng
t
he
net
w
ork
c
onfig
ur
at
i
on
a
nd
the
cl
us
te
r
co
nfi
gurati
on
at
t
wo
ba
re
-
m
et
al
nodes
co
ns
ti
tuti
ng
a
10
G
Et
he
rn
et
netw
ork.
The
D
ock
e
r
ne
twork
config
ur
at
io
n
i
s
div
ide
d
into
Do
c
ke
r
Li
n
ux
Ov
e
rlay
,
Wea
ve
Net,
Fla
nn
e
l,
Ca
li
co
with
IP
-
in
-
IP
,
a
nd
Ca
li
co
without
the
I
P
-
in
-
I
P
c
onfig
ur
a
ti
on
.
The
cl
us
t
er
c
onfig
ur
at
io
n
is
de
vid
e
d
in
to
et
cd
f
or
D
oc
ker
Lin
ux
O
ve
rlay
,
Kube
rn
et
s
for
Weav
e
Net,
Flann
el
,
Ca
li
co
with
I
P
-
in
-
I
P,
and
Ca
li
co
without
the
IP
-
in
-
IP
c
onfig
ur
at
io
n,
a
nd
10G
et
hernet f
or b
a
re
-
m
et
al
.
Table
1.
T
est
-
be
d
cl
ust
er spec
ific
at
ion
No
d
e
CPU
Co
re
Me
m
o
r
y
Co
n
troller
Intel(R)
Xeo
n
(R)
CPU E55
3
0
@
2.4
0
GHz
8
16G
W
o
rk n
o
d
e 1
Intel(R)
Xeo
n
(R)
CPU E56
4
5
@
2.4
0
GHz
1
2
7G
W
o
rk n
o
d
e 2
Intel(R)
Xeo
n
(R)
CPU E56
4
5
@
2.4
0
GHz
12
7
G
W
o
rk n
o
d
e 3
Intel(R)
Xeo
n
(R)
CPU
E56
2
0
@
2.4
0
GHz
8
1
5G
W
o
rk n
o
d
e 4
Intel(R)
Xeo
n
(R)
CPU E56
2
0
@
2.4
0
GHz
8
1
5G
Table
2
.
T
est
-
be
d netw
ork
c
onfi
gurati
on
Do
ck
er
Netwo
rk Co
n
f
i
g
u
ration
Clu
ster Co
n
f
ig
u
rat
io
n
1
Do
ck
er
Linu
x
Ov
e
rlay
etcd
2
W
eave
Net
Ku
b
ernetes
3
Flan
n
el
4
Calico
(
with
IP
-
in
-
IP)
5
Calico
(
with
o
u
t I
P
-
in
-
IP)
6
Bare
-
m
etal
Ho
st n
etwo
rk with
10
G
Ethern
et
Fo
r
th
e
ba
ndw
idth
te
st,
we
ar
e
us
in
g
Ip
e
rf3
[
1
4
]
ba
ndwi
dth
to
ol.
Ip
e
rf3
i
s
a
ba
ndwi
dth
m
easur
m
ent
too
l
to
m
easure
the
m
axi
m
u
m
achievable
band
width
on
IP
netw
ork.
A
s
show
n
i
n
Fi
g
ure
4
,
Ca
li
co
without
the
IP
-
in
-
IP
c
onfig
ur
at
io
n
pe
rfor
m
s
m
uch
bette
r
tha
n
ot
her
overlay
ne
tworks
.
U
pon
te
sti
ng
with
Ip
e
rf3
,
the
ba
ndwi
dth
of
t
he
Ca
li
co
overlay
net
wor
k
with
ou
t
IP
-
in
-
IP
is
8,500
M
bi
ts/
s,
wh
ic
h
is
as
high
as
t
he
value
for
bar
e
m
et
al
that
pe
rfor
m
s
9,3
00
M
bits/
s.
W
e
belie
ve
t
his
res
ult
is
be
cause
Ca
li
co
use
s
the
Ca
li
co
dr
i
ve
r
wh
ic
h
is
loa
de
d
i
nto
the
ke
r
ne
l,
but
oth
e
r
overlay
s
olu
ti
on
s
use
t
he
def
a
ul
t
Linux
Brid
ge
dri
ve
r
i
n
t
he
kernel.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
2
,
A
pr
i
l
2020
:
1505
-
1514
1
510
Howe
ver,
with
the
IP
-
in
-
IP
config
ur
at
io
n,
Ca
li
co
ex
hi
bits
a
band
width
of
4,600
Mbi
ts/
s,
that
is,
ha
lf
of
8,500
Mb/s.
