-
8/13/2019 QoSCloud
1/23
QoS for CloudComputing
PIREGRID THEMATIC DAYMay, 10thn 2011
University of Pau
Prof. Congduc Pham http://www.univ-pau.fr/~cpham
Universit de Pau, France
PIREGRIDEFA53/08
-
8/13/2019 QoSCloud
2/23
2
What is Quality ofService?
!Quality of service is the abilityto provide different priority todifferent applications, users,or data flows, or to guarantee
a certain level of performance
!QoS criteria are numerous andis highly dependant of theapplication... !Throughput, Delay, jitter, Loss rate
!... Or of the end-user!Image resolution, sound quality,
appropriate language, ...
-
8/13/2019 QoSCloud
3/23
3
Common ServiceSpecification
! Loss: probability that a flows data islost
! Delay: time it takes a packets flow toget from source to destination
! Delay jitter: maximum differencebetween the delays experienced by twopackets of the flow
! Bandwidth: maximum rate at which thesource can send data
! QoS spectrum:est Effort Guaranteed
-
8/13/2019 QoSCloud
4/23
4
QoS for Cloud
!The shared cloud assumption
Manyusers, with
various
profile and
different
needs!
-
8/13/2019 QoSCloud
5/23
5
Applications profile
Traditional
scientificcomputing
applications
Business/
financialapplications &
simulations
Interactive
applicationswith user
feedback
Environmental
and dataprocessing/
analysis
Urgent
computing forpublic safety,
disaster relief
Storage and
databasesoperations/
queries
Time scale
Bounded in time
SporadicContinous
Data injection
rate
Assuming that
received data must
be processed
Defines the
minimum datainjection rate that
is required by theapplication
-
8/13/2019 QoSCloud
6/23
6
Urgent disaster relief
Imote2
Multimedia board
Time scale
Continous
Data injection
rate
Min=25 fps
-
8/13/2019 QoSCloud
7/23
7
How to provide QoS?
!Many ways to provide QoS!Scheduling, admission control,
trafic control, dynamic resourceprovisioning,
How to take into account the various applications profiles?
How to protect users from misbehaving applications?
How to handle urgent demands?
Regulate (adapt & control) the data injection rate into thecomputing resources
-
8/13/2019 QoSCloud
8/23
8
QoS: general picture
Network
bandwidth isassumed to
NOT BEING thebottleneck
PU
buffers
Regulation element
Regulate the data
injection rate intothe Processing
Units (PU)
-
8/13/2019 QoSCloud
9/23
9
Traffic and ServiceCharacterization
! Definitions! Cloud Infrastructure (CI)! Processing Units (PU)
! To quantify a service one has two know!
Flows traffic arrival! Service provided by the CI, i.e., resources
reserved at PU
! Regulation will ne done by an envelopeprocess, borrowed & adapted from thenetwork community
! Ideas is to! Bound the data injection rate to! isolate users from each others and! to provide QoS enforcement at flow level.
-
8/13/2019 QoSCloud
10/23
10
Traffic Envelope(Arrival Curve)
! Maximum amount of service that a flow canrequest during an interval of time t
slope = max average rate
b(t) = Envelope
slope = peak rate
t
Burstiness Constraint
-
8/13/2019 QoSCloud
11/23
11
Traffic Envelope(traffic shaping)
time
bps Peak rate 1
Peak rate 2
Mean rate 1
Mean rate 2
Traffic are variable
by nature, musttake into account
burstinessconstraints
time
bits
Arrival curve
time
bps
Use an envelope
process to boundthe data injection
rate while allowingfor variable, bursty
traffic
-
8/13/2019 QoSCloud
12/23
12
Ex: Token Bucket (1)
! Characterized by three parameters (b, R, C)! b token depth! R average arrival rate! C maximum data injection rate
! A bit is transmitted only when there is anavailable token
!When a bit is transmitted exactly one token isconsumed
R tokens per second
b tokens
!C bps
regulatortime
bits
b*C/(C-R)
slope C
slope R
-
8/13/2019 QoSCloud
13/23
13
Token Bucket (2)
-
8/13/2019 QoSCloud
14/23
14
Token Bucket (3)
-
8/13/2019 QoSCloud
15/23
15
Arrival curve
-
8/13/2019 QoSCloud
16/23
16
Per-hop Reservation withToken Bucket
! Given b,r,R and per-hop delay d ! Allocate bandwidth raand buffer
space Basuch that to guarantee d
bits
b
slope rArrival curve
d
Ba
slope ra
-
8/13/2019 QoSCloud
17/23
17
Service Model
! The QoS measures (delay,throughput,loss, cost) depend on offered traffic,and possibly other externalprocesses.
! A service model attempts tocharacterize the relationship betweenoffered traffic, delivered traffic, andpossibly other external processes.
external process
Network elementoffered traffic
delivered traffic
(connection oriented)
-
8/13/2019 QoSCloud
18/23
18
Arrival and DepartureProcess
Network ElementRin Rout
Rin(t) = arrival process= amount of data arriving up to time t
Rout
(t) = departure process
= amount of data departing up to time t
bits
t
delay
buffer
-
8/13/2019 QoSCloud
19/23
19
Delay and Buffer Bounds
t
S(t) = service curve
E(t) = Envelope
Maximum delay
Maximum buffer
bits
-
8/13/2019 QoSCloud
20/23
20
Token Bucket support inworkflows (1)
Support of superscalar pipeline models
Work done with R. Tolosana, J. Banares and O. Rana
-
8/13/2019 QoSCloud
21/23
21
Token Bucket support inworkflows (2)
TB QoS is introduced
seamlessly intoworkflow specifications
with the Renew tools
Work done with R. Tolosana, J. Banares and O. Rana
-
8/13/2019 QoSCloud
22/23
22
Conclusions
!Clouds will be shared cloudsdriven by economicalconstraints
!For some applications,availability of resources andisolation are of primeimportance (urgent computing)
!QoS for clouds is already anecessary and hot topic inresearch community
-
8/13/2019 QoSCloud
23/23
23
Perspectives
!Add more parameters to the TBmodel
!Excess burst size!Advanced mark vs. drop policy
!Dynamic configuration of TBparameters at each stage of theprocessing path
!Take into account data inflationbehaviors
!Generalized the usage ofenvelope processes,comparison,