Overview

18-345: Introduction to
Telecommunication Networks • What is QoS?
Lectures 20: Quality of Service
• Queuing discipline and scheduling
Peter Steenkiste
• Traffic Enforcement
Spring 2015
www.cs.cmu.edu/~prs/nets-ece
• Integrated services
1 2

What is QoS? Why Should we Consider QoS?

• Current Internet supports best effort packet delivery
– Sufficient for most applications – “elastic” applications
• What is the basic objective of network
– But some applications require or can benefit from a “higher” design?
level of service
– Maximize total bandwidth? Minimize latency?
• “Higher” quality of service can mean that bounds are
provided for one or more performance parameters. – Maximize user satisfaction – the total utility
– Bandwidth: fast data transfers, video given to users
– Delay, jitter: telephony – Maximize profit?
– Packet loss, bit error rate: update services
• QoS can also mean that a user gets “better”
• What does utility vs. bandwidth look like?
treatment (than other users) – Shape depends on application
– But no guarantees are given – Must be non-decreasing function
4

1
 Performance versus
Utility Curve Shapes
Satisfaction
User U Elastic U Hard real-time
Satisfaction

No longer
Unacceptable Matters
BW BW

U Delay- or Rate-adaptive
Does equal
allocation of
Stay to the right and you
are fine for all curves bandwidth maximize
total utility?
Service Level BW
6

Utility curve – Elastic traffic Utility Curves – Inelastic traffic

U Delay-adaptive U Hard real-time
U Elastic

BW BW
Bandwidth

Does equal allocation of

Does equal allocation of
bandwidth maximize total utility?
bandwidth maximize total utility?
7 8

2
 Quality of Service
Inelastic Applications
versus Fairness
• Continuous media applications
– Lower and upper limit on acceptable • Traditional definition of fairness: treat all users
performance. equally.
– BW below which video and audio are not – E.g., share bandwidth on bottleneck link equally
intelligible • QoS: treat users differently.
– Internet telephones, teleconferencing with high – For example, some users get a bandwidth guarantee, while
delay (200 - 300ms) impair human interaction others have to use best effort service
– Sometimes called “tolerant real-time” since they • The two are not in conflict
can adapt to the performance of the network – All else being equal, users are treated equally
• Hard real-time applications – Unequal treatment is based on policies, price:
• Administrative policies: rank or position
– Require hard limits on performance
• Economics: extra payment for preferential treatment
– E.g. control applications
9

QoS Analogy:
How to Provide QoS?
Surface Mail
• Admission control limits number of flows
• The defaults if “first class mail”. – You cannot provide guarantees if there are too many flows
– Usually gets there within a few days sharing the same set of resources (bandwidth)
– Sufficient for most letters – For example, telephone networks - busy tone
– This implies that your request for service can be rejected
• Many “guaranteed” mail delivery services:
• Traffic enforcement limits how much traffic flows can
next day, 2-day delivery, next day am, …..
inject based on predefined limits.
– Provide faster and more predictable service at a – Make sure user respects the traffic contract
higher cost – Data outside of contract can be dropped (before entering the
– Providers differentiate their services: target network!) or can be sent at a lower priority
specific markets with specific requirements and • Scheduling support in the routers guarantee that
budgets users get their share of the bandwidth.
• Why don’t we do the same thing in networks? – Again based on pre-negotiated bounds

3
 QoS Framework What is a flow?
Traffic • Defines the granularity of QoS and fairness
Packet
Enforcement Scheduling – TCP flow
C H J – Traffic to or from a device, user, or network
B
I – Bigger aggregates for traffic engineering purposes
A
• Flows are defined using a packet classifier
– Classifier uses a set of fields in the packet header
Admission G to generate a flow ID
Control
– Example: (src IP, dest IP, src port, dest port,
E H
protocol)
F D – Or: (src prefix, dest prfix)
14

Admission Control Admission Control

• If U(bandwidth) is concave • If U is convex  inelastic
 elastic applications applications
– U(number of flows) is no longer
– Incremental utility is decreasing U monotonically increasing
Elastic
with increasing bandwidth – Need admission control to maximize Delay-adaptive
U
– Is always advantageous to have total utility
more flows with lower bandwidth • Admission control  deciding
• No need of admission control; when adding more people would
BW reduce overall utility
This is why the Internet works!
– E.g., bandwidth or latency
BW
• Not so for delay-adaptive and guarantees
real-time applications – Basically avoids overload
15 16

4
 Overview Queuing Disciplines
• Each router must implement some queuing
• What is QoS? discipline
• Queuing allocates both bandwidth and
• Queuing discipline and scheduling buffer space:
– Bandwidth: which packet to serve (transmit)
• Traffic Enforcement next
– Buffer space: which packet to drop next (when
• Integrated services required)
• Queuing also affects latency
17 18

Network Queuing Disciplines Alternative Drop Policies

• Avoid lockout and full queue problems
• First-in-first-out (FIFO) + drop-tail
• Random drop and drop front policies
– Simplest choice - used widely in the Internet
– FIFO means all packets treated equally – Drop random packet or packet that the head of the
– Drop-tail: new packets gets dropped when queue is queue is full and a new packet arrives
– Important distinction: – Solve the lock-out problem but not the full-queues
• FIFO: scheduling discipline problem
• Drop-tail: drop policy
• Random Early Discard (RED) and Explicit
• Alternative is to do Active Queue Management
Congestion Notification (ECN) slow down
– To improve congestion response
– Support fairness in presence of non-TCP flows
receivers before queues are full
– To give flows different types of service – QoS – See TCP lectures