For
t
his
reas
on,
in
cr
os
s
dif
fer
e
nt
s
ubnet
en
vir
onm
ents,
the
net
work
pe
rfor
m
ance
is
50%
of
that
of
bar
e
m
et
a
l i
n
any
one
sub
net
env
i
ronm
ent.
Fig
ure
4
.
Ba
nd
width t
est
for n
et
work
i
ng so
l
ut
ion
3.2.
Thr
ough
put
tes
t
Ba
sed
on
this
pe
rfor
m
ance
c
om
par
ison
,
w
e
s
el
ect
ed
Ca
li
co
to
c
onstr
uct
a
c
loud
e
nvir
on
m
ent
for
H
PC
serv
ic
es
.
The
perform
ance
ve
rificat
ion
of
t
he
en
vir
on
m
e
nt
i
n
w
hich
H
PC
se
rv
ic
es
c
an
be
pro
vid
e
d
was
com
par
ed
with
the
m
easur
ed
th
r
oughput
of
TCP/
U
D
P
[
21
]
us
in
g
a
H
PC
pe
rfor
m
ance
be
nc
hm
ark
too
l
–
HP
CB
E
NCH
[
22
]
.
Th
is
te
st
co
m
par
es
bar
e
m
et
al
,
Ca
li
co,
and
C
anal
w
hich
inc
orp
or
at
es
the
Ca
li
c
o
secur
it
y
poli
cy
f
or
Flan
nel.
A
ll
the
te
sts
were
done
after
set
ti
ng
MTU
=
9000
on
al
l
dev
ic
e
s
s
uch
as
t
he
s
witc
h,
ho
st
,
a
nd
c
onta
iner.
We
c
ho
se
Fla
nn
el
f
rom
sever
al
net
work
i
ng
s
olu
ti
on
s
t
hat
us
e
t
he
Lin
ux
B
rid
ge
ke
r
nel
dr
i
ver
be
ca
us
e
it
is
easy
to
be
integrate
d
with
K
uber
netes
a
nd
c
onfi
gure
with
Ca
li
co
’s
secu
rity
poli
cy
as
C
anal
netw
orkin
g.
T
her
e
fore,
us
i
ng
only
Ca
li
co
a
nd
Ca
nal,
we
c
an
al
s
o
ac
qu
ire
get
the
be
st
re
su
lt
s
with
out
te
sti
ng
oth
e
r netw
orking so
l
utio
ns
.
As
sho
wn
i
n
Figure
5,
the
TC
P
t
hroug
hput
of
Ca
li
co
is
e
qual
to
80
%
of
th
at
of
ba
re
m
et
a
l,
bu
t
Ca
na
l
is
eq
ual
to
onl
y
20%
of
the
bar
e
m
et
al
val
ue.
F
or
U
DP
,
Ca
li
co’
s
th
r
oughput
is
e
qual
to
70
%
of
that
of
bar
e
m
et
al
, b
ut for
Ca
nal, it i
s equ
al
to
only
30%
.
I
n ad
diti
on, t
he
loss
-
rate
of Cal
ic
o’
s
UDP is
alm
os
t 0%, w
hi
ch
is
m
or
e stable
tha
n
that
of
bar
e
-
m
et
al
. Canal ha
s a
25
%
loss
-
r
at
e w
hic
h
is i
nferior i
n
te
rm
s o
f
sta
bili
ty
.
Fig
ure
5
.
TCP/
UDP th
rou
ghput t
est
(
M
b/s)
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N:
20
88
-
8708
Perf
orma
nce
analysis
of c
on
t
ain
er
-
base
d ne
tw
or
ki
ng
Soluti
on
s
for
hi
gh
…
(
Sang B
oem L
im
)
1511
4.
EVAL
UA
TI
O
N
To
e
valuate
the
HP
C
ser
viceab
il
ity
of
the
s
el
ect
ed
Ca
li
co
net
work
i
ng,
we
c
om
par
ed
it
with
bar
e
m
et
a
l,
Singularit
y,
ke
y
-
value
st
or
e
net
wor
k,
an
d
s
war
m
.
