1
1
The R
ela
tive Q
ueuin
gD
ela
y
of
Para
llel Packet
Sw
itches
Hag
it A
ttiya
CS
, Tec
hnio
n
Join
t wor
k w
ith D
avid
Hay
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
2
Outp
ut
Queued S
witches
Cel
ls a
re q
ueue
d at
th
e ou
tput
ports
whe
n th
ey a
rriv
e
Sw
itch
mus
t ope
rate
at
agg
rega
tein
put
(=ou
tput
) lin
e ra
te!
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
3
Oth
er
Sw
itch A
rchitectu
res
Inpu
t que
uing
(IQ
)C
ells
are
que
ued
at th
e in
putp
orts
Hea
d of
line
bloc
king
Virtu
alou
tput
-que
uing
(VO
Q)
Each
inpu
t hol
ds a
sep
arat
e (v
irtua
l) qu
eue
for e
ach
outp
utC
ompl
ex (=
expe
nsiv
e) s
ched
ulin
g al
gorit
hms
Com
bine
d in
put-o
utpu
t que
uing
(CIO
Q)
Que
uing
bot
h in
inpu
ts a
nd o
utpu
tsM
emor
y / c
ontro
l in
two
loca
tions
Stil
l req
uire
mem
ory
/ dec
isio
nsp
eed
≥in
put l
ine
rate
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
4
Para
llel Packet
Sw
itch
(PPS)
[Iye
ran
d M
cKeo
wn]
3-st
age
Clo
s ne
twor
k w
ith K
mid
dle
stag
e pl
anes
Fixe
d si
ze p
acke
ts
(cel
ls)
switc
hed
in p
aral
lel
thro
ugh
the
plan
esfra
gmen
tatio
n an
d re
asse
mbl
y ou
tsid
e th
e sw
itch
Pla
ne
1
Pla
ne
K
1 2 NN21
. . .. . .
. . .
2
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
5
Para
llel Packet
Sw
itch
(PPS)
Eac
h pl
ane
is a
nN☓
N s
witc
hW
ith in
put r
ate
r < R
Can
be
an o
utpu
t-qu
eued
sw
itch
The
over
-cap
acity
of
the
switc
h is
its
spee
dup
S =
Kr/R
RRR
Rr
r rr
Pla
ne
1
Pla
ne
K
1 2 NN21
. . .. . .
. . .
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
6
Para
llel Packet
Sw
itch
(PPS)
Dem
ultip
lexo
rat t
he in
put
port
deci
des
whe
re to
sen
d an
inco
min
g ce
llC
ells
arr
ive
and
leav
e at
di
scre
te ti
me
slot
s A
t lea
st R
/r tim
e-sl
ots
betw
een
two
cells
from
an
inpu
t to
a pl
ane
from
a p
lane
to a
n ou
tput
As
in In
vers
e M
ultip
lexi
ng
for A
TMR
RRR
rr r
r
Pla
ne
1
Pla
ne
K
1 2 NN21
. . .. . .
. . .
