Transcript Week10_qos - The University of Sydney

1
NETS3303
Networked Systems
Bjorn Landfeldt, The University of Sydney
2
Section 2
IP QoS
Bjorn Landfeldt, The University of Sydney
3
Outcomes
• Understanding components of IP QOS
– What they do
– Why they are used or proposed
• Have knowledge of some case study
technologies
• Understanding the relevance to MM
delivery
Bjorn Landfeldt, The University of Sydney
4
IP QoS
• Today’s Outline
–
–
–
–
–
–
What is QoS?
Types of traffic
IntServ
Signalling
Queuing and Scheduling
DiffServ
Bjorn Landfeldt, The University of Sydney
5
QoS ?
• Many definitions in literature
• My definition is:
– “A perceived level of quality of a service or
function in relation to the wanted or expected
level of quality”
• In this course, application behaviour
depending on network performance
Bjorn Landfeldt, The University of Sydney
6
IP QoS
• IP provides only Best Effort service:
–
–
–
–
No guarantees full stop
No guaranteed packet delivery
No guaranteed time
No guaranteed order
• IP is ignorant of packet content
• No “Flows” in IP
• Compare telephony network
Bjorn Landfeldt, The University of Sydney
7
QoS
Get lost
Lost speech:
“ing”, “is easy here honey”
Internet
Network parameters
•Packet loss
•Delay
•Jitter
Bjorn Landfeldt, The University of Sydney
Getting
lost is
easy here
honey.
8
Where did
he go?
QoS
Delay 1000 ms
Silence
Internet
Network parameters
•Packet loss
•Delay
•Jitter
Bjorn Landfeldt, The University of Sydney
Getting
lost is
easy here
honey.
9
QoS
What the
G ettinglos tis
easyhere h on ey
Internet
Network parameters
•Packet loss
•Delay
•Jitter
Bjorn Landfeldt, The University of Sydney
Delay 1000 ms
Getting
lost is
easy here
honey.
10
Types of Traffic
• Different applications generate different
types of traffic e.g.
–
–
–
–
Web pages (delay sensitive)
FTP (BW sensitive)
Streamed Media (BW sensitive)
Conversational Multimedia (delay and BW)
Bjorn Landfeldt, The University of Sydney
11
Building blocks
Routers
Edge Router
End host
Networ
k
Region
Routers
Edge Router
Networ
k
Region
•End – to – end signalling
•Routers: Queuing and Scheduling
•Edge Routers: Add admission control
•ALandfeldt,
defined
set of ofrules
Bjorn
The University
Sydneyor classes to request
End host
12
IntServ
• Provides a set of service classes per flow
– Guaranteed Service
• Hard guarantees (Conversational MM)
– Controlled Load
• Same behaviour as lightly loaded BE network
(adaptive MM etc.)
– Best Effort
• All other types of traffic
Bjorn Landfeldt, The University of Sydney
13
Is there a problem with
the per-flow
specification?
Bjorn Landfeldt, The University of Sydney
14
RSVP
• Create notion of flow:
– E2E Signalling
• IETF proposal
– Resource Reservation Protocol, RSVP
• Allows Applications to make reservations
Bjorn Landfeldt, The University of Sydney
15
RSVP
Router
Router
Router
OK
Flow
End Host
End Host
Can I get?
Can I get?
Can I get?
• App fills in Traffic specification (T-Spec)
• Each router: admission control
• If requirements met: make reservations
Bjorn Landfeldt, The University of Sydney
Can I get?
16
Why is signalling
receiver-based?