Sin
gi
la
rity
sh
ari
ng
the
ho
st
net
works
a
nd
w
hich
is
sp
eci
fical
ly
de
sign
e
d
for u
se
with
H
PC
ser
vi
ces
[
23
-
25
]
.
We
c
on
st
ru
ct
e
d t
w
o
nodes
for r
unni
ng
HP
L
[
26
]
that
is
a
porta
ble
i
m
ple
m
entat
ion
of
the
hi
gh
-
pe
rfor
m
ance
Lin
pack
be
nc
hm
ark
f
or d
ist
rib
uted
-
m
e
m
or
y
co
m
pu
te
rs.
In
order
to
run
H
PL
accu
ratel
y,
we
need
t
o
set
the
pa
rtit
ion
in
g
blo
c
ki
ng
factor
(N
B
)
a
nd
t
he
m
e
m
or
y
us
a
ge
factor
(NS).
In
this
e
valuati
on
i
n Fi
gure
6
,
we
te
ste
d
va
rio
us
conditi
ons
us
i
ng
NS
a
nd
NB
f
act
or
.
We
are
usi
ng
NS
facto
rs
of
80%
a
nd
90
%,
w
hich
m
e
ans
that
HP
L
us
es
up
to
8
0%
a
nd
90%
of
a
vaila
ble
m
e
m
or
y.
Fo
r
portio
ning
blo
c
k
siz
e,
we
are
us
in
g
64
by
te
s,
128
byte
s
,
an
d
25
6
byte
s.
T
he
te
st
r
es
ul
t
sh
ows
ba
re
m
et
al
,
the Doc
ker co
nt
ai
ner
with Cal
ic
o
an
d K
ub
e
r
netes, a
nd Si
ngularit
y offer
si
m
il
ar levels o
f
perform
ance.
Fig
ure
6
.
MPI
band
width t
est
resu
lt
s
We
co
ns
tr
uct
ed
tw
o
no
des
for
r
unni
ng
par
al
le
l
MPI
ta
sk
s
to
c
he
ck
the
band
width
us
i
ng
the
osu
-
m
ic
ro
-
ben
c
hm
ark
s
MPI
be
nch
m
ark
to
ol
[
2
7
]
Figure
7
.
T
he
te
st
re
su
lt
s
s
how
that
ba
re
m
et
a
l,
the
Do
c
ker
c
onta
iner
with
C
al
ic
o
an
d
K
ubern
et
es,
a
nd
Si
ngularit
y
al
l
of
fer
sim
il
ar
le
ve
ls
of
perform
ance
i
n
te
rm
s
of
ba
ndwidth
a
nd
la
te
ncy.
I
n
a
ddit
ion
,
we
c
onstruc
te
d
four
nodes
(
with
72
co
re
s)
for
r
unn
i
ng
par
al
le
l
MPI
ta
s
ks
to
check
t
he
M
PI
al
l
-
to
-
al
l
pe
rs
on
al
iz
ed
e
xc
ha
ng
e
la
te
ncy
te
st
Fig
ure
8
an
d
t
he
MPI
al
l
gath
e
r
per
s
onal
iz
ed
e
xch
a
nge
la
te
nc
y
te
st
Fig
ur
e
9
w
hich
inc
ur
s
the
hi
gh
e
st
com
m
un
ic
at
ion
l
oad
am
on
g
al
l
the
f
unct
ion
s
pro
vid
e
d
by
the
ben
c
hm
ark
to
ol
.
The
te
st
res
ult
sh
ows
t
hat
ba
re
m
et
al
,
the
Do
cke
r
c
onta
ine
r
wit
h
Ca
li
co
and
K
uber
netes,
a
nd
Singularit
y
al
l
offer
th
e
sam
e
le
vels
of
la
te
nc
y
per
f
or
m
anc
e
even
i
n
a
m
ulti
ple
-
node st
ru
ct
ur
e
.
Fig
ure
7
.
MPI
band
width t
est
resu
lt
s
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
2
,
A
pr
i
l
2020
:
1505
-
1514
1512
Figure
8
.
MPI
al
l
-
to
-
al
l pe
rs
onal
iz
ed
e
xc
hange lat
ency te
st
resu
lt
s
Figure
9
.
MPI
al
lgather
pe
rsonali
zed e
xc
hange lat
ency te
st
resu
lt
s
5.
