All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
006,
#####
Perfo
rman
ce co
mpari
son be
twee
n a 2
phase
s and
a 3 p
hases
Nego
tiatio
n pro
tocol
Antoi
ne Pi
chot,
Aleja
ndro
Gaspa
rAp
ril, 2
007
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
2| Pe
rf comp
betw
een a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
Agenda
1. Pr
oblem
2. Mo
del
3. Re
sults
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
3| Pe
rf comp
betw
een a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
The P
roblem
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
4| Pe
rf comp
betw
een a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
The context : Co-allocation
Multiple Computing farms
A (G)MPLS network
How to reserve two (or m
ore) reso
urces at
the same time ?
I.e. : A network connexion & a CPU
TORI
NO
GENO
VA
ALES
SAND
RIA PI
SA
MIL
ANO MIL
ANO2
BRES
CIA
SAVO
NA
BOLO
GNA
VERO
NAVICE
NZA VE
NEZI
A
FIREN
ZE
PERU
GIAAN
CONA
PESC
ARA
L’AQ
UILA
ROM
A
ROM
A2
NAPO
LISA
LERN
O
CATA
NZAR
O
POTE
NZA
BARI TA
RANT
O
CAGL
IARI
SASS
ARI
PALE
RMO
MES
SINA
REGG
IO C
.
CPUs
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
5| Pe
rf comp
betw
een a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
Solutions
•VIO
LA like (2Phases comm
it protocol)
•WS-Agreement based
Negotiation (3Phases Comm
it Protocol)
•Cf O.Waeldrich&
W. Ziegler draft @OGF18
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
6| Pe
rf comp
betw
een a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
3 Phases Comm
it Protocol vs 2 Phases
•Capability and availability check
•Pre
-reservation
•Resou
rce are rese
rved w
ith short reservation lifetim
e•
(No p
enalt
ies if
reserv
ation
is ca
ncell
ed at
this
stage
)
•Comm
itment
•Re
source
are r
eserve
d whe
never
need
ed•
(pena
lties
if can
celle
d)
Phase
2
Phase
1
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
7| Pe
rf comp
betw
een a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
VIOLA architecture
Aka A
RGON
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
8| Pe
rf comp
betw
een a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
VIOLA reservation process
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
9| Pe
rf comp
betw
een a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
VIOLA’s Meta-Scheduler reservation
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
10| P
erf co
mp be
tween a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
The M
odel
Expla
ined
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
11| P
erf co
mp be
tween a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
Model origin
Gurba
niV.K
.,Jag
adee
sanL.,
Mend
iratta
V.B.,
“Cha
racter
izing s
ession
initia
tion p
rotoc
ol (SI
P) ne
twork
perfo
rman
ce an
d rel
iability”,
Intern
ation
al ser
vice a
vaila
bility
sympo
sium,
April
2005
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
12| P
erf co
mp be
tween a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
Model explained(1)
Job requests �
Clients of the
queuing network
Time spent in a w
aiting queue �
Time spent in a state i
nsidethe
MS
A queue �
A state i
nsidethe
MS
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
13| P
erf co
mp be
tween a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
Model Explained (2), Notation convention
Meta-
Sche
duler
Crea
teAgre
emen
tµ 1
1/µ1is
the a
verag
e tim
e tak
en by
the M
eta-Sc
hedu
ler to
�re
ceive
the C
reate
Agre
emen
t mes
sage,
�pro
cess
it, an
d �ta
ke ac
tion
A Loc
al Sc
hedu
ler
Subm
itµ 2
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
14| P
erf co
mp be
tween a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
Model Explained (3), Example
Example synchronous
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
15| P
erf co
mp be
tween a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
2 Phases Negotiation protocol Model
Asynchronous
Synchronous
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
16| P
erf co
mp be
tween a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
3 Phases Negotiation protocol Model
Asynchronous
Synchronous
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
17| P
erf co
mp be
tween a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
Service Time & Error Probability example
After a few measures on a [email protected]
Hz
o��
�����
���
����
���…………………………………………………
���
o��
�����
� �
�����
������………………………………………………………
�����
��
o��
� ����
����
���……………………………………………………………
�����
��
�����
��
Probability to cancel a rese
rvation : 10%
Probability to need to look after scheduling horizon : 10%
Those
values can be modifie
d to take m
ore realistic values.
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
18| P
erf co
mp be
tween a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
Limiting Queue
)))(
((
1/2
2CB
AM
Qµ+
+−
=λ
ρ
For example (3 Phases Asynchronous) :
Arriv
al rat
e
Servi
ce ra
te
Prob
abilit
y to l
oop o
r to g
et an
erro
r
=−
=
Queues
Nb
kk
kN
11ρρ
Load
of
Queu
e k
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
19| P
erf co
mp be
tween a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
Resul
ts
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
20| P
erf co
mp be
tween a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
Maxim
um Job request Arrival Rate in the
Meta-Scheduler Alm
ost N
O IMPA
CT on
Arriv
al Rate
NEED
for
ASYN
CHRO
NOUS
Imple
mentati
on
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
21| P
erf co
mp be
tween a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
Mean number of Jobs in the
Meta-Scheduler
Threshold
2 Ph
ases
Threshold
3 Ph
ases
Synchronous
Asynchronous
Perfo
rmance loss
of 3P
NP ov
er 2P
NP
Less than 50
%
xPNP
: x Ph
ases N
egoti
ation
Proto
col
Arriv
al Rate
= 63 j
ob/s
Delay
~ 240m
s
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
22| P
erf co
mp be
tween a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
Conclusions
Need for an Asynchronous implem
entation (obvious)
Performance loss of a 3 Phases over a 2 Phases Negotiation protocol is
�less than 10% on the m
aximum job request arrival rate
�less than 50% on the total jo
b request processing delay and m
emory
requirement
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
23| P
erf co
mp be
tween a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
References
Article to be publishe
d soon…
by the end of the year
For a similar model use
d in a different context :
Gurba
niV.K
.,Jag
adee
sanL.,
Mend
iratta
V.B., “
Chara
cteriz
ing se
ssion
initia
tion
protoc
ol (SI
P) ne
twork
perfo
rman
ce an
d relia
bility”
, Inter
natio
nal se
rvice
avail
abilit
y sym
posiu
m, Ap
ril 20
05
All Ri
ghtsRe
serve
d©Alc
atel-L
ucent 2
007
24| P
erf co
mp be
tween a
2 Phase
s and a
3 Phases
Negotia
tion P
rotocol
| Apri
l 2007
antoi
ne.pi
chot@
alcate
l-luce
nt.fr