Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
23
,
No.
3
,
Septem
ber
2021
, pp.
1419
~
1431
IS
S
N: 25
02
-
4752, DO
I: 1
0.1
1591/i
j
eecs
.v
23
.i
3
.
pp
1419
-
1431
1419
Journ
al h
om
e
page
:
http:
//
ij
eecs.i
aesc
or
e.c
om
Su
rv
ey o
n: A vari
ety of AQM
algorith
m s
ch
emas an
d in
tellige
nt
tec
hn
iqu
es
d
eveloped fo
r c
ongesti
on
co
nt
ro
l
Amar A
. M
ahawi
sh, Hass
an J. H
as
s
an
Com
pute
r
Engi
n
ee
ring
Depa
r
tment,
Univ
ersity
of
Technol
og
y
,
Ir
a
q
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
v
ed
Ma
y
24
,
2021
Re
vised
Ju
l
30
,
2021
Accepte
d
Aug
7
,
2021
The
cong
esti
on
on
the
interne
t
i
s
the
m
ai
n
issue
tha
t
aff
ects
the
p
erf
orm
anc
e
of
tra
nsiti
on
data
over
th
e
net
w
ork.
An
al
gori
th
m
for
conge
stio
n
cont
rol
is
req
uire
d
to
ke
ep
an
y
net
work
ef
f
ic
i
ent
and
r
el
i
able
for
tra
nsf
er
traffic
da
ta
of
the
users.
Man
y
Algorit
hm
s
had
bee
n
suggested
over
the
y
e
ars
to
improve
the
con
trol
of
co
ngesti
on
th
at
oc
c
urs
in
the
ne
twor
k
such
as
drop
tail
pa
cke
ts.
Rec
en
tly
th
ere
a
re
m
an
y
al
gor
ithm
s
have
bee
n
deve
lop
ed
to
ov
erc
om
e
th
e
dra
wbac
k
o
f
th
e
drop
ta
i
l
pro
ce
dure
.
One
of
the
important
al
gor
it
hm
s
deve
lop
ed
is
a
ct
iv
e
queu
e
m
ana
gement
(AQ
M)
tha
t
p
rovid
es
eff
i
ci
en
t
conge
stion
cont
r
ol
b
y
r
educ
ing
d
rop
pac
ke
ts,
thi
s
te
chni
qu
e
consi
der
ed
as
a
base
for
m
an
y
o
the
r
cong
esti
on
cont
rol
al
gori
th
m
s
sche
m
a.
It
works
at
the
net
work
cor
e
(ro
ute
r)
for
cont
ro
lling
the
drop
and
m
ark
ing
of
pa
c
ket
s
in
the
route
r
'
s
buffe
r
bef
ore
the
conge
stion
in
c
ept
ion
.
In
this
study
,
a
comprehe
nsive
surve
y
is
done
on
the
AQM
Algorit
hm
sc
hemas
tha
t
proposed
and
m
odifi
c
at
ion
the
se
al
gori
thms
to
achie
ve
the
b
est
pe
rform
anc
e
,
the
class
ifi
c
at
ion
of
AQ
M
al
gorithm
s
base
d
on
qu
eue
le
ng
th,
queu
e
del
a
y
,
or
both.
Th
e
adv
an
ta
ges
and
li
m
it
a
t
ions
of
ea
ch
al
g
orit
hm
have
be
e
n
discussed.
Also,
deba
t
e
th
e
intelligent
tec
hnique
s
pr
oce
du
re
with
AQ
M
a
lgori
thm
to
ac
hi
eve
op
ti
m
izati
on
in
per
for
m
anc
e
of
a
lgorithm
oper
at
ion
.
Final
l
y
,
th
e
compari
son
has
bee
n
discussed
among
al
gorit
h
m
s
to
find
the
wea
kness
and
powerful
of
each
one
b
ase
d
on
di
ffe
ren
t
m
et
ri
cs.
Ke
yw
or
ds:
AQM
Fu
zzy
lo
gic
GA
P
ID co
ntr
oller
RED
This
is an
open
acc
ess arti
cl
e
un
der
the
CC
B
Y
-
SA
l
ic
ense
.
Corres
pond
in
g
Aut
h
or
:
Am
ar A
. Ma
ha
wish
Com
pu
te
r
E
ng
i
neer
i
ng D
e
par
t
m
ent
Un
i
ver
sit
y o
f Te
ch
no
l
og
y
Ba
ghda
d
-
571
-
15
-
11, Ira
q
Em
a
il
:
ce.19
.
17@
gr
a
d.u
otech
no
l
og
y.e
du.iq
1.
INTROD
U
CTION
The
inter
net
tod
ay
widely
use
d
by
m
any
a
pp
li
cat
io
ns
an
d
pro
vid
es
a
va
riet
y
of
inf
orm
at
ion
us
in
g
sta
nd
a
rd
com
m
un
ic
at
ion
protoc
ols
suc
h
a
s
TCP/
IP
.
The
increasi
ng
use
r’
s
dem
and
on
the
i
nter
net
m
akes
a
chall
enge
in
pr
ov
i
ding
se
rv
ic
es
an
d
qual
it
y
of
ser
vice
“
QoS”.
T
he
c
onge
sti
on
pro
blem
i
s
the
m
ai
n
fact
or
t
hat
aff
ect
s
inter
net
per
f
orm
ance.
The
co
ngest
io
n
phen
om
ena
oc
cur
wh
e
n
m
any
us
ers
s
en
d
a
la
rg
e
am
ou
nt
of
data
on
t
he
sam
e
l
i
nk
t
hat
excee
de
d
it
s
capaci
ty
.
The
c
ongesti
on
e
ff
ect
QoS
by
increase
l
at
ency
and
drop
of
sen
ding
pack
et
s.
The
c
ongestio
n
c
on
t
ro
l
“C
C”
is
a
m
echan
ism
and
te
chn
i
que
us
ed
to
re
duce
c
ongestio
n
on
the
wire
d
and
wireless
ne
tworks
as
in
[1]
by
us
i
ng
r
ou
ti
ng
i
nfor
m
at
ion
strat
e
gy
to
e
nh
a
nce
t
he
internet
perf
orm
ance.
The
CC
al
l
ow
s
a
se
nd
e
r
t
o
a
djust
it
s
rate
bas
ed
on
m
easur
ing
the
co
ngest
ion
sta
tus
in
t
he
netw
ork
.
T
here
ar
e
two
a
ppr
oac
he
s
to
s
olv
e
this
pro
blem
:
pr
eve
nt
co
ngest
io
n
befor
e
it
occ
urs
or
detect
s
an
d
rem
ov
e
co
ng
est
ion
after it
occ
urs.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
23
, N
o.
3
,
Se
ptem
ber
2
0
21
:
14
19
-
14
31
1420
The
t
ran
sm
issio
n
c
on
tr
ol
prot
oco
l/
inter
net
prot
oco
l
"
TCP/
I
P"
is
the
inte
rnet
'
s
pr
im
ary
pr
oto
c
ol.
T
he
TCP
co
ngest
io
n
c
on
tr
ol
wor
k
in
the
sam
e
c
on
ce
pt
of
a
ddit
ive
-
inc
rease,
m
ul
ti
plica
ti
ve
-
decr
ease
"A
IM
D"
[
2]
.
The
AI
MD
ha
s
two
li
nes
,
th
e
first
one
is
c
al
le
d
the
ef
fici
ency
li
ne,
a
nd
the
seco
nd
li
ne
is
cal
le
d
the
f
ai
rn
ess
li
ne.
For
optim
al
con
tr
ol,
the
two
fl
ow
s
m
us
t
reach
the
opti
m
al
po
int,
the
intersect
io
n
po
int
of
the
ef
fici
ency
li
ne
an
d
t
he
fai
rn
ess
li
ne.
T
he
TCP
us
e
s
an
e
nd
to
e
nd
fl
ow
co
nt
ro
l
to
av
oi
d
c
ongestio
n
[
3]
.
T
he
TCP h
as
f
our
ph
a
ses
of
co
ngest
io
n
co
ntr
ol
al
go
rithm
s,
slow
sta
rt
"SS",
congesti
on
av
oid
a
nce
"C
A",
fast
retransm
i
t
,
an
d
fast rec
overy
.
Ther
e
a
re
tw
o
par
am
et
ers
bes
ides
S
S
a
nd
C
A,
the
c
onge
sti
on
wind
ow
"c
wnd”
a
nd
a
dve
rtise
window
“rwn
d”.
The
c
wnd
is
sen
der
-
side
that
sp
eci
fy
the
nu
m
ber
of
pac
kets
th
e
send
e
r
can
s
end
befor
e
rec
ei
vin
g
ackno
wled
ge
“
ACK”,
w
hile
t
he
r
wnd
is
rec
ei
ver
-
side
that
determ
ine
the
nu
m
ber
of
pac
kets
that
can
a
ccept.
The
TCP
co
ntr
ol
sel
ect
s
the
m
ini
m
u
m
of
cwnd
a
nd
r
w
nd
for
tra
ns
m
issio
n
data.
The
T
CP
sen
der
us
e
d
s
l
ow
sta
rt
and
c
onge
sti
on
av
oida
nc
e
beh
a
vior
as
s
how
n
in
Fig
ur
e
1.
I
n
this
cas
e
the
TCP
slo
wly
send
pac
ke
ts
into
netw
ork
t
o
est
im
at
e
the
capac
it
y o
f
li
nks th
r
ough ro
utin
g
,
t
his pr
ocedu
re
us
e
d
in
orde
r
t
o
a
vo
i
d
c
onges
ti
ng
the
netw
ork
.
T
he
s
low
sta
rt
th
res
ho
l
d
“ssth
r”
pa
ram
et
er
allow
s
the
co
ntro
ll
er to
sp
eci
fy w
hic
h
phase
is
SS
or
CA
.
the
SS
i
ncr
eas
ing
e
xpone
ntial
ly
by
duplica
te
the
cw
nd
value
wh
il
e
th
e
CA
was
inc
reasin
g
li
near
l
y
by
increasin
g
c
w
nd
by one.
To d
et
erm
ine the
ne
w
sst
hr
,
the c
wnd divi
de
in
half. This
ne
w sst
hr
deci
ded
wh
e
n
al
l
pack
et
s
dro
pp
e
d.
T
he
TCP
da
ta
transm
issi
on
in
unkn
own
ne
twork
ca
pacit
y,
so
the
SS
c
an
quic
kly
discove
r
the n
et
wor
k
li
nk ca
pacit
y avai
la
ble for safe t
r
ansm
issi
on
.