CONCL
US
I
O
N
In
t
he
prese
nt
s
tud
y,
we
a
ddre
ssed
c
on
ta
i
ner
netw
orkin
g
s
ol
utions
on
m
ultip
le
hosts
f
or
H
PC
ser
vices.
Ba
sed
on
t
he
res
ults
of
these
te
sts,
we
be
li
eve
that
Ca
li
co
offe
rs
the
bes
t
perform
ance
wh
il
e
a
c
om
par
at
ive
an
al
ysi
s
re
ve
al
ed
t
hat
it
is
the
easi
est
m
ean
s
of
c
onfig
ur
i
ng
a
n
H
PC
e
nvir
onm
ent.
The
m
ai
n
con
t
rib
ution
of
t
he
pr
ese
nt
stu
dy
was
the
te
sti
ng
of
t
he
pe
rfor
m
ance
of
a
real
su
pe
rc
om
pu
te
r
-
base
d
HP
C
cl
us
te
r.
Ba
sed
on
our
perform
ance
com
par
ison
a
na
ly
sis,
we
ha
ve
propose
d
the
best
co
ntainer
-
base
d
netw
orkin
g
s
ol
ution
f
or
H
PC
ser
vices,
at
ta
inin
g
e
xcell
ent
resu
lt
s
w
hich
are
c
om
par
abl
e
with
t
hose
f
or
ba
re
m
et
al
.
In
the
f
uture,
we
will
determ
ine
the
serv
ic
e
value
by
exec
utin
g
ne
twork
-
inten
sive
par
al
le
l
jo
bs
with
wh
ic
h we ca
n evaluat
e
our fi
nd
i
ngs.
ACKN
OWLE
DGE
MENTS
This
stu
dy
was
perform
ed
as
a
su
bp
roject
of
t
he
K
IS
T
I
pro
j
e
ct
entit
le
d
“
The
Nati
onal
S
up
e
rco
m
pu
ti
ng
Infr
ast
ru
ct
ur
e
Con
st
ru
ct
io
n
a
nd Se
rv
ic
e
[K
-
19
-
L
02
-
C01
-
S
01
]
”.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N:
20
88
-
8708
Perf
orma
nce
analysis
of c
on
t
ain
er
-
base
d ne
tw
or
ki
ng
Soluti
on
s
for
hi
gh
…
(
Sang B
oem L
im
)
1513
REFERE
NCE
S
[1]
B
.
Gourav
,
et
a
l
.
,
"A
Novel
Ap
proa
ch
for
C
luste
ring
Big
Da
ta
base
d
on
MapR
educ
e
,
"
Inte
rnat
ional
Journal
o
f
El
e
ct
rica
l
and
C
omputer
Engi
n
e
erin
g
,
vo
l. 8, no. 3, pp. 1711
-
171
9,
Jun 2018
.
[2]
R.
Rajak,
"
A
C
om
par
at
ive
Stud
y
:
Ta
xonom
y
of
High
Perform
anc
e
Com
puti
ng
(
HP
C),
"
Inte
rnati
o
nal
Journal
o
f
El
e
ct
rica
l
and
C
omputer
Engi
n
e
erin
g
,
vo
l. 8, no. 5, pp. 3386
-
339
1,
Oct
2018.
[3]
A.
Cel
est
i,
et
al
.
,
"Ev
al
ua
ti
ng
A
lt
ern
at
iv
e
Daa
S
Soluti
ons
in
Pri
vat
e
and
Public
OpenStac
k
Clo
uds."
Soft
ware
-
Pract
i
ce
and
E
x
perie
nc
e
,
vol. 47
,
no
.
9
,
pp
.
1185
-
120
0,
Sep
2017
.
[4]
K.
Suo,
Y.
Zha
o
,
W
.
Ch
en
and
J.
Rao,
"A
n
Anal
y
s
is
and
Empiri
ca
l
Stud
y
of
Con
ta
in
er
Ne
tworks,"
IE
EE
IN
FOCOM
2018
-
IE
EE Co
nfe
renc
e
on
Co
mputer
Comm
unic
ati
ons
,
Honolu
lu,
HI
,
2018
,
pp
.
189
-
197.