Bjorn Landfeldt, The University of Sydney
17
Admission Control
• Token Bucket (rate r, size b)
– Start with full bucket
– If enough tokens in bucket accept packets and
remove tokens
– Tokens keep filling with rate r
Bjorn Landfeldt, The University of Sydney
18
Queuing
• Traditional queuing: FIFO, one input one
output
• Need to separate traffic into classes
• Need to give different priority to different
classes
• Need to manage the different queues
Bjorn Landfeldt, The University of Sydney
19
QoS Router
• Standard QoS Router Components
–
–
–
–
–
Routing Policy (rules for classification)
Routing table (Where to send packets)
Input Lines (where packets com in, no queue)
Output queues (where packets wait to be sent)
Classifier (puts packets into queues acc. to
policy)
– Scheduler (decides which queue to empty)
Bjorn Landfeldt, The University of Sydney
20
Scheduling
• The scheduler assigns resources to tasks
• In a computer: divide CPU runtime to
processes
• In a router: divide available BW (output
queues) to packets
• Operates based on router policy
Bjorn Landfeldt, The University of Sydney
21
FCFS
• Work Conserving (if packet waiting, serve)
• Klienrock Conservation Law
 q
N
n 1
n
n
C
ρ = link utilisation
q = mean scheduler delay
C = a constant
If delay for one flow is lowered, the delay
for one or more other flows must increase
Bjorn Landfeldt, The University of Sydney
22
Non Work Conserving
• Scheduler can be idle even if packets
waiting
• Switches packets to
– The right destination
– At the right time
• Reduces jitter
• Makes traffic predictable
Bjorn Landfeldt, The University of Sydney
23
Scheduling Requirements
• Easy to implement
– Simple makes fast
– Few states allows HW
implementation
• High speed routing
• Fairness
– Local means global
– Protect from other
misbehaving flows
Bjorn Landfeldt, The University of Sydney
• Performance bounds
– Per flow bounds
• Deterministic guaranteed)
• Statistical
– Data rate, jitter, delay,
loss
• Admission Control
– Easy to implement
– efficient
24
Scheduling choices
• Priority levels
– How many
– Serve higher priority
queues first? (can
cause starvation)
• Work conserving?
– Delay/jitter control
required?
– Extra cost acceptable?
Bjorn Landfeldt, The University of Sydney
• Flow Aggregation
– Granularity?
• Per flow
• Per application
• Per terminal
• Per queue policy
– FCFS?
– Look inside each
packet and decide?
– Performance/overhead
25
Priority Queuing
• K queues
– 1≤k≤K
– Queue k+1 higher prio. than queue k
– Higher prio. served first
• Simple implementation
• Low processing overhead
• No fairness, low prio. queues can be starved
Bjorn Landfeldt, The University of Sydney
26
WFQ
• Round robin scheme
– Estimate time to send packet (finish number)
– Tag packet with finish number
– Serve packet with smallest finish number
• Regardless of queue
• Weights can be assigned to enable
prioritisation
• Implemented by manufacturers
Bjorn Landfeldt, The University of Sydney
27
CBQ
• Assigns fractions of
BW to class nodes
• Values minimum
• Nodes can borrow
unused BW
• Priority to flows
within a class
Bjorn Landfeldt, The University of Sydney
Root
X
RT
40 %
60 %
100 %
NRT
20 %
Y
40 %
28
Question:
Can we do QoS
management without
Queuing / Scheduling?
Bjorn Landfeldt, The University of Sydney
29
DiffServ
• IntServ per-flow scalability problem
• Solution: aggregate flows
– Treat classes not individual flows
– Thus, tables kept small
• IP TOS field becomes DSCP
– 6 bit identifier of class
Bjorn Landfeldt, The University of Sydney
30
DiffServ domain
Ingress
Router
PHB
Core
Router
Core
Router
PHB
Egress
Router
PHB
Dimensioned to meet
Ingress router admission
control
Bjorn Landfeldt, The University of Sydney
31
DiffServ PHB
• Expedited Forwarding, EF
– Highest priority
– WFQ suitable
• Assured Forwarding, AF
– Three drop probability classes
– Graceful behaviour
Bjorn Landfeldt, The University of Sydney
32
What if two DiffServ
domains have different
definitions of what a
DSCP translates to?
Bjorn Landfeldt, The University of Sydney
33
The QoS stair
QoS
Level
A
Bjorn Landfeldt, The University of Sydney
B
C
D
Domain
34
DiffServ:
Scales well
Statistical guarantee only
Bjorn Landfeldt, The University of Sydney
35
Summary
• IP, no flows, no traffic separation
• Different types of traffic, different needs
• QoS management:
–
–
–
–
Classification
Signalling
Admission control
Queuing/scheduling
• IntServ, DiffServ, RSVP
Bjorn Landfeldt, The University of Sydney
36
Reading
• Deeper understanding:
– RFCs 2205-2216, 2474-2475
Bjorn Landfeldt, The University of Sydney
37
Other areas
•
•
•
•
•
•
•
MPLS
VLANS
Working Intserv and Diffserv together
QoS in 3G, bearer services, signalling etc.
Service Level Agreements
Billing and business models
Fibre, channel allocations DWDM etc.
Bjorn Landfeldt, The University of Sydney