Figure
1. TCP
congesti
on c
on
trol
beh
a
vior
The
TCP
us
es
the
sli
din
g
wi
ndow
m
et
ho
d
to
thr
ottl
e
the
send
e
r
rate,
w
hi
ch
is
a
certai
n
nu
m
ber
of
pack
et
s
that
th
e
send
e
r
can
s
end
befo
re
AC
K
is
receive
d.
This
m
et
ho
d
is
us
ed
to
m
axim
iz
e
the
util
izati
on
of
band
width.
Th
e
init
ia
l
value
of
cw
nd
is
init
ia
l
window
"I
W
".
T
he
I
W
val
ue
sp
eci
fies
the
siz
e
of
data
transm
issi
on
,
wh
et
her
slo
w
or
a
ggres
sive
da
ta
transm
issi
on
.
T
he
reas
on
beh
i
nd
inc
reasing
I
W
as
f
ound
in
[
4]
to
10
du
e
to
i
nter
net
traff
ic
,
m
os
tl
y
is
web
traff
ic
w
hich
has
short
-
li
ved
connecti
ons.
The
increa
sin
g
I
W
is
help
fu
l t
o
inc
re
ase inte
r
net
perform
ance b
y p
r
ob
e
the a
vaila
bl
e b
an
dwidth
on th
e
n
et
wor
k.
The
Q
ueu
i
ng,
m
ark
in
g,
a
nd
dro
pp
i
ng
[
5]
a
re
f
ound
i
n
a
ny
net
w
ork
sys
tem
us
ing
fi
nite
m
e
m
or
y
"buff
e
r"
to
pro
te
ct
it
sel
f
fr
om
congesti
on
c
ollapse.
T
he
que
uing
m
eans
the
pack
et
s
buf
fere
d
in
a
qu
e
ue
whe
n
sp
ace
in
that
buf
fer
a
vaila
ble
.
The
syst
em
dr
ai
ns
it
s
buffer
base
d
on
sc
hedule
al
gorit
hm
s
(su
ch
as
f
irst
-
in
-
first
-
ou
t
"FI
FO"
).
If
the
sen
de
r
se
nds
data
at
a
rate
higher
th
an
outg
oing
li
nk
ca
pacit
y,
the
pack
et
s
that
re
side
in
the
buf
fer
w
il
l
increase,
ca
us
in
g
a
bu
ff
e
r
ov
e
rf
l
ow,
wh
ic
h
pro
du
ces
c
onge
sti
on
in
t
he
netw
ork
.
T
he
act
ive
qu
e
ue
m
anag
e
m
ent
“AQ
M”
al
gorithm
s
pr
opose
d
to
sta
rt
m
ark
in
g
or
dro
p
the
pa
cket
be
fore
the
qu
e
ue
fu
ll
to
handle
co
ngest
ion
.
The
se
nd
er
inter
pr
et
s
t
he
dr
oppe
d
pa
cket
as
a
sig
na
l
to
wh
ic
h
c
onge
sti
on
occ
urs,
an
d
reducin
g
it
s
rate
is
req
uir
ed
.
The
ty
pe
of
se
r
vi
ce
“ToS”
is
spe
ci
fied
in
the
s
econd
byte
of
the
IPv
4
hea
der
an
d
the
traff
ic
cl
ass
in
the
IP
v6
head
e
r
to
cl
assify
the
traff
ic
ty
pes
flow
.
F
or
ex
am
ple,
crit
ic
al
network
traff
ic
needs
low
la
te
ncy
li
ke
vo
ic
e
or
stream
ing
m
edia,
w
hile
non
-
c
riti
cal
netwo
r
k
traf
fic
nee
ds
best
ef
fort
serv
ic
e
s
li
ke
file
t
ra
ns
fe
r or
we
b
tra
ff
ic
.
This
rev
ie
w
st
ud
y
will
exa
m
ine
the
dev
e
lop
m
ent
of
A
QM
al
gorit
hms
an
d
t
heir
a
dvanta
ge
a
nd
dr
a
w
back
of
e
ach
al
gorithm
.
Sect
ion
2
has
a
br
ie
f
desc
ription
of
A
QM
and
cl
assi
ficat
ion
al
gorithm
s.
Th
e
Sect
ion
3,
s
ome
essenti
al
al
gorithm
s
ba
sed
on
queue
le
ngth
ha
ve
bee
n
di
scusse
d.
S
ect
io
n
4
li
sts
al
gori
thm
s
that
achieve
d
t
he
be
st
pe
rform
ance
w
hen
di
ff
ere
nt
ty
pes
of
se
r
vices
flo
w
ba
sed
on
qu
euin
g
delay
.
S
ect
ion
5
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Su
r
vey
on:
A
v
ar
ie
ty
o
f A
QM
algorit
hm sc
he
ma
s
an
d
i
ntell
i
gen
t t
ec
hn
i
que
s d
evel
op
e
d…
(
Ama
r
A.
Ma
hawi
sh
)
1421
stud
ie
d
the
al
gorithm
s
that
ut
il
iz
e
the
ad
vant
ages
of
both
c
la
sses
of
A
QM
schem
as
base
d
on
qu
e
uing
l
engt
h
and
qu
e
uing
delay
.
Sect
ion
6
descr
i
bed
the
intel
li
gen
t
op
ti
m
iz
at
ion
procedure
s
to
en
han
ce
t
he
AQ
M
al
gorithm
s.
Sect
ion
7
c
om
par
ison bet
ween d
isc
us
se
d
al
gori
thm
s an
d
the
c
on
cl
us
io
n
is t
he
last
secti
on.
2.
AC
TI
VE Q
U
EUE M
A
NAGEMENT
AQM
A
LG
ORIT
HMS
The
act
ive
qu
eue
m
anag
em
e
nt
"AQ
M"
[
6]
is
a
te
chn
ique
us
ed
to
m
anag
e
co
ngest
ion
in
netw
ork
dev
ic
es
(s
uc
h
as
r
ou
te
r’
s
que
ue)
be
fore
t
he
buff
e
r
fu
ll
.
Th
e
A
QM
wa
s
de
velo
ped
to
el
i
m
inate
the
dro
p
-
ta
il
“DT”
FIFO
queue
draw
bac
ks.
First,
the
DT
dr
aw
bac
ks
di
scard
the
new
incom
ing
pac
ke
ts
wh
e
n
the
buf
fer
ov
e
rf
l
ow,
so
the
co
ng
e
sti
on
no
ti
ficat
io
n
is
done
w
he
n
th
e
qu
e
ue
bec
om
es
fu
ll
.
Seco
nd,
the
DT
m
ay
no
t
achieve f
ar
nes
s
queue
“F
Q",
w
hic
h
m
eans
a
few
co
nnect
io
ns
siz
e
d
the
qu
eue b
uffe
r,
which p
re
ve
nts
the
othe
r
new
c
onnecti
ons.
T
hird,
the
la
rg
e
pac
kets
burst
m
ay
increase
the
la
te
ncy
of
a
s
m
al
l
bu
r
st
of
pack
et
s
.
Finall
y,
the
DT
m
akes
con
t
ro
l
lo
op
sy
nchr
on
iz
at
io
n.
The
AQ
M
has
been
im
pr
ove
d
[7]
to
e
nhance
the
pe
rfor
m
ance
of
internet
work.
The
A
QM
wa
s
desi
gn
e
d
to
ac
hieve
le
ss
pa
cket
los
s,
l
ow
qu
e
uing
del
ay
,
an
d
high
l
ink
capaci
ty
util
iz
at
ion
.
AQM
ens
ur
es
dr
oppi
ng
or
m
ark
i
ng
pac
kets
be
f
or
e
a
r
oute
r
’s
buff
e
r
bec
om
e
f
ull.
A
QM
w
ork
as
qu
e
ue
siz
e
regulat
or
in
c
on
t
r
ol
theo
ry
po
i
nt
of
vie
w.
The
input
to
AQ
M
al
go
rithm
is
l
ev
el
of
c
onges
ti
on
in
qu
e
ue
a
nd
the
ou
t
pu
t
t
he
pro
ba
bili
ty
of
m
ark
ing
or
dro
ppin
g
the
pac
kets.
AQ
M
se
nd
ear
ly
feed
bac
k
t
o
sen
der
for
increase
or
red
uc
e
the
send
i
ng
pack
et
s
to
avo
i
d
dro
pp
ing
pac
kets
as
well
as
hig
h
util
iz
e
the
avail
able
band
width.
Ther
e
a
re
m
any
AQ
M
schem
as
was
pr
opose
d
to
achie
ve
these
goal
s.
T
he
Ra
ndom
Earl
y
Detect
ion
was
the
earli
es
t
al
go
rithm
design
e
d
as
A
QM
congesti
on
co
ntr
ol.
The
AQM
al
go
rithm
s
can
be
cl
assifi
ed
int
o
three
m
ai
n
m
e
thods
to
m
eas
ur
e
net
w
ork
de
vices'
con
gest
ion
Fig
ure
2.
Th
e
first
m
et
h
od
is
base
d
on
qu
e
ue
le
ng
th
“
QL”
,
wh
ic
h
m
easur
es
how
m
any
pack
et
s
reside
i
n
the
router'
s
qu
eue.
T
he
sec
ond
m
et
ho
d
is
ba
sed
on
the
m
eantim
e
t
hat
a
pac
ket
s
pe
nds
in
a
qu
e
ue
.
The
t
hir
d
m
et
hod
is
ba
sed
on
both
t
he
le
ngth
a
nd
tim
e
delay
of
pack
et
s
i
n
t
he qu
e
ue.
Figure
2. Cl
assifi
cat
ion
of AQ
M al
gorithm
s
3.
AQM
ALGO
RITH
MS
B
A
SED ON
QUE
UING LE
N
G
TH
In
this
A
QM
schem
as,
the
c
on
t
ro
ll
er
reacti
on
is
base
d
on
m
easur
ing
th
e
net
work
de
vi
ce'
s
qu
eue
.
This
sc
hem
a
is
the
earli
est
al
gorithm
to
ad
opt
AQ
M beha
vi
or
t
o
m
ark
/dr
op p
ackets befo
r
e
the
que
ue
f
ul
l.
The
ov
e
rall
advant
age
of
these
sc
hem
as
is
e
lim
i
nating
t
he
DT
m
echan
ism
'
s
dr
aw
bac
k
an
d
im
pr
ov
in
g
the
QoS
of
the
inter
net.