[5]
Y
oungki
Park,
H
y
unsik
Yang,
Younghan
Kim
,
"P
erf
orm
an
ce
C
om
par
ison
of
Co
nta
in
er
N
et
worki
ng
T
ec
hnolog
ie
s
",
Journal
of
Kore
a
Information
an
d
Comm
unic
ati
o
ns Soc
iety
,
vol
.
4
4,
no
.
01,
pp.
015
8
-
0170,
Jan
201
9
[6]
P
asqua
le
Sa
lza,
Filom
ena
Ferru
c
ci
,
"S
pee
d
Up
Ge
net
i
c
Algori
thms
in
the
Cloud
Us
ing
Software
Con
ta
in
ers",
J
ournal
of
Fu
ture
Gen
erati
on
Comput
er
Syste
ms
,
vol
.
92
,
pp.
672
-
681,
Se
p
2019
[7]
M.
De
B
ene
di
cti
s
,
et
al
.
,
"Inte
g
rity
V
eri
fi
cation
o
f
Docke
r
Contain
ers
for
a
Li
ghtwe
ight
Clo
ud
Env
ir
onm
ent
.
"
Fu
ture
Gene
ration
Com
pute
r Sy
st
ems
,
v
ol.
97
,
pp
.
236
-
2
46,
Aug 2019
.
[8]
M.
K.
Hus
sein,
et
a
l
.
,
"A
Pla
cem
ent
Archi
t
ectu
re
for
a
Con
ta
in
er
as
a
Serv
ic
e
(
Caa
S)
in
a
C
lou
d
Envi
ronm
ent."
Journal
of
Clou
d
Computing
,
vo
l.
8
,
no
.
1
,
De
c
2
019.
[9]
Max
Alauna,
Er
i
c
Vial,
Nuno
Ne
ves,
Ferna
ndo
M.V.
Ramos
,
"S
ecure
Multi
-
C
loud
Network
Virtual
iz
a
ti
on",
Journal
of
Computer
N
etwor
ks
,
vol. 161,
pp.
45
-
60
,
Oc
t
2
019
[10]
M
arc
o
De
Ben
edi
c
ti
s,
Anton
io
Li
o
y
,
"Int
egr
ity
V
eri
fi
ca
t
ion
of
Docke
r
Conta
in
ers
for
a
L
ight
wight
C
lou
d
Envi
ronm
ent
",
J
ournal
of Fut
ure
Gene
ration
Co
mputer
Syste
ms
,
vol.
97
,
pp
.
236
-
246,
A
ug
2019
[11]
C.
R
amon
-
Cortes
,
e
t
al.
,
"Tra
nsp
are
nt
Orch
estra
t
i
on
of
Ta
sk
-
B
ase
d
Para
ll
e
l
Appli
ca
t
ions
in
Cont
a
ine
rs
Pl
at
form
s."
Journal
of
Gr
id Com
puti
ng
,
vo
l.
16,
no
.
1
,
pp
.
13
7
-
160,
Feb
2018
.
[12]
.
Buh
,
et
al.
,
"A
dapt
ive
Netwo
rk
-
Tra
ffi
c
B
alan
ci
ng
on
Mult
i
-
Core
So
ftwa
r
e
Networki
ng
De
vic
es,
"
Compute
r
Net
works
,
vo
l. 6
9,
pp
.
19
-
34
,
Au
g
2014.
[13]
Z.
Zh
ang,
et
a
l.
,
"La
rk:
An
Ef
fe
ct
iv
e
Approac
h
for
Software
-
De
fine
d
Networki
n
g
in
High
Throu
ghput
Com
puti
ng
Cluste
rs,"
Fu
ture
Gene
rat
ion
Co
mputer
Syste
ms
,
vol.
72
,
pp
.
105
-
117,
Ju
l
2017
.
[14]
J.
Stru
y
e,
et
al.
,
"A
ss
essing
the
Value
of
Con
ta
i
ner
s
for
NF
Vs
:
A
Deta
iled
Ne
t
work
Perform
anc
e
Stud
y
,
”
in
13
th
Inte
rnational
Co
nfe
renc
e
on
Networ
k
and
Ser
vi
c
e
Manage
m
ent
,
CNSM
2017
,
20
17,
pp
.
1
-
7.
[15]
J.
La
ng
emak, "Docker
N
et
workin
g
Cookbook,
"
Packt P
ubli
shing,
2016.
[16]
J.
Zha
ng
,
et
a
l.