19 20

5
Problems in Achieving fairness An Example
• In the Internet, fairness is only achieved • 1 UDP (10 Mbps) and 31 TCPs sharing a 10
if all flows play by the rules Mbps line
• In practice: most sources must use
TCP or be “TCP friendly” UDP (#1) - 10 Mbps UDP (#1)

– most sources are cooperative TCP (#2)

.
TCP (#2)
.
. .
– most sources implement . .
TCP (#32) TCP (#32)
homogeneous/compatible control law Bottleneck link
(10 Mbps)
• compatible means less aggressive than TCP
• What if sources do not play by the rule?
21
22

Throughput of UDP and TCP Fairness Goals

Flows With FIFO
10 • Allocate resources fairly
9
• Isolate ill-behaved users
8
– Router does not send explicit feedback to
Throughput (Mbps)

7
6 source
5 FIFO
– Still needs e2e congestion control
4
3 • Still achieve statistical multiplexing
2
– One flow can fill entire pipe if no contenders
1
0 – Work conserving  scheduler never idles link
1 3 5 7 9 11 13 15 17
Flow Number
19 21 23 25 27 29 31
if it has a packet
24
23

6
 What is Fairness? Max-min Fairness
• Allocate user with “small” demand what it
• At what granularity?
wants, evenly divide unused resources to
– Flows, connections, domains?
“big” users
• What if users have different RTTs/links/etc.
– Should it share a link fairly or be TCP fair? • Formally:
• Maximize fairness index? • Resources allocated in terms of increasing demand
– Fairness = (xi)2/n(xi2) 0<fairness<1 • No source gets resource share larger than its demand
• Sources with unsatisfied demands get equal share of
• Basically a tough question to answer! resource
• Good to separate the design of the mechanisms from
definition of a policy
– User = arbitrary granularity
• We will focus on max-min fairness – just an example
25 26

Implementing Max-min Bit-by-bit RR

Fairness
• Single flow: clock ticks when a bit is
• Generalized processor sharing transmitted. For packet i:
– Fluid fairness – Pi = length, Ai = arrival time, Si = begin transmit
time, Fi = finish transmit time
– Bitwise round robin among all queues
– Fi = Si+Pi = max (Fi-1, Ai) + Pi
• Why not simple round robin? • Multiple flows: clock ticks when a bit from all
– Variable packet length  can get more active flows is transmitted  round number
service by sending bigger packets – Can calculate Fi for each packet if number of flows
– Unfair instantaneous service rate is know at all times
• Why do we need to know flow count?  need to know A
• What if arrive just before/after packet departs?  This can be complicated
27 28

7
 Bit-by-bit RR Illustration Fair Queuing
• Mapping bit-by-bit schedule onto packet
• Not feasible to
transmission schedule
interleave bits on
real networks • Transmit packet with the lowest Fi at any
– FQ simulates bit-by- given time
bit RR – How do you compute Fi?

29 30

Example: Throughput of TCP

Bit-by-bit RR Example and UDP Flows With Fair
Flow 1 Flow 2 Output Queueing Router
0.45

0.4

0.35
F=10

Throughput (Mbps)
F=8 0.3

F=5 Flow 1 Flow 2 0.25

Output
(arriving) transmitting FQ
0.2

0.15

Cannot preempt packet F=10 0.1

currently being transmitted
0.05

F=2 0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31
Flow Number
31 32

8
 Fair Queuing Tradeoffs WFQ Illustration
• Complex computation
Flow 1
– Classification into flows may be hard W1
– Must keep queues sorted by finish times W2
Flow 2
– dR/dt changes whenever the flow count changes
I/P O/P
• Complex state – must keep queue per flow
• Hard in routers with many flows (e.g., backbone routers) W3
• Flow aggregation is a possibility (e.g. do fairness per domain)
• FQ can control congestion by monitoring flows
• Weighted fair queuing can give flows a different fraction
of the bandwidth - controlled by a weight Wi Flow n Wn
– Bandwidth of flow i is Wi / ∑ Wj
Variation: Weighted Fair Queuing (WFQ)
33
34

Traffic Enforcement:
Overview
Token Bucket Filter
Tokens enter bucket
at rate r
• What is QoS?
Bucket depth b:
• Queuing discipline and scheduling capacity of bucket

Operation:
• Traffic Enforcement – If bucket fills, tokens are discarded
– Sending a packet of size P uses P
tokens
• Integrated services – If bucket has P tokens, packet sent
at max rate, else must wait for
35 tokens to accumulate 36

9
 Token Bucket Operation Token Bucket Characteristics
• Can characterize flow using a token bucket:
Tokens Tokens Tokens
smallest parameters for which no packets will
be delayed
• On the long run, rate is limited to r
Overflow • On the short run, a burst of size b can be sent
• Amount of traffic entering at interval T is
bounded by:
Packet Packet
– Traffic = b + r*T
Enough tokens  Not enough tokens
packet goes through,  wait for tokens to • Information useful to admission algorithm
tokens removed accumulate
38
37

Token Bucket Traffic Enforcement: Example

• Parameters • r = 100 Kbps; b = 3 Kb; R = 500
– r – average rate, i.e., rate at which tokens fill the bucket (b)
– b – bucket depth
Kbps
(a)

3Kb
– R – maximum link capacity or peak rate (optional parameter) 2.2Kb
• A bit is transmitted only when there is an available token

r bps T = 0 : 1Kb packet arrives T = 2ms : packet transmitted

Maximum # of bits sent
bits b = 3Kb – 1Kb + 2ms