Si
m
ul
ta
neo
usl
y,
the
ge
ner
al
disadv
a
ntage
is
t
he
pro
blem
a
ti
c
t
un
i
ng
of
the
pa
ram
et
er
to
ach
ie
ve
an op
ti
m
al
r
esult.
3.1.
Rand
om
early
de
tecti
on
-
red
RED
is
desig
ned
as
AQ
M
te
chn
i
qu
e
[8
]
,
[
9]
to
perform
pack
et
qu
eue
m
anag
em
ent
to
a
vo
i
d
dr
a
w
back
of
DT
a
nd
im
pr
ove
the
pac
ket
transm
issi
on
durin
g
netw
ork
co
ngest
io
n.
The
RE
D
al
gorithm
com
pu
te
s
the
a
v
era
ge
qu
e
ue
s
iz
e
“
avg
”
an
d
com
par
ed
it
wi
th
the
s
pecific
thres
ho
l
d.
T
he
RED
determ
ined
th
e
pro
bab
il
ist
ic
m
ark
/
drop
f
or
ea
ch
new
i
nc
om
i
ng
pac
kets
if
t
he
avg
le
ss
tha
n
the
m
ini
m
um
threshold
“
min
th
”
the
pro
ba
bili
ty
drop
“
P
d
”
is
zero
a
nd
t
here
are
no
pac
ke
ts
d
r
op
w
hen
the
av
g
exce
eded
t
he
m
axi
m
u
m
thres
ho
l
d
“
max
th
”
the
pro
ba
bili
ty
reaches
to
one,
an
d
al
l
new
i
nc
om
in
g
pack
et
s
dro
pp
e
d.
The
av
g
queu
e
le
ng
th
can
be c
al
culat
ed
by u
s
ing
e
xpone
ntial
w
ei
ghte
d
m
ovin
g
a
ver
a
ge
“
E
W
MA”
alg
ori
th
m
as shown i
n
(
1
)
.
(1)
Wh
e
re
W
q
is
prede
fine
d
wei
ght
queue
c
oe
ffi
ci
ent
and
is
t
he
que
ue
le
ngth.
Wh
e
n
the
va
lue
of
av
g
betwee
n
min
th
and
max
th
the
pro
bab
il
it
y
assigne
d
to
inc
omi
ng
pac
kets
to
dro
p
these
pac
kets
can
be
cal
culat
es
by
(
2
)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
23
, N
o.
3
,
Se
ptem
ber
2
0
21
:
14
19
-
14
31
1422
(2)
Wh
e
re
the
m
ax
p
is
a
m
axi
m
um
dr
oppi
ng
prob
a
bili
ty
.
The
RED
al
gorith
m
is
sh
ow
n
i
n
Figure
3.
Th
e
dro
p
pro
ba
bili
ty
in
(
2
)
dr
ive
d
from
se
m
ico
ntin
uous
piec
ewise
li
near
f
un
ct
io
n
[
10]
of
the
ave
rag
e
qu
e
ue
le
ng
th
as
s
how
n
in
Fig
ure
3
,
s
o
w
hen
t
he
av
g
increa
se
the
P
d
increases
unt
il
reach
to
m
axim
u
m
value
(which
is
P
d
)
a
nd the
n al
l i
nco
m
ing
pac
kets
will
d
r
op.
Figure
3. RED
al
gorithm
The
ad
va
ntage
s
of
the
RED
al
gorithm
hav
e
ov
e
rco
m
e
the
prob
le
m
of
t
he
f
ull
qu
e
ue
and
global
synch
ronizat
io
n.
T
he
disa
dv
a
ntages
are
t
he
com
plex
tu
ni
ng
par
am
et
er,
unfair
que
ue
by
hav
i
ng
the
sam
e
dro
p
rate
pe
rfor
m
to
al
l
traff
ic
s
ty
pe
beca
us
e
it
ca
nnot
diff
e
re
ntiat
e
betwee
n
T
oS
a
nd
RED
unable
t
o
sta
bili
ze
the
com
pu
ta
ti
on
of
avg
val
ue
between
tw
o
the
t
hr
es
holds
w
he
n
tra
ff
ic
l
oad
r
apidly
ch
an
ge.
The
RED
al
gorithm
hav
e
b
ee
n de
ve
lop
e
d
to
im
pr
ov
e
their
p
e
rfo
rm
ance as sho
wn in ne
xt p
a
r
agr
a
phs.
The
fair
ra
ndom
early
detection
“
FRED
”
[
11
]
was
desi
gned
to
el
i
m
inate
RED'
s
unfai
r
drop
w
he
n
diff
e
re
nt
traff
i
cs
ty
pes
are
use
d.
I
n
RED
,
al
l
con
necti
on
s
will
hav
e
th
e
sa
m
e
loss
rate
wh
en
qu
e
ue
le
ng
th
exceede
d
the
thres
hold.
T
he
slow
rate
co
nn
ect
ion
us
i
ng
le
ss
than
it
s
fair
band
width
w
il
l
hav
e
the
sam
e
pack
et
loss
rate
as
a
hig
he
r
rate
conn
ect
ion
,
w
hich
i
s
unfair
util
iz
ing
the
band
width
f
or
this
ki
nd
of
c
onnecti
on.
The
F
RED
al
lo
ws
f
ai
rn
ess
f
or
a
di
ff
e
ren
t
ty
pe
of
traff
ic
c
on
nec
ti
on
by
assig
ni
ng
m
ini
m
u
m
p
ackets
of
eac
h
flo
w
that
would
e
na
ble
buf
fer
i
ng
be
fore
any
dro
p.
The
FR
ED
m
ai
ntains
av
g
f
or
each
flo
w
ty
pe,
if
av
g
exce
eded
the
th
res
ho
l
ds
the
dr
op
pr
ob
a
bili
ty
will
cal
c
ul
at
e
to
m
ark
/
dro
p
pack
et
s
of
this
flo
w
so
that
the
FRED
will
pen
al
iz
e t
he
a
ggressi
ve flo
w.
The
wei
gh
te
d
rand
om
early
detect
ion
“
WR
ED”
[
12]
was
desig
ne
d
to
dr
op
pac
kets
bas
ed
on
traf
fic
flo
w
weig
ht
or
pri
or
it
y
in
a
ddit
ion
to
m
ulti
pl
e
virt
ual
qu
e
ue
s
an
d
queue
th
reshold
f
or
eac
h
virtu
al
que
ue
.
T
he
IP
pr
ece
de
nce
in
the
IPv4
he
ader
is
us
e
d
t
o
gi
ve
pr
io
rity
for
eac
h
pac
ke
t
and
disti
ng
uish
am
ong
se
rv
ic
es
flo
ws.
T
his
pr
ecedenc
e
valu
e
increases
or
decr
ease
s
base
d
on
the
im
po
rtance
of
the
pack
et
.
The
w
ork
of
WRED
is
t
o
dr
op
l
ow
e
r
-
pri
or
i
ty
pack
et
s
by
s
et
ti
ng
va
rio
us
dro
p
-
pro
bab
il
it
y
fu
nc
ti
on
s
f
or
each
pri
ori
ty
le
vel.
Each
t
raffic
cl
ass
ha
s
a
virt
ual
qu
e
ue
a
nd
a
qu
e
ue
thre
s
ho
l
d.
The
CI
S
CO
r
outi
ng
platfor
m
s
sup
port
this
al
gorithm
in
re
cent series
produ
ct
s
[
13
]
.
An
a
doptiv
e
ra
ndom
early
det
ect
ion
“ARE
D
”
was
sug
geste
d
by
[14
]
,
[
15]
to
so
lve
RE
D
par
am
et
er
tun
in
g.
T
he
A
RED
pe
rfor
m
s
sel
f
-
co
nf
i
guri
ng
of
RE
D
para
m
et
ers
based
on
m
ixed
traffi
c
ty
pes.
The
ARE
D
adoptin
g
aut
otu
ne
of
m
ax
p
to
keep
av
g
ta
r
ge
t
range
bet
we
en
mi
n
th
an
d
m
ax
t
h
.
T
her
e
a
re
m
any
m
od
ific
at
ion
s
bu
il
d
ove
r
AR
ED
[16]
by
m
a
king
it
m
or
e
r
obus
t
f
or
de
ploym
ent
in
a
network
dev
ic
e.
T
his
m
od
ific
at
io
n,
f
or
exam
ple,
no
t
on
ly
keep
i
ng
t
he
avg
betwe
e
n
tw
o
th
res
ho
l
ds
bu
t
t
o
kee
p
av
g
in
half
wa
y
betwee
n
mi
n
th
and
max
th
t
o
in
crea
se
the
perf
or
m
ance
of the al
gorithm
.
Anothe
r
m
od
ific
at
ion
of
RE
D
is
gen
tl
e
ran
dom
early
detect
ion
“GRE
D”
[17
]
,
[
15]
,
wh
ic
h
use
d
gen
tl
e
pa
ram
eter
s
to
im
pr
ove
RED'
s
beh
a
vio
r
.
The
GRE
D
avo
i
d
a
sud
de
n
cha
nge
in
t
he
dro
p
pro
ba
bili
t
y
from
to
1
when
t
he
av
g
exc
eeded
the
.
Th
e
m
od
ific
at
ion
on
RED
im
pl
e
m
ent
by
add
ing
th
ree
thres
ho
l
ds
min
th
,
m
ax
th
,
a
nd
double
m
ax
th
.
I
n
GRED
gra
du
al
va
ryi
ng
of
dro
p
pac
kets
pr
obabili
ty
fr
om
m
ax
p
t
o
1
wh
e
n
t
he
av
g
e
xceed
e
d
max
th
to
do
ub
le
m
ax
th
.
T
he
t
hr
ee
thres
hol
d
le
ve
l
min
th
,
m
ax
th
,
and
do
ub
le
max
th
are
us
e
d
to
gi
ve
m
or
e
robust
to
tu
ning
m
ax
p
an
d
av
g
par
a
m
et
ers.
The
G
RED
was
desi
gn
e
d
t
o
el
im
in
at
e
th
e
pro
blem
of
sta
bili
zi
ng
the
av
g
val
ue
at
a
cer
ta
in
le
vel.
The
a
dap
ti
ve
GRE
D
“A
GRE
D”
in
[
18
]
was
de
ve
lop
e
d
from
GRED
to
increase
p
er
form
ance
by
dete
rm
ining
the
i
niti
al
dr
op u
si
ng a
sp
eci
fic
f
or
m
ula.