,
"O
n
Achie
ving
t
he
Maximum
Strea
m
ing
Ra
te
in
H
y
brid
W
ir
ed/W
ire
le
ss
Overl
a
y
Ne
tworks,"
IE
EE
Wirel
ess Comm
unic
ati
ons
Letters
,
vol
.
8
,
no
.
2
,
pp
.
472
-
475
,
Apl
2
019.
[17]
S
andr
ine
Va
ton,
Olivi
er
Brun,
Maxime
Mouche
t,
Pablo
B
el
z
ar
ena
,
Isabe
l
Am
igo,
Balakrishna
J.
Prahbhu,
Thi
e
r
r
y
Chonave
l
,
"J
oin
t
Minim
iz
a
ti
on
of
Monitori
ng
Cost
and
Del
a
y
in
Overl
a
y
Ne
tworks:
Optimal
Polic
i
es
with
a
Markovia
n
Appr
oac
h",
Journal
o
f
Ne
twork
and
S
yste
ms
Manage
ment
,
vo
l. 27, i
ss
ue
1,
pp.
188
-
23
2,
Jan
2019
[18]
Min
-
Ho
Ha,
Zali
Yang
,
Jasm
in
e
Siu
Lee
L
am,
"P
ort
Perform
a
nce
in
Cont
ai
n
e
r
Tra
nsport
Log
isti
cs:
A
m
ult
i
-
stake
hold
er
Pers
pec
t
ive
",
Journa
l
of
Tr
ansfport Policy
,
vol. 73, pp
.
25
-
40,
Jan
2019
[19]
Tuka
sz
Makow
ski,
Paola
Gros
so,
"Eva
lu
at
ion
of
Virtua
l
ization
and
Tr
aff
ic
Filt
eri
ng
M
et
ho
ds
for
Conti
aner
Networks",
Jour
nal
of
Fut
ure
Ge
nerati
on
Comput
er
Syste
ms
,
vol
.
93,
pp
.
345
-
357
,
Apr 2019
[20]
A.
V.
Bara
nov
,
et
a
l.
,
"M
e
thods
of
Jobs
Con
ta
i
ner
izati
on
for
Superc
o
m
pute
r
W
orkloa
d
Mana
ger
s,"
Lobache
v
skii
Journal
of
Math
emati
cs
,
vol
.
40
,
no.
5,
pp.
525
-
5
34,
Ma
y
2019.
[21]
S.
F.
Sz
il
a
ag
y
i
,
et
al
.
,
"Throug
hput
Perform
an
ce
Com
par
ison
of
MP
T
-
GRE
a
nd
MP
TCP
in
t
he
Fast
Et
h
ern
e
t
IPv4/IPv6
Envi
ronm
ent
,
"
Journal
of
Telecomm
unic
ati
o
ns
and
Inf
orm
ati
on
Technol
ogy
,
vo
l.
2018
,
no.
2,
pp
.
53
-
5
9
,
2018.
[22]
B.
Huang
,
e
t
al
.
,
"H
pcb
enc
h
-
A
Li
nux
-
Based
Network
Benc
h
m
ark
for
High
Perform
anc
e
Ne
tworks,"
in
19
t
h
Inte
rnational
Sy
mpos
ium
on
Hig
h
Pe
rform
anc
e
Computing Syst
e
ms
and
Applicati
ons, H
PCS
2005
,
pp
.
65
-
71
,
2005
.
[23]
C.
Yong,
e
t
a
l.
,
"P
roposal
of
Co
nta
in
er
-
Based
HP
C
Struct
ure
s
a
nd
Perform
anc
e
Anal
y
sis
,
"
Journ
al
of
Informatio
n
Proce
ss
ing
Syst
e
ms
,
vol. 14, no.
6,
pp
.
1398
-
140
4,
2018
.
[24]
P
.
China Venka
n
na
Varm
a, Ve
nk
at
a
Kal
y
a
n
Chak
rav
art
h
y
K
.
,
V.
Vall
i
Kum
ari
,
S.
Visw
ana
dha
Ra
ju,
"A
na
l
y
sis
of a
Network
IO
Boo
tl
en
ec
k
in
Big
D
at
a
Envi
ronm
ents
Based
on
Do
ck
er
Cont
ai
ner
s"
,
J
ournal
of
Bi
g
Da
ta
R
ese
arch
,
vol.
3,
pp
.