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
7
PPS P
erf
orm
ance F
igure
s
The
lag
behi
nd a
shad
ow o
utpu
t-que
ued
switc
hO
pera
ting
at ra
te R
Rec
eivi
ng th
e sa
me
traffi
cR
elat
ive
queu
ing
dela
y (R
QD
) Im
pact
of q
ueui
ng w
ithin
the
switc
h N
egle
ct th
e im
pact
of d
iffer
ent p
ropa
gatio
n de
lays
Com
pete
tive
(rel
ativ
e) m
easu
reBu
t add
itive
(rat
her t
han
mul
tiplic
ativ
e)
Rel
ated
to th
ece
ll de
lay
jitte
r, an
d bu
ffer s
ize
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
8
This
Work
How
(lac
k of
) inf
orm
atio
n af
fect
s th
e R
QD
Low
eran
d up
per b
ound
s
Dep
end
on d
emul
tiple
xor t
ype
With
/ w
ithou
t buf
fers
Cen
traliz
ed: G
loba
l inf
orm
atio
nFu
lly-d
istri
bute
d: O
nly
loca
l inf
orm
atio
n u-
Rea
l tim
e (u
-RT)
:Loc
al in
form
atio
n an
d gl
obal
info
rmat
ion
olde
r tha
n u
time
slot
s
3
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
9
PPS w
/ C
entr
alize
d C
ontr
ol
[Iye
ran
d M
cKwo
en]
Find
free
link
s be
twee
n in
puts
an
d pl
anes
at a
rriva
l tim
eC
alcu
late
cel
l dis
patc
hing
tim
ein
th
e sh
adow
OQ
sw
itch
Find
free
link
s be
twee
n pl
anes
an
d ou
tput
s at
this
tim
eD
ispa
tch
cell
thro
ugh
a co
mm
on
plan
e (a
t the
targ
et ti
me)
Mus
t exi
stsi
nce
spee
dup
≥2
Zero
RQ
D, b
ut…
Rel
ies
on k
now
ing
the
stat
us o
f al
l inp
ut-p
orts
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
10
Tra
ffic
Restr
ictions
Low
erbo
unds
hol
d ev
en fo
r lea
kybu
cket
flow
s:
No
mor
e th
anτ
R+B
cel
ls s
harin
g an
inpu
t-or
ou
tput
-por
t arr
ive
to th
e sw
itch
durin
g τ
time.
⊳Al
so u
nder
Adve
rsar
ialQ
ueui
ng T
heor
yre
stric
tions
.
Num
ber
of ce
lls d
estined
fo
r a
cert
ain o
utp
ut
For
any
tim
e in
terv
al,
the
scat
tere
dar
ea is
larg
er t
han
th
e gre
y ar
ea b
y at
most
B
time
cells
R
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
11
Concentr
ation S
cenari
o
(Sim
plified)
Assu
me:
Buffe
r at o
utpu
t jis
em
pty
Exa
ctly
m c
ells
for j
arriv
e du
ring
[t,t+
m)
All c
ells
are
sen
t thr
ough
pla
ne k
RQ
D≥
m(R
/r-1)
⊳Ex
act s
et-u
p ne
eded
in th
e ge
nera
l cas
eo
Acco
unt f
or b
urst
so
Pass
age
of ti
me
Can
als
o (lo
wer
) bou
nd th
e ce
ll de
lay
jitte
r.
. . .. . .
. . .j
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
12
Fully D
istr
ibute
d,
Buff
erl
ess L
ow
er
Bound
RQ
D ≥
N[(R
/r)-1
]for
an
unpa
rtitio
ned
algo
rithm
All
dem
ultip
lexo
rs m
ay s
end
a ce
ll to
out
put j
thro
ugh
the
sam
e pl
ane
kTh
e pr
oof c
reat
estra
ffic
that
sen
ds N
cells
th
roug
h th
e sa
me
plan
e to
the
sam
e ou
tput
W
ithin
shor
t tim
eC
lean
buf
fers
initi
ally
Sta
gger
the
cells
to a
void
bur
sts
App
lyth
e co
ncen
tratio
n le
mm
a.
4
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
13
Low
er
Bound o
n t
he
Avera
ge R
QD
The
prev
ious
pro
of c
reat
es
traffi
c in
whi
ch s
ome
cell
has
RQ
D ≥
N[(R
/r)-1
]S
ame
RQ
D fo
r fol
low
ing
cells
with
the
sam
e de
stin
atio
n Si
nce
the
switc
h m
ust p
rese
rve
orde
r with
in fl
owFo
r a s
uffic
ient
ly lo
ng tr
affic
aver
age
RQ
D ≥
N[(R
/r)-1
]-εfo
rarb
itrar
ily s
mal
l ε
Cel
l with
hig
h R
QD
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
14
d-Part
itio
ned A
lgori
thm
s
RQ
D ≥
d[(R
/r)-1
]if
dde
mul
tiple
xors
sen
d a
cell
dest
ined
for o
utpu
t j
thro
ugh
plan
e k
sam
e pr
oof…
Inpu
tcon
stra
ints
impl
y d
≥N
/S
RQ
D≥
[N/S
][(R
/r)-1
]Ev
en if
pla
nes
are
a-pr
iori
parti
tione
d
k
j
. . .. . .