T
he
i
niti
al
dro
p
sta
rts
from
max
p
to
0.5
as
t
he
avg
le
ngth
m
ov
es
f
r
om
max
th
to
double
ma
x
th
.
En
han
c
ed
A
GRED
“
EA
G
RED”
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Su
r
vey
on:
A
v
ar
ie
ty
o
f A
QM
algorit
hm sc
he
ma
s
an
d
i
ntell
i
gen
t t
ec
hn
i
que
s d
evel
op
e
d…
(
Ama
r
A.
Ma
hawi
sh
)
1423
[19]
was
pr
opos
e
d
to
im
pr
ov
e
the
res
pons
e
of
the
router
’s
buf
fer
t
o
drop
pac
kets
as
well
as
this
al
gorithm
giv
es
be
tt
er
pe
rfor
m
ance
by
red
uci
ng
delay
and
pack
et
los
s.
The
EA
GRE
D,
unli
ke
the
AG
RE
D
an
d
GRED
,
wh
e
n
us
i
ng
th
e
fixe
d
value
of
que
ue
weig
ht
(e
.g.,
0.0
02)
to
cal
culat
e
the
av
g
w
hich
m
a
y
le
ad
del
ay
in
conges
ti
on
res
pons
e
.
The
sta
ti
c
pr
obabili
ty
m
ay
c
ause
high
dro
p
pac
kets
w
he
n
the
drop
pr
obabili
ty
value
is
increase
d,
wh
il
e
the
net
work
'
s
pe
rform
ance
decr
ea
s
es
due
to
c
ongestio
n
w
he
n
the
pro
ba
bili
ty
value
is
sm
al
l.
The
dynam
ic
al
go
rithm
was
pr
op
ose
d
to
cal
c
ulate
dro
p
pac
kets
b
ase
d
on
c
ongestio
n
sta
tus.
The
d
ynam
ic
GRE
D
“DG
RE
D”
[
20]
sa
m
e
as
GR
ED
by
ha
ving
double
m
ax
th
but
prov
i
des
a
f
ast
er
respo
ns
e
to
congesti
on
even
t
s
and
e
nh
a
nces
the
netw
ork'
s
pe
rfor
m
ance
when
us
in
g
a
dif
f
eren
t
ty
pe
of
s
erv
ic
es
fl
ow.
The
D
GRED
de
pends
on
the
th
ree
-
sta
te
m
ark
ov
m
od
ulate
d
bern
oull
i
arr
ival
pr
oc
ess
(MM
BP
-
3)
to
su
pport
thr
ee
ty
pes
of
the
traff
ic
flo
w
by
pe
rfo
r
m
ing
co
rr
el
at
ion
am
ong
thes
e
netw
ork
tra
ffi
cs.
An
e
xte
ns
ion
of
D
GRED
is
sta
bili
zed
dynam
ic
GRED
“S
D
G
RED”
[
21
]
,
[
22]
to
prov
i
de
dynam
ic
al
l
y
st
abili
zi
ng
the
av
g
betwee
n
min
th
and
m
ax
th
.
This
dynam
ic
al
go
ri
thm
us
es
dy
na
m
ic
increase
or
decr
ease
of
max
th
an
d
double
m
ax
th
value
by
sta
bili
zi
ng
the
av
g
value
a
r
ound
min
th
to
adj
us
t
the cal
culat
io
n of d
rop
ping
prob
a
bili
ty
.
3.2.
CHOKe
Ther
e
a
re
diff
e
ren
t
ty
pes
of
c
onnecti
on
fl
ow
[7]
,
the
flo
w
that
res
ponds
to
congesti
on
c
ontr
ol
sign
a
l
cal
le
d
TCP
-
f
riend
ly
flo
ws,
wh
il
e
the
oth
e
r
flo
ws
do
not
respond
to
re
du
ce
the
fl
ow
rate
wh
e
n
co
ngest
io
n
occurs
t
his
na
m
ed
unres
ponsi
ve
fl
ow
s
or
aggressi
ve
fl
o
ws
s
uc
h
as
U
DP
vid
e
o
or
voic
e
co
nnect
io
n.
T
he
fairn
e
ss
que
ue
“FQ”
will
no
t
achieve
w
hen
t
hese
tw
o
ty
pes
of
flo
w
s
har
i
ng
the
sam
e
bandw
i
dth
.
As
a
r
esult,
the agg
ressive
flo
w
will
size t
he
m
os
t avail
able b
a
ndwi
dth
wh
e
n
c
onge
sti
on o
cc
urs.
To
sa
ve
TCP
-
f
rien
dl
y
flo
ws
a
nd
to
ac
hieve
FQ
am
ong
di
fferent
flo
w
c
onnecti
on
ty
pes,
the
CH
O
K
e
al
gorithm
[23]
was
desi
gn
e
d.
The
CH
O
Ke
al
go
rithm
fo
l
lows
the
dro
p
cand
i
date
pac
ket
proce
dure.
This
proce
dure
w
ork
w
hen
t
he
a
ver
a
ge
que
ue
siz
e
avg
e
xce
eded
t
he
,
so,
the
arr
ivi
ng
new
pac
ket
is
com
par
ed
with
a
ra
ndom
ly
s
el
ect
pack
et
f
r
om
the
buff
e
r,
the
sel
ect
ed
pa
cket
cal
le
d
c
and
i
date,
if
t
he
y
are
from
the
sa
m
e
flo
w
ty
pe,
bo
t
h
of
t
hem
are
droppe
d.
Othe
rwi
se,
the
can
dida
te
pack
et
is
ke
pt
in
the
buf
fe
r,
a
nd
the
la
te
st
ar
riv
i
ng
pa
cket
is
dr
oppe
d
base
d
on
pro
bab
il
it
y
va
lue.
T
he
dro
p
pro
bab
il
it
y
val
ue
is
c
om
pu
te
d
as
in
the
RED
al
gor
it
h
m
,
wh
ic
h
is
aff
ect
ed
by
the
buff
e
r'
s
conge
sti
on
le
vel.
As
in
RED,
t
her
e
are
no
pac
kets
dro
p
if
the
av
g
le
ss
t
han
min
th
or
al
l
the
ne
w
ar
rivi
ng
p
a
c
kets
do
pe
if
the
av
g
pas
sed
the
m
ax
th
. Th
e
a
ggressiv
e
flo
w
has
m
or
e
packet
s
in
the
buffer
than
the
TC
P
-
f
rien
dly
flo
w
w
hen
c
onge
sti
on
occ
urs.
S
o
the
cha
nce
to
sel
ect
cand
i
date
pac
ke
t
fr
om
agg
res
sive
flo
w
high
er
than
the
othe
r
flo
w
ty
pes
and
this
high
drop
rate
from
this
ty
pe
of
flo
w.
T
he
a
dv
a
ntage
of
th
is
al
gorithm
has
a
low
proce
ssing
c
os
t.
Th
e
disad
va
ntage
is
the
pre
-
flo
w
ty
pe
inf
or
m
at
ion
ne
eded
f
or
it
s
pr
oc
ess.
T
he
C
HOKe
is
a
sta
te
le
ss
te
ch
nique
a
bout
t
he
num
ber
of
a
ggressi
ve
flo
ws
and pr
oduces l
ess p
e
rf
orm
ance when m
ulti
pl
e ag
gr
es
sive
flow
s
prese
nt.
The
C
HOKe
-
RH
“C
H
O
Ke
with
recent
dr
op
hist
or
y
”
al
gorithm
[24]
was
desig
ne
d
to
pr
otect
the
TCP
-
f
rien
dly
flow
s
a
nd
el
im
i
nate
the
lim
i
tation
s
CH
OK
e
al
gorithm
,
wh
ic
h
sp
a
ns
f
ro
m
sta
te
le
ss
to
stateful
.
The
CH
O
Ke
-
RH
us
es
t
wo
ph
a
ses
of
com
par
is
ons,
the
first
phase
is
th
e
init
ia
l
co
m
par
iso
n,
a
nd
t
he
second
ph
a
se
is
the
pen
al
ty
fo
r
ag
gre
ssive
flo
ws.
T
his
te
chn
i
qu
e
s
tore
histo
ry
inf
or
m
at
ion
that
has
rece
ntly
droppe
d
pack
et
s
f
lo
w
-
ids a
nd the
n use
it
to
pe
naliz
e t
he
a
ggress
i
ve
f
lows
.
3.3.
H
ash t
abl
e a
nd
circ
ula
r buffer
-
HTC
B
The
H
TCB
[25]
is
a
sta
te
fu
l
congesti
on
c
ontr
ol
al
gorithm
desig
ne
d
to
i
m
pr
ov
e
the
A
QM
sta
te
le
ss
schem
as
'
per
f
orm
ance
.
The
s
ta
te
le
ss
al
go
rit
hm
s
featur
es
a
re
sim
ple,
le
ss
processi
ng
re
qu
i
rem
ent
and,
fe
w
s
tora
ge
resou
rc
es
nee
d.
Th
e
st
at
el
ess
al
gorith
m
's
per
f
orm
ance
re
du
ce
d
w
he
n
m
anag
in
g
a
n
a
ggressi
ve
fl
ow,
s
o
the stat
efu
l
w
a
s sug
gested
to e
lim
inate
the l
ack
of m
anag
ing ag
gressi
ve
fl
ow
duri
ng con
gestio
n
c
on
tr
ol
.
The
HTCB
c
onsist
s
of
tw
o
pa
rts,
a
fi
xed
si
ze
o
f
has
h
ta
ble
“HT”
a
nd
ci
r
cular
buf
fer
“C
B”
.
T
he
HT
is
a
data
st
ru
c
ture
t
hat
rec
ords
a
bs
tract
i
nfor
m
at
ion
ab
ou
t
aggressi
ve
fl
ow
co
nnect
ions.
T
his
in
form
at
io
n
cal
culat
es
by
has
hing
an
d
w
il
l
con
ta
in
the
IP
a
ddress
a
nd
po
rt
num
ber
of
both
peer
s
in
ad
diti
on
t
o
flo
w
pack
et
s
rate
a
nd
dr
op
rate
of
aggressi
ve
fl
ow.
Wh
il
e
the
CB
records
inf
or
m
at
ion
of
th
e
ne
w
ar
riving
pac
kets
su
c
h
as
I
P
ad
dr
ess
an
d
port
nu
m
ber
for
both
peers.