24
-
28
,
Ap
r
2016
[25]
G
.
Calarc
o
,
M.
C
asoni,
"O
n
th
e
Ef
fec
t
ive
ness
of
Linux
Conta
in
ers
f
or
Network
Virt
ual
i
za
t
ion",
Journal
of
S
imulat
io
n
Mode
ll
ing
Prat
i
ce
and
Theory
,
v
ol.
31
,
pp
.
169
-
1
85,
Feb
2013
[26]
T.
Sterling,
et
a
l.
,
"A
Surv
e
y
:
Runti
m
e
Softwa
re
S
y
stems
for
High
Perform
an
ce
Com
puti
ng.
"
Superc
omput
in
g
Fronti
ers and
In
novat
ions
,
vol
.
4
,
no
.
1
,
2017
,
pp
.
48
-
68,
2017.
[27]
S.
Hunold
and
A.
Carpe
n
-
Am
ar
ie
,
"R
epr
oduci
b
l
e
MP
I
Benc
hm
ark
ing
is
Stil
l
Not
as
Ea
s
y
as
You
Thi
nk,
"
in
IE
E
E
Tr
ansacti
ons on Paralle
l
and
Dis
tribut
ed
Syste
ms
,
vol
.
27
,
no
.
12
,
pp.
3617
-
3630
,
1
Dec
.
2016
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
2
,
A
pr
i
l
2020
:
1505
-
1514
1514
BIOGR
AP
H
I
ES
OF
A
UTH
ORS
P
r
of.
San
g
Boe
m
Lim
re
ce
iv
ed
his
PhD
in
Co
m
pute
r
Sci
enc
e
from
Florida
St
at
e
Univer
si
t
y
i
n
2003.
Previousl
y,
he
leade
d
Kor
e
a
e
-
Sc
ie
nc
e
Proj
ec
t
as
t
ec
hni
ca
l
Te
am
L
ea
d
er
a
t
Korea
Insti
tut
e
o
f
Scie
nc
e
Techno
log
y
Inform
atio
n
(KIS
TI)
Supe
rco
m
puti
ng
C
en
te
r.
He
cur
ren
t
l
y
is
an
assis
ta
nt
profe
ss
or
for
Ko
nkuk
Univer
si
t
y
.
His
rese
arc
h
in
t
ere
sts
include
hi
gh
-
per
form
anc
e
computing
and
cl
oud
computing
.
Dr.
Guoh
ua
Li
re
ceive
d
her
P
hD
in
Int
erd
isciplina
r
y
IT
from
Konkuk
Univ
e
rsit
y
in
2018
.
She
comple
te
d
her
M
.
S.
degr
e
e
from
Konkuk
Un
ive
rsit
y
,
KO
REA,
in
2013.
She
par
ticipated
in
cl
oud
computin
g
applic
at
ion
fo
r
HP
C
service
par
t
-
t
ime
tr
ai
nin
g
at
Kore
a
Inst
it
ute
of
Scie
n
c
e
Te
chno
log
y
Info
rm
at
ion
(KIS
TI)
as
a
studen
t
r
ese
arc
h
er
from
20
13
to
2015.
She
al
so
p
art
i
ci
pa
ted
in the
d
evelopm
ent
of
con
tainer
-
base
d
HP
C
cl
ou
d
servic
e
platfor
m
as
a co
-
rese
ar
che
r from
2016
t
o
2017.
She
cur
re
ntly
is
a
post
-
do
ct
ora
l
r
ese
arc
h
er
for
KIS
TI.
Her
rese
ar
ch
int
e
res
ts
include
h
igh
-
per
form
anc
e
co
m
puti
ng
and cl
o
ud
computing.
Dr.
Joon
Woo
r
ec
e
ive
d
his
Ph.
D.
at
Chungn
am
Nati
ona
l
Univer
sit
y
in
2018.
H
e
was
invol
ved
in
buil
ding
infr
astr
uct
ure
for
the
HP
C
servic
e
at
KIS
TI
Superc
om
puti
ng
Infra
struct
ur
e
Center
.
He
cur
r
entl
y
is
a
sen
ior
rese
a
rc
her
for
KIS
TI.
His
rese
ar
ch
intere
st
includes
HP
C,
co
n
ta
in
erize
d
HP
C,
and
HP
C
Cloud.
Evaluation Warning : The document was created with Spire.PDF for Python.