. . .
d inputs
use
k
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
15
Fully D
istr
ibute
d,
Buff
erl
ess U
pper
Bound
Frac
tiona
l tra
ffic
disp
atch
(FTD
) alg
orith
m[I
yer
and
McK
eown
, 200
1] [K
hoti
msk
y, K
rish
nan,
200
1]
Div
ide
the
flow
of c
ells
into
blo
cks
of le
ngth
R/r
Two
cells
in th
e sa
me
bloc
k ar
e se
nt th
roug
h di
ffere
nt p
lane
sne
eds
a sp
eedu
p ≥
2H
as R
QD
≈(R
/r)N
Proo
fsar
e fla
wed
.I&
M b
ound
que
uing
in th
e m
ultip
lexo
r (ou
tput
) an
d ig
nore
queu
ing
in th
e pl
anes
K&
K ig
nore
per
iods
whe
n th
e P
PS
is b
usy
and
the
shad
ow is
idle
, and
vic
e ve
rsa
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
16
New
Pro
of
for
FT
D
Our
pro
of m
etho
dolo
gy is
bas
ed o
n bo
undi
ng th
e im
bala
nce
betw
een
plan
esAp
plic
able
to o
ther
type
s of
dem
ultip
lexo
rsA
ssum
esa
glob
al F
CFS
disc
iplin
ebo
th o
n th
e P
PS
and
the
shad
ow s
witc
hce
llsar
rivin
gaf
ter a
cel
l cdo
not
dela
y it
We
prov
e th
at s
ome
cell
c th
at s
uffe
rs m
axim
um
RQ
D is
notq
ueue
d in
the
mul
tiple
xor
It su
ffice
s to
bou
nd th
equ
euin
g w
ithin
the
plan
es
alth
ough
ther
e qu
euin
g in
mul
tiple
xors
is p
ossi
ble
5
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
17
How
man
y ce
lls (a
bove
fair
shar
e) a
re s
ent t
hrou
gh p
lane
k,
dur
ing
a tim
e in
terv
al [t
1,t2
]
Imba
lanc
eat
tim
e t,
∆=
max
imum
of ∆
jk (t)
over
all
j, k
and
t.
The Im
bala
nce F
acto
r
),
()
,(
),
(2
12
12
1t
tA
Rrt
tA
tt
jk j
k j−
=∆
),
(max
)(
11
tt
tk j
tk j
∆=
∆
Num
ber o
f cel
ls
dest
ined
forj
Num
ber o
f cel
ls
dest
ined
for j
and
sent
th
roug
h pl
ane
k
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
18
Theo
rem
:
In fa
ct, d
epen
ds o
n th
e m
ax tr
ansi
ent b
urst
Max
imum
num
ber o
f cel
ls w
ith s
ame
dest
inat
ion
arriv
ing
at th
e sa
me
time-
slot
Can
bou
nd th
e R
QD
by
boun
ding
∆
Imbala
nce F
acto
r and R
QD
()
++
∆≤
N1
,0max
rRRQD
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
19
Pro
of
for
FT
D
Hen
ce, t
he m
axim
um
RQ
Dis
Also
the
aver
age
rRrR
N
tt
ARr
rRrR
rR
tt
A
tt
ARr
rR
tt
A
tt
ARr
tt
A
tt
ARr
tt
At
t
jN i
ji
jN i
ji
jN i
kj
i
jk j
k j
1
),
(1
),
(
),
()
,(
),
()
,(
),
()
,(
),
(
21
12
1
21
12
1
21
12
1
21
21
21
−=
−
−+
≤
−
≤
−=
−=
∆
∑∑∑
=
→
=
→
=→
()1+
NrR
Asym
ptot
ical
ly m
atch
es th
e lo
wer
bou
ndC
lubN
et le
ctur
eM
ar-1
6-05
Par
alle
lPac
ket S
witc
h20
Eas
ier t
o st
art c
alcu
latio
ns w
hen
both
apl
ane
and
the
shad
ow s
witc
h ar
e id
le a
t tim
e-sl
ot
Pre
viou
s ce
lls a
lread
y le
ftW
ill n
otde
lay
futu
re c
ells
Nee
d to
con
side
r bus
ype
riods
in w
hich
eith
erth
e pl
ane
orth
e sh
adow
sw
itch
are
not i
dle
The
plan
eop
erat
esat
low
er ra
tePl
ane
busy
per
iod
can
belo
nger
The
plan
ere
ceiv
eson
ly a
sub
seto
f the
traf
fic
Shad
owsw
itch
busy
per
iod
is lo
nger
Busy P
eri
ods
6
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
21
1-Real
-T
ime, B
uff
erl
ess
Low
er
bound
If th
ere
are
no b
urst
sO
nly
one
inpu
t get
s a
cell
for o
utpu
t jFu
ll in
form
atio
n ab
out t
raffi
c to
out
put j
Can
sim
ulat
e ce
ntra
lized
alg
orith
m
With
bur
stin
ess
fact
or [N
/K]-1
RQ
D ≥
[N/S
][1-(
R/r)
]C
an b
e ex
tend
edto
ave
rage
RQ
D
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
22
Matc
hin
g U
pper
Bound
Div
ide
plan
esin
toeq
ual-s
ize
sets
Two
phas
es, e
ach
usin
g on
ly p
lane
s of
one
set
Aph
ase
ends
whe
n m
any
of it
s pl
anes
are
not
ba
lanc
ed(∆
≠0)
W
ith s
uffic
ient
spe
edup
, the
plan
es o
fthe
oth
er s
et
are
bala
nced
whe
n th
e ph
ase
ends
Spe
edup
S=8
yiel
dsR
QD
<4N
Asym
ptot
ical
ly m
atch
ing
the
low
er b
ound
(fo
r con
stan
t S)
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
23
Can R
andom
ization H
elp
?
Spre
adce
lls a
cros
s pl
anes
Cho
ose
at ra
ndom
to w
hich
elig
ible
pla
ne to
sen
d a
cell
Som
e ce
llsu
ffers
high
RQ
D
With
low
pro
babi
lity
(but
non
-neg
ligib
le)
Am
plifi
catio
n:m
any
trial
s un
til th
is h
appe
ns w
ith v
ery
high
pr
obab
ility
Perp
etua
te th
is h
igh
RQ
DSi
nce
the
switc
h m
ust o
bey
per-
flow
FC
FSAl
mos
t sam
e av
erag
eR
QD
as
dete
rmin
istic
alg
orith
ms
Rel
ies
on a
n ad
aptiv
ead
vers
ary
Wha
t hap
pens
whe
n th
e ad
vers
ary
is n
on-a
dapt
ive?
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
24
PPS w
ith B
uff
ers
at
the
Input-
Port
s
Del
ay is
the
sum
of q
ueui
ng in
the
inpu
t-po
rts a
nd in
the
plan
esD
emul
tiple
alg
orith
m is
mor
e fle
xibl
e| i
nput
buf
fer |
≥u:
RQ
D ≤
u B
lack
box
redu
ctio
n to
CP
A| i
nput
buf
fer |
< u
: RQ
D >
[N/S
][1-(
R/r)
], ev
en w
ith 1
-RT
algo
rithm
Fully
dis
tribu
ted
algo
rithm
s ar
e ∞
-RT
7
Clu
bNet
lect
ure
Mar
-16-
05P
aral
lelP
acke
t Sw
itch
25
Futu
re R
esearc
h
Ran
dom
izat
ion
agai
nst o
bliv
ious
, non
-ada
ptiv
ead
vers
ary
Mul
ticas
t tra
ffic
Par
tial,
out-d
ated
info
rmat
ion
abou
t the
sw
itch
Cre
dit-b
ased
con
trol
Som
e de
gree
of s
ynch
roni
zatio
n
26
T H
A N
K
Y O
U !