T
his
al
gorithm
'
s
work
by
in
sp
e
ct
ing
the
ne
w
arr
ivi
ng
pack
et
s
if
it
be
longs
to
fl
ow
found
in
H
T,
then
the
pac
ke
t
will
be
dr
oppe
d
pa
ssed
on
pro
ba
bili
ty
.
In
the
seco
nd
phase
of
t
his
al
gorith
m
that
if
the
f
low
ty
pe
of
ar
riving
pack
et
i
s
not
f
ound
in
HT,
s
o
t
he
ar
riving
pack
et
will
b
e
com
par
ed wit
h a ra
ndom
ly
sel
ect
ed
pa
cket
from
CB
if its sam
e flo
w
t
ype
will
r
egiste
r
in HT.
4.
AQM
ALGO
RITH
MS
B
A
SED ON
QUE
UING
DELA
Y
The
A
QM
al
gorithm
s
base
d
on
que
ue
l
eng
t
h,
su
c
h
a
s
RED
,
s
uffe
red
f
ro
m
diffi
cult
tun
i
ng
par
am
et
ers
fo
r
diff
e
ren
t
ty
pe
s
of
fl
ow
se
r
vices
connecti
on
s
(su
c
h
as
TC
P
-
f
rien
dly
and
unres
pons
i
ve)
.
Thes
e
s
erv
ic
es
ha
ve
div
e
rse
se
nd
i
ng
data
rate
sp
e
eds,
var
ia
nce
r
ound
trip
ti
m
e
“R
TT”
an
d
li
nk
ty
pes
(
su
c
h
a
s
fibe
r
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
23
, N
o.
3
,
Se
ptem
ber
2
0
21
:
14
19
-
14
31
1424
op
ti
c
an
d
sat
el
li
te
).
Also
,
the
RED
m
anag
e
s
the
qu
e
ue
le
ng
t
h,
w
hich
i
m
pl
ic
it
l
y
aff
ect
s
la
te
ncy.
The
based
qu
e
ue
le
ng
t
h
al
gorithm
s
hav
e
a
la
c
k
to
i
m
pr
ove
the
pe
rfor
m
ance
of
la
te
ncy
-
sen
sit
ive
ap
plica
ti
ons
(or
serv
ic
es
)
t
hat
hav
e
RTT.
T
he
refor
e
,
m
any
al
gorithm
s
have
recently
been
sugg
est
e
d
to
i
m
ple
m
ent
con
tr
ol
congesti
on
bas
ed
on
la
te
ncy
or
RTT
t
o
determ
ine
the
le
ve
l
of
co
ngest
ion
instea
d
of
dep
e
ndin
g
o
n
qu
e
ue
le
ng
th
and eli
m
inati
ng
the
pr
ob
le
m
s in
que
ue
base
d
al
gorithm
.
4.1.
Clos
e
d
-
l
oop c
on
t
rolle
r
These
c
ontr
oller
was
us
e
d
in
m
any
researc
h
areas
s
uc
h
as
DC
m
oto
rs
[26
]
,
[
27]
,
R
oboti
cs
[
28
]
,
F
uel
Ce
ll
[29]
,
an
d
oth
e
rs
fiel
ds.
Also
in
A
QM
co
ng
e
sti
on
ne
t
work
[
30
]
us
e
d
the
c
on
tr
olle
rs
by
us
in
g
c
ontr
ol
syst
e
m
feed
ba
ck
su
c
h
as
pro
portio
nal,
integral,
a
nd
/
or
de
rivati
ve
“PI
D”.
The
co
ntro
l
s
yst
e
m
m
od
el
i
n
A
Q
M
fo
c
us
e
d
on
sta
bili
zi
ng
the
va
riat
ion
of
dif
f
eren
t
a
pp
li
cat
ion
flo
w
a
nd
r
edu
ci
ng
the
st
eady
-
sta
te
er
ror.
T
he
con
t
ro
l
syst
em
us
es
co
ns
ta
nt
gain
t
o
a
dju
st
the
A
QM
par
a
m
et
er
to
achie
ve
high
perf
orm
ance
in
c
ontr
olli
ng
congesti
on.
The
pro
portio
nal
integral
c
on
t
ro
ll
er
e
nh
a
nced
“P
IE”
[
31]
was
desi
gned
to
pe
rfor
m
congesti
on
con
t
ro
l wit
h
fe
edb
ac
k
co
ntr
ol
loop
c
on
nect
ion
ba
sed
on
la
t
ency.
T
he
PI
E
w
ork
u
ses
t
hr
e
e
basic
ste
ps
: r
andom
dro
pp
i
ng,
dro
p
pro
bab
il
it
y
updatin
g,
an
d
la
te
ncy
cal
cu
la
ti
on
,
as
sho
wn
in
Fi
gure
4.
Th
e
dro
p
pa
ckets,
accor
ding
to
pr
ob
a
bili
ty
dr
op
“
P
d
”
cal
culat
e
by
(
3
)
,
will
affe
ct
based
on
la
te
ncy
value,
w
hethe
r
it
increases
or
decr
ease
dro
pping
.
T
he
P
I
fac
tors
are
t
wo
c
onsta
nt
gai
n
and
fo
r
P
rop
or
ti
on
al
a
nd
In
te
gral
resp
ect
ively
,
us
e
d
to
ad
just
the
optim
al
c
al
culat
ion
of
pro
bab
il
it
y
drop
a
nd
re
duce
error
st
ud
y
st
at
e
“e”,
wh
e
re
error
cal
culat
ed
by
the
diff
e
re
nce
betwee
n
ta
r
get
queue
q
t
an
d
current
queu
e
q.
Wh
il
e
the
l
at
ency
cal
culat
ion
i
s
done
by
us
in
g
on
e
of
t
wo
wa
ys:
ei
ther
usi
ng
li
tt
le
’s
la
w
(
current
queue
delay
=
que
ue
byte
le
ng
t
h
/
de
qu
e
ue
rate
) or
by
us
in
g othe
r
ways
s
uch a
s
record
Ti
m
e
-
Sta
m
p
the p
ac
kets whe
n i
ns
erti
ng i
n
th
e queue
, th
e
n u
se this
Ti
m
eStam
p
to
ob
ta
in
la
te
ncy
wh
e
n
this
pac
ke
t
le
aves
the
qu
eue
.
Sam
e
as
the
RED,
the
PI
E
dro
p
al
l
pack
et
s
wh
e
n
it
exceeded
the
ta
r
get
la
te
ncy.
The
PI
E
al
gorithm
's
adv
a
ntages
are
inh
e
rite
d
f
ro
m
the
cl
assical
PI
con
t
ro
ll
er
m
eth
od,
w
hich
el
im
inate
s
the
ste
ady
-
sta
te
erro
r,
autotunin
g
pa
ram
et
er
based
on
c
ongesti
on
le
vel,
and achie
ves
st
abili
ty
.
(3)
(a)
(b)
Figure
4. The
PI
E
st
ru
ct
ur
e;
(
a)
detai
ls view
,
(
b)
ge
ner
al
vie
w of
PI co
ntr
ol
le
r
The
d
el
ay
-
b
a
se
d
P
I
c
on
t
ro
ll
er
en
han
ce
d
by
a
dap
ti
ve
CH
OKe
“D
-
PA
C”
[
32]
,
wa
s
de
sig
ne
d
as
delay
base
al
gorithm
to
ac
hieve
fair
ness
queue
am
ong
va
rio
us
se
r
vices
fl
ows
ty
pe
s,
by
us
i
ng
A
dap
ti
ve
CH
O
K
e
an
d
PI
c
ontrolle
r
.
The
a
da
ptive
CHO
Ke
us
e
d
t
o
pen
al
iz
e
a
gg
ressive
flo
ws
to
discipli
ne
fa
irness
with
slo
w
rate
flo
ws,
w
hile
th
e
PI
us
e
d
t
o
ke
ep
the
la
te
ncy
acce
ptable
with
diff
e
re
nt
flo
w
se
rv
ic
es
an
d
el
i
m
inate
the
ste
ady
-
sta
te
err
or.
As
m
entioned
a
bo
ve
ab
out
CHO
Ke
m
echan
ism
work,
sel
ect
in
g
one
or
m
or
e
pack
et
as
a
ca
ndidate
to
com
par
e
with
the
ne
w
inco
m
ing
pack
et
if
m
at
ch
then
discards
the
can
did
at
e
an
d
inc
om
ing
pack
et
s.
Wh
il
e
the P
I
cal
culat
es the lat
ency a
nd pr
ob
a
bili
ty
o
f
drop a
s m
ention
a
bove
i
n
t
he
P
IE
al
gorith
m
.
The
PD
-
A
QM
[
33
]
,
[
34]
c
ontr
o
ll
er
al
gorit
hm
was
us
e
d
base
d
on
TCP
wind
ow
siz
e
and
cu
rr
e
nt
qu
e
ue
le
ngth
within
ti
m
e.
T
he
val
ue
of
cu
r
ren
t
que
ue
le
ngth
q(t)
chan
ge
based
on
the
num
ber
of
a
ppli
cat
ions
flo
w,
wi
ndow
siz
e
w(
t
)
and
,
RTT
(
4
)
,
wh
e
r
e
the
RTT
is
m
easur
ed
by
both
prop
a
gatio
n
dela
y
an
d
queui
ng
delay
.
The
dro
p
pro
bab
il
it
y
p
is
m
easur
e
d
ba
sed
on
cu
rr
e
nt
qu
eue
q
an
d
ta
rg
et
q
t
val
ue
(
qu
e
ue
le
ngth
e
rror).
The
pro
portiona
l
and
de
rivati
ve
gain
is
us
e
d
to
el
i
m
inate
t
he
er
r
or
a
nd
al
so
to
cal
culat
e
the
dr
op
pro
ba
bili
ty
(
5
)
,
wh
e
re
thes
e
gain
e
ff
ect
by
window
siz
e,
qu
e
uing
delay
and,
li
nk
ca
pa
ci
ty
.
This
al
gorithm
pr
ov
i
des
high
li
nk
util
iz
at
ion
,
f
ast
er
sett
li
ng
tim
e, an
d
a
sli
ght o
ve
rsho
ot.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Su
r
vey
on:
A
v
ar
ie
ty
o
f A
QM
algorit
hm sc
he
ma
s
an
d
i
ntell
i
gen
t t
ec
hn
i
que
s d
evel
op
e
d…
(
Ama
r
A.
Ma
hawi
sh
)
1425
(4)
(5)
4.2.
CoDel
The
c
on
t
ro
ll
ed
delay
act
ive
queue
m
anag
em
ent
“C
oDel
A
QM
”
[
35]
,
sa
m
e
as
PI
E
was
us
e
d
qu
e
ue
delay
(lat
ency)
to
i
m
ple
m
ent
congesti
on
c
ontrol
an
d
ke
ep
delay
in
the
queue
as
m
ini
m
um
as
po
ssible
.
Th
e
CoD
el
is
par
a
m
et
er
-
le
ss
based
on
the
con
c
ept
of
tw
o
keys:
avo
idin
g
ba
d
queue
(a
dds
delay
)
and
e
nhanc
i
ng
good
qu
e
ue
(h
igh
ba
ndwidt
h
util
iz
at
ion
).
T
he
go
al
of
Co
Del
is
hi
gh
ba
ndwidt
h
util
izati
on
with
m
i
nim
a
l
delay
.
This
al
gorithm
i
m
p
leme
nts
by
cal
c
ulati
on
the
s
ojo
ur
n
tim
e
(the
tim
e
that
pack
et
s
ha
ve
sp
e
nt
in
the
qu
e
ue
)
of
a
st
and
i
ng
qu
e
ue
(the
num
ber
of
pack
et
s
en
qu
eues).
If
the
s
ojour
n
pac
ket’
s
tim
e
exceed
ed
t
he
ta
rg
et
que
uing
delay
tim
e,
then
the
al
gorith
m
sta
rts
droppi
ng
pac
kets.
T
he
ta
rg
et
ti
m
e
a
ct
s
as
a
thres
hold
th
at
sp
eci
fies
that
t
he
m
axi
m
u
m
qu
e
uing
delay
can
be
acce
pt
able
ab
ove
the
al
gorithm
m
a
rk
/
drop
pac
ket
s
an
d
t
arg
et
tim
e
m
e
asur
e
from
int
erv
al
ti
m
e.
The
interval
ti
m
e
sh
ould
be
at
le
ast
an
RTT
to
av
oid
pac
ket
dro
p
m
isbehav
i
or
a
nd
not
a
lot
hi
gh
e
r
t
han
RTT
to
pr
e
ve
nt
the
unne
cessary
de
la
y
in
detect
i
on
c
ongestio
n.
So
the
su
it
able s
ugges
te
d
inter
val ti
m
e is the m
axi
m
um
RTT a
m
ong
al
l s
har
in
g ap
plica
ti
on
s c
onne
ct
ion
.
Fair
q
ue
ue
C
oDel
“F
Q
-
C
oDel
”
[36
]
,
[
37]
,
is
an
exte
ns
i
on
of
t
he
C
oDel
al
gorithm
that
prov
i
des
fairn
e
ss
am
ong
diff
e
re
nt
flo
ws
se
rv
ic
es
c
onnecti
ons.
Al
so
,
it
’s
co
ns
i
de
red
a
hy
br
i
d
al
gorithm
of
pack
et
sche
du
le
r
an
d
qu
e
ue
m
anag
e
m
ent
to
av
oid
co
ng
e
sti
on
co
ntr
ol.
T
he
F
Q
-
CoD
el
is
nea
rly
par
am
et
er
-
le
ss
t
o
i
m
pr
ove
the
p
e
rfor
m
ance
an
d effici
ency o
f
c
ongestio
n
c
ontrol.
T
his
al
gori
thm
cl
assifi
es
the
i
nco
m
ing
pac
kets
(the
pac
kets
flow
ser
vices
di
sti
nguish
e
d
ba
se
on
IP
ad
d
re
ss
an
d
port
nu
m
ber
of
both
peer
s
pa
ram
eter
s)
in
m
ul
ti
ple
diff
er
ent
qu
e
ues
,
th
en
ap
ply
the
CoD
el
on
eac
h
queue
.
The
FQ
-
C
oDel
cont
ai
ns
two
pa
rt
s,
the
sche
du
le
r
to
se
le
ct
the
que
ue
for
deque
ue
t
he
pac
kets,
a
nd
the
Co
Del
to
pe
rfor
m
the
co
ngest
io
n
c
on
t
ro
l
in
t
he
sel
ect
ed
que
ue. The
sch
e
dule
r part
giv
es
prio
r
it
y and
fairness
to
lo
w rat
e se
r
vices c
onnecti
on
s
.
An
a
dap
ti
ve
C
oD
el
with
inte
rv
al
tu
ning
“A
CoD
el
-
IT”
an
d
adap
ti
ve
Co
D
el
with
ta
rg
et
and
inte
rv
al
tun
in
g
“
AC
oDel
-
TIT”
al
gorit
hm
s
wer
e
pro
pose
d
f
or
a
uto
t
unin
g
t
he
par
am
et
er
to
sta
bili
ze
que
uing
delay
and
increase
the
perform
ance
of
the
netw
ork
[
38
]
.
Wh
il
e
the
ACoDel
-
TIT
has
a
n
add
it
io
nal
en
ha
nced
perform
ance
by
reducin
g
the
dro
p
pac
kets
and
high
ba
ndwidth
util
iz
at
i
on.
Bot
h
al
gor
it
h
m
s
hav
e
th
e
sam
e
proce
dure
of
C
oD
el
for
cal
culat
ing
dro
p
pro
bab
il
it
y.
The
Co
Del
has
fix
par
am
et
er,
w
hile
the
a
da
ptive
al
gorithm
has
an
autotunin
g
par
am
et
er
(interv
al
an
d
ta
rg
et
tim
e)
based
on
netw
ork
co
ndit
ion
s
.
In
th
e
ACoDel
-
IT
on
ly
the
interval
has
been
a
da
ptive
,
w
hile
AC
oD
el
-
TIT
both
interval
a
nd
t
arg
et
ti
m
e
hav
e
aut
o
tun
in
g.
ACoD
el
-
TIT
ca
n
m
or
e
eff
ect
iv
el
y
sta
bili
ze
the
qu
euei
ng
delay
and
pro
vid
e
hig
he
r
pe
rfo
rm
a
nce
in
li
nk
util
iz
at
ion
than
t
he ACo
D
el
-
IT.
5.
AQM
ALGO
RITH
MS
B
A
SED ON
QUE
UING LE
N
G
TH A
ND D
E
LAY
Ma
ny
serv
ic
e
s
and
a
pp
li
ca
ti
on
s
w
ork
over
the
netw
ork,
wh
ic
h
f
or
m
s
heteroge
neous
fl
ow
connecti
ons
[
39]
.
T
he
IPv4
he
ader
has
T
OS
an
d
tra
ff
ic
cl
ass
in
t
he
IPv6 heade
r
t
o
diffe
ren
ti
at
e
these
s
erv
ic
es.
The
heter
ogen
ei
ty
resu
lt
s
f
rom
diff
eren
t
flo
w
rate
as
well
as
the
dif
fer
e
nt
end
to
en
d
d
el
ay
(
rou
nd
tri
p
ti
m
e
“R
TT”).
F
or
exam
ple
,
the
m
edia
strea
m
i
ng
a
ppli
cat
ion
(e.g.,
vid
e
o)
has
a
hi
gh
e
r
flo
w
rate
tha
n
oth
er
app
li
cat
io
ns
,
w
hile
the
sat
el
lit
e
com
m
un
ic
at
i
on
has
a
la
rg
e
RTT
than
ot
he
r
connecti
ons.
Du
e
to
this
va
r
ia
ti
on
of
flo
w
rate
a
nd
RTT,
t
he
c
ongestio
n
c
on
t
ro
l
al
go
rithm
s
base
d
on
t
he
delay
of
a
pa
cket
are
re
qu
ir
ed
to
i
m
pr
ove alg
ori
thm
s
'
p
erfor
m
a
nce
base
d on queue
len
gth
,
s
uc
h
as
RED
schem
a.
The
en
ha
nced
rand
om
early
detect
ion
“E
N
RED”
[40]
wa
s
us
ed
to
m
ini
m
iz
e
the
qu
e
ue
delay
and
loss
rate
o
f
t
he
pac
ket
by
kee
ping
t
he
a
ver
a
ge
queue
siz
e
avg
m
ini
m
a
l.
The
l
ow
pass
f
il
te
r
was
us
e
d
t
o
m
ake
the
que
ue
m
or
e
sta
ble
by
al
lowi
ng
t
he
al
gorithm
to
ta
ke
act
ion
to
be
m
eaningfu
l.
R
eact
ing
fa
ste
r
l
eads
t
o
os
ci
ll
at
ion
s
a
nd
insta
bili
ty
wh
il
e
res
pondin
g
m
or
e
sl
owly
m
akes
the
sys
tem
ta
rd
y.
The
ENRE
D
de
pe
nd
s
on
qu
e
ue
wei
gh
t
ba
sed
on
que
ue
siz
e
and
bu
rst
delay
in
a
qu
e
ue
.
This
al
gorith
m
red
uces
the
avg
of
the
qu
e
ue
by
us
in
g
a
sm
all q
ueu
e
size, s
o, this le
ads
to
le
s
s d
el
ay
as
well
as a lo
w
l
os
s
ra
te
.
Delay
-
co
ntr
olle
r
ra
ndom
early
detect
ion
“
D
cR
ED”
[
41]
is
an
e
xtension
of
t
he
RED
al
gorithm
by
us
in
g
the
delay
featur
e
wh
il
e
pr
ese
rv
i
ng
th
e
or
igi
nal
cha
r
act
er
su
c
h
as
min
th
an
d
max
th
.
The
m
od
ific
at
ion
in
DcRED
u
si
ng
a d
el
ay
of a
p
a
cket in t
he q
ue
ue
to
calc
ulate
the dr
op pr
oba
bili
ty
.
An
a
dap
ti
ve
AQ
M
[
42]
wo
r
k
base
d
on
trade
-
off
bet
ween
qu
e
uing
delay
s
and
li
nk
util
iz
at
ion
“TO
DU
”
.
The
TODU
go
al
is
to
achieve
l
ow
qu
e
ue
delay
w
hen
dif
fer
e
nt
ty
pes
of
ser
vice
s
are
co
nnect
ed
ove
r
the
netw
ork.
In
TODU
,
t
he
vir
tual
qu
e
ue
was
m
ai
ntain
ed,
w
hich
has
a
capa
ci
ty
s
m
al
le
r
than
the
act
ual queu
e
.
The
m
ark
/dr
op
packet
s
in
the
act
ual
queue
are
do
ne
w
he
n
the
virt
ual
qu
eue
ove
rf
l
ow
s
,
wh
il
e
the
a
rr
i
ving
pack
et
s
int
o
the
act
ual
qu
eue
will
up
date
the
sta
te
of
the
virtu
al
qu
e
ue.
Wh
e
n
the
ne
w
pack
et
ar
ri
ves
with
a
siz
e
that
exceeds
the
avail
abl
e
capaci
ty
in
a
virtu
al
queue
,
this
pack
et
w
il
l
m
ark
/dr
op
t
o
the
act
ual
queue;
oth
e
rw
ise
,
it
adds a
nd the
n u
pdat
es the
v
irt
ua
l qu
e
ue.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
23
, N
o.
3
,
Se
ptem
ber
2
0
21
:
14
19
-
14
31
1426
The
RE
D
-
e
xp
on
e
ntial
te
chn
i
qu
e
“R
ED
-
E”
[43]
was
pro
pose
d
as
a
m
odifie
d
RED
al
gor
it
hm
.
The
RED
-
E
use
s
non
le
ane
r
beh
a
vior
to
dro
p
pa
ckets
to
m
anag
e
heter
ogene
ou
s
ser
vices
flo
w
ty
pes.
T
he
RE
D
-
E
us
e
d
min
th
,
max
th,
and
insta
ntaneous qu
e
ue
l
eng
t
h
to m
easur
e the con
gestion
le
vel.
Wh
en
avg
e
xceed
t
he
min
th
the drop pr
ob
a
bili
ty
calc
ulatio
ns
b
ase
d on
pa
ckets ar
riving
r
at
e o
r
RT
T by
u
sing
t
he
ex
pone
ntial
f
orm
ula. Th
e
higher
ar
rivi
ng
rate
m
eans
that
the
flow
ha
s
a
hig
he
r
del
ay
in
the
qu
e
ue
.
In
(
6
)
an
d
(
7
)
s
how
the
R
ED
-
E
m
od
ifie
d
from
(
1
)
of
RE
D
(
r
e
m
ov
e
the
de
pe
nd
e
nt
on
t
he
sta
ti
c
value
of
max
p
),
w
her
e
the
C
is
the
pa
ckets
arr
ivi
ng
rate t
ha
t base
on the
RTT o
f
the
f
l
ows.
(6)
(7)
6.
AQM W
ITH
INTEL
LIGE
NT OPTI
MIZ
ATIO
N
To
im
pr
ov
e
th
e
AQ
M
al
gori
thm
schem
as,
an
intel
li
gen
t
proce
dure
was
com
bin
ed
with
it
.
Ma
ny
intel
li
gen
t
asp
ect
s
wer
e
avai
la
ble
su
ch
as
m
achine
le
arn
i
ng,
ne
ur
al
net
work,
ge
netic
al
gorithm
and
fu
zzy
value.
T
he
inte
ll
igent
m
et
ho
d
is
us
ed
to
en
ha
nce
the
A
QM
per
f
orm
ance
by
op
ti
m
iz
ing
the
sel
ect
and
tun
in
g
par
am
et
er dur
i
ng h
et
e
roge
neous
tra
ff
ic
flo
ws
to red
uce t
he de
la
y m
or
e tha
n
tra
diti
on
al
A
QM.
6.1.
AQM
w
ith n
e
ural
netw
ork alg
orith
m
The
ne
ural
ne
twork
“
NN
”
m
i
m
ic
s
the
hum
an
br
ai
n
-
behavio
r
f
or
le
ar
ni
ng
a
nd
decisi
on
m
aking;
recently
,
NN
was
us
e
d
with
the
A
QM
a
lgorit
hm
to
achieve
t
he
que
ue
le
ngth
sta
bi
li
ze
with
hi
gh
l
in
k
util
iz
at
ion
.
N
N'
s
essenti
al
par
ts
are
inputs
,
weig
hts,
act
ivati
on
f
unct
io
n,
an
d
outp
ut,
wh
ere
the
act
ivati
on
functi
on
perfor
m
s
su
m
m
ing
c
al
culat
ion
w
hile
wei
gh
t
is
up
dated
to
ac
hie
ve
op
ti
m
iz
ation
decisi
on.
Th
e
N
N
can
be
cl
assifi
ed
base
d
on
ne
uro
n
interc
on
necti
ons
(s
uch
as
Feedforwa
rd,
Feed
bac
k,
and
Re
c
urren
t
)
an
d
weig
ht
update
le
arn
in
g
(e
.g.,
su
pe
r
vised
,
un
su
pe
r
vised
,
rei
nfor
cem
ent
le
arn
i
ng).
m
any
research
st
ud
ie
s
hav
e
been
a
ppli
ed
t
he
N
N
with
A
QM
to
e
nh
a
nc
ed
co
ngest
io
n
con
t
ro
l
perfor
m
ance
and
pre
dict
fu
t
u
re
c
on
gestio
n
le
vels.
The
fee
d
-
for
ward
ne
ur
al
netw
ork
A
Q
M
“FF
NN
-
A
QM”
[
44]
w
as
de
sig
ned
to
deal
wit
h
heter
og
y
nous
traf
fics
fl
ow
an
d
to
pr
e
dict
th
e
f
uture
qu
e
ue
le
ngth
value
.
The
FF
NN
-
A
QM
weig
hts
updat
e
base
d
on
tim
e
-
va
ryi
ng
to
ac
hieve
sta
bili
ze
qu
e
uein
g
le
ngth.
It
adopts
sel
f
-
le
ar
ning
to
predict
f
utu
r
e
qu
e
ue
le
ng
th
us
in
g
a
n
act
ivati
on
f
unct
ion
(in
this
stud
y,
act
ivati
on
functi
on
wa
s
a
so
ft
sig
n).
As
sho
wn
i
n
F
igure
5,
the
fee
db
ac
k
c
on
t
ro
l
al
go
rith
m
,
as
descr
ibe
d
in
t
he
pre
vious
sect
io
n,
m
app
e
d
to
the NN
co
n
tr
oller
to
cal
culat
e
the
pro
bab
il
it
y
dro
p.
T
her
e
is
two
i
nput
(tar
ge
t
qu
e
ue
an
d
cu
rr
e
nt
queue
le
ngth
),
w
he
re
the
con
sta
nt
gain
us
e
d
as
wei
gh
t
a
nd
t
his
weig
ht
up
da
te
by
us
in
g
gradient
desce
nt.
The
sta
bili
ty
of
qu
e
ue
le
ng
t
h
in
this
al
gorith
m
is
achieve
d
by
m
ai
ntainin
g
the
current
que
ue
le
ng
t
h
cl
os
e
to
the
ta
rg
et
qu
e
ue
.
In
this
al
gor
it
h
m
,
in
add
it
ion
to
sta
bili
ty
,
the
set
tl
ing
tim
e
beco
m
es
faster,
as
well
as
hi
gh
ut
il
iz
ation
of
ba
ndwidt
h
a
nd
l
ow
de
la
y.
I
n
[45]
,
the
exp
li
ci
tl
y
cong
est
ion
no
ti
ficat
ion
“EC
N”
w
as
us
e
d
with
Re
inforcem
ent
le
arn
i
ng
a
nd
the
decisi
on
m
aking
base
d
on
in
ferred
rest
of
pat
h
co
ng
est
i
on
to
tun
i
ng
the
a
lgorit
hm
par
am
et
er
(w
hich
is
based
on
m
arko
v
decisi
on
proce
ss
“M
DP
”
).
T
his
al
gorit
hm
has
a
set
of
le
vel
co
ngest
io
n
sat
e
as
well
a
s
a
set
of
ac
ti
on
s
to
achieve t
he
ta
r
get p
a
ram
et
er.
(a)
(b)
Figure
5. Co
nvert the
A
QM f
e
edb
ac
k
c
ontr
oller to
ne
ur
al
c
ontr
oller
;
(a
)
A
QM f
ee
dback
con
t
ro
ll
er
,
(b) Neu
ral
network
c
on
tr
oller
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Su
r
vey
on:
A
v
ar
ie
ty
o
f A
QM
algorit
hm sc
he
ma
s
an
d
i
ntell
i
gen
t t
ec
hn
i
que
s d
evel
op
e
d…
(
Ama
r
A.
Ma
hawi
sh
)
1427
6.2.
AQM wi
th
gene
tic a
l
gori
th
m
A
ge
netic
al
go
rithm
“GA
”
te
chn
iq
ue
is
use
d
to
searc
h
op
ti
m
iz
ation
f
ro
m
the
po
pula
ti
on
(
best
so
luti
on)
, whic
h
f
ollo
ws
D
ar
win'
s n
at
ural
e
vo
l
ution t
he
or
y
. Th
e
popula
ti
on r
e
pr
e
sents as
a chrom
os
om
e
(su
c
h
as
bin
a
ry
enc
odin
g),
an
d
t
he
op
e
rati
ons
use
d
in
G
A
are
reprod
uction,
sel
ect
ion
,
c
ro
s
so
ve
r
a
nd,
m
utati
on
.
m
any rese
arc
h and st
udie
s in
netw
ork
c
onge
sti
on
c
ontrol
w
it
h
G
A help,
s
uc
h
as
AQM
with
GA.
The
ge
netic
alg
ori
thm
with
pro
portion
al
a
nd
inte
gr
al
co
ntr
oller
“GA
-
PI
”
[
46
]
was
desig
ne
d
to
op
ti
m
iz
e
the
PI
c
ontr
oller'
s
so
l
ution.
As
discusse
d
ab
ove
,
t
he
PI
co
ntr
oller
is
cl
ose
d
-
l
oop
fee
dback
to
el
i
m
inate
the
e
rror
stu
dy
sta
te
“e”
caus
ed
by
la
te
ncy
of
di
ff
e
ren
t
flo
w
c
onnecti
ons
,
a
nd
this
act
io
n
by
us
in
g
two
facto
rs
co
ns
ta
nt
gai
n
K
p
and
K
i
.
I
n
G
A
-
PI
al
gorithm
us
ed
G
A
to
opti
m
iz
e
the
sel
ection
of
these
fa
ct
or
s
,
as
sho
wn
in
Fi
gure
6.
T
he
tw
o
P
I
facto
rs
re
pr
ese
nted
as
bi
nar
y
c
oncat
ena
ti
on
,
G
A
opera
ti
on
s
w
ork
by
sel
ect
higher
fitnes
s
after
that
the
cro
ss
over
a
nd
m
uta
ti
on
app
li
ed
that
al
low
a
var
ie
ty
of
popula
ti
on
(
K
p
an
d
K
i
)
to
sel
ect
the
best
.
Als
o
the
re
i
s
m
any
oth
er
stud
y
in
t
his
f
ie
ld
su
c
h
as
a
nt
col
on
y
op
ti
m
iz
at
ion
“ACO”
t
o
op
ti
m
iz
e
tun
in
g
PID AQM p
aram
et
ers
[47]
.
W
hile other
st
ud
y use
d
pa
rtic
le
swar
m
o
pti
m
iz
at
ion
(
PSO)‐base
d
PI
D
(
pro
portio
nal
–
inte
gr
al
–
de
rivati
ve
)
“PS
OP
I
D”
[48]
.
The
P
SO
us
e
d
to
fin
d
bes
t
posit
ion
from
globa
l
po
sit
io
n
by
us
i
ng
vel
ocity
an
d
po
sit
io
n.
The
se
featu
re
of
P
SO
us
e
d
to
sel
ect
approp
riat
e
par
am
et
ers
of
PI
D
’s
gain
to
reac
h
z
ero
ste
a
dy
sta
te
error
i
n
dif
fe
ren
t
ty
pe
of
se
rv
ic
es.
The
al
gorithm
resu
lt
sh
ow
lo
w
in
ris
e
tim
e
and re
duced
th
e sett
li
ng
ti
m
e
and this m
ean lo
w
d
el
ay
a
nd fa
ste
r
set
tl
ing
ti
m
e.
Figure
6. G
A
-
PI
c
ongestio
n con
t
ro
ll
er
6.3.
AQM wi
th
F
uz
z
y
Log
i
c a
lg
orit
hm
The
fu
zzy
l
og
i
c
“FL”
have
m
e
m
ber
sh
ip
va
lues
be
twee
n
0
a
nd
1
w
hile
crisp
val
ue
has
on
ly
tw
o
value
on
ly
0
a
nd
1
or
tr
ue
a
nd
false. FL
f
ollow
s
t
he
if
-
el
se r
ule
that
ta
kes
the
decisi
on b
a
sed
on
im
pr
eci
se
an
d
vague
ness
i
nfo
rm
ation
.
Ma
ny
w
orks
im
plem
ent
the
FL
w
it
h
the
A
QM
a
lgorit
hm
to
en
han
ce
the
co
ngest
io
n
con
t
ro
l
par
am
et
er
an
d
achie
ve
the
opti
m
iz
a
ti
on
res
ult.
T
he
FL
ta
kes
a
c
risp
value
as
i
nput
(
for
e
xa
m
ple,
x
value
)
the
n
tr
ansfo
rm
s
it
into
a
fu
zzy
set
by
s
plit
ti
ng
i
t
into
dif
fer
e
nt
sta
ges
as
in
Table
1
to
optim
iz
e
decisi
ons
[49]
.
Table
1.
Fu
zzy
m
e
m
ber
sh
ip val
ue
Stag
e
Descripti
o
n
LN
MN
SN
Z
SP
MP
LP
x
is
Lar
g
e
Neg
ativ
e
x
is M
ed
iu
m
Neg
a
tiv
e
x
is S
m
all
Negativ
e
Zer
o
x
is S
m
all
Pos
itiv
e
x
is M
ed
iu
m
Po
sitiv
e
x
is L
arge
Po
sitiv
e
The
f
uzzy
pro
portio
nal
integ
ral
de
rivati
ve
(
FPID)
c
ontr
oller
[
50
]
,
[
51]
w
as
desig
ne
d
to
op
ti
m
iz
e
the
PI
D
-
AQ
M
al
gorithm
par
am
e
te
rs,
a
nd
these
pa
ram
et
ers
are
three
co
ns
ta
nt
gain
K
p
,
K
i
a
nd
K
d
.
Wh
e
re
the
FL
adjuste
d
the
se
par
am
et
ers
to
achieve
opti
m
al
con
t
ro
l
on
c
ongestio
n
w
he
n
a
dif
fer
e
nt
t
ype
of
s
er
vice
flo
ws
us
e
d
ov
e
r
a
ne
twork
.
T
he
e
rror
“e”
is
t
he
di
ff
ere
nce
betw
een
act
ual
an
d
desire
d
que
ue
le
ng
t
h,
an
d
e
c
is
th
e
change
i
n
fl
ow
rate,
bo
t
h
e,
a
nd
e
c
us
e
d
as
input
to
FL
bloc
k,
a
nd
the
t
hree
co
ns
ta
nt
gain
as
optim
iz
in
g
ou
t
pu
t
of this
blo
c
k
as
shown i
n
Fi
gu
re
7.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
23
, N
o.
3
,
Se
ptem
ber
2
0
21
:
14
19
-
14
31
1428
Figure
7. FP
ID co
ng
e
sti
on contr
oller
Ther
e
are
m
any
oth
er
stu
dy
that
m
ixed
two
or
m
or
e
op
ti
m
i
zat
ion
te
ch
nique
to
im
pr
ove
op
ti
m
iz
ation
of
AQ
M
par
a
m
et
er.
In
[
52]
fu
zzy
pr
opor
ti
on
al
-
I
nteg
ral
(
FPI)
co
ntr
oller
was
desi
gn
e
d
with
ge
netic
al
gorith
m
(GA)
us
e
d
to
t
un
i
ng
t
he
F
PI
par
am
et
ers.
Th
e
FPI
ca
n
ac
hi
eve
go
od
pe
rfor
m
ance
su
c
h
as
the
de
sired
qu
e
ue
le
ng
th
, fast
r
es
pons
e
and
high
li
nk
util
iz
at
ion.
Wh
il
e the
GA
u
se
d
t
o op
ti
m
i
ze in selec
t t
he
FP
I
p
a
ram
et
ers.
In
[
53]
us
ed
F
L
con
tr
oller
wi
th
par
ti
cl
e
swa
rm
op
tim
iz
at
ion
(PSO
),
s
ocial
sp
ider
optim
izati
on
(SSO
)
and
a
nt
col
ony
op
ti
m
iz
a
ti
on
(A
CO
)
al
gori
thm
s
then
m
a
ke
com
par
iso
n
to
show
how
op
ti
m
iz
e
PI
D
gains
.
Wh
il
e
in
[
54
]
desig
n
a
li
near
qu
a
dr
at
ic
(L
Q
)
-
se
rvo
co
ntr
ol
le
r
as
an
AQ
M
and
it
s
par
am
et
ers
tun
e
d
by
us
in
g
the
par
ti
cl
e
sw
arm
op
tim
iz
at
i
on
(
PS
O)
m
et
ho
d,
w
hich
pro
vi
de
m
or
e
sta
bili
zat
ion
with
fa
ste
r
set
tl
ing
tim
e
and
sm
a
ll
d
el
ay
.
7.
COMP
AR
I
S
ON
BE
TWE
E
N
THE
DISC
US
SE
D ALG
ORI
TH
MS
This
re
view
c
om
par
e
the
al
gorithm
s
discuss
ed
ab
ove
with
each
ot
her
t
o
analy
ze
the
im
pr
ov
em
ent
perform
ance
of
these
al
gorithm
s.
In
[55]
the
c
om
par
ison
ha
ve
bee
n
hol
d
based
on
vary
ing
the
pro
pa
gati
on
delay
an
d
the
li
nk
ca
pacit
y.
Wh
il
e
in
this
st
ud
y
,
the
A
QM
schem
as
are
br
ie
fly
discuss
e
d
an
d
c
om
par
ed
base
d
on
sev
e
ral
m
etr
ic
s,
as
s
how
n
in
Table 2.
T
he
Lo
w
“L”,
M
oderate
“M
”, H
i
gh
“
H”,
a
nd
V
ery
High
“
V.H
”
hav
e
been
use
d
t
o
a
ssign
the
value
of
a
s
pecific
a
lgorit
hm
deal
with
the
m
et
ri
c.
Fair
ness
m
eans
how
a
n
al
gorithm
achieves
fairness
a
m
on
g
dif
fer
e
nt
flo
ws
(
TCP
-
f
rien
dly
or
UDP
ag
gr
e
ssive
flo
w),
w
hile
the
com
pl
exity
of
tun
in
g
par
am
eter
s
of
this
al
go
rithm
.
Each
al
gorithm
can
achieve
th
e
Th
r
ough
pu
t,
Lin
k
ut
il
iz
ation
,
a
nd
Loss
rate o
f pack
et
s.
Finall
y, the
stabil
it
y of
a
n
al
gorithm
o
btain
ed wh
e
n flo
w dr
am
at
ic
ally c
hange
d wit
hin
tim
e.
Table
2.
C
om
par
iso
n of A
QM
algorit
hm
s
AQM
Metho
d
Alg
o
rith
m
na
m
e
Fairnes
s
Co
m
p
lex
it
y
Thro
u
g
h
p
u
t
Link
utilizatio
n
Los
s Rate
Qu
eu
e Stability
Qu
eu
ein
g
L
en
g
th
RED
L
H
L
M
H
L
FRED
H
V.H
H
H
L
M
W
RED
L
H
M
M
M
M
ARED
L
H
M
H
M
H
GRED
L
H
M
L
M
M
EAGRE
D
L
H
H
M
M
M
DGRED
L
H
H
M
L
H
SDGRED
L
V.H
H
M
L
V.H
CHOKe
M
M
M
M
M
M
CHOKe
-
RH
H
V.
H
H
M
L
H
HTCB
H
V.H
H
M
M
L
Qu
eu
in
g
Delay
PIE
M
M
M
H
L
V.H
D
-
PAC
H
H
M
H
L
V.H
PD
-
AQ
M
M
H
M
H
M
H
Co
Del
M
L
M
H
M
L
FQ
-
Co
Del
V.H
M
H
M
M
L
ACo
Del
-
IT
M
M
H
H
M
H
ACo
Del
-
T
IT
H
M
H
H
M
H
Qu
eu
in
g
Leng
th
&
Delay
ENRED
M
M
M
M
M
H
DcRED
M
M
H
M
L
M
TODU
L
H
H
H
L
H
RED
-
E
L
M
L
M
M
H
Ther
e
is
not
e
asy
to
say
wh
i
ch
al
go
rithm
i
s
best
beca
us
e
each
al
gorith
m
is
dev
el
op
e
d
to
deal
with
app
li
cat
io
ns
,
s
erv
ic
es,
an
d
fl
ow
rate.
Th
e
m
ai
n
cl
assifi
cat
ion
of
A
QM
s
chem
as
was
ei
ther
base
d
on
qu
e
uing
Evaluation Warning : The document was created with Spire.PDF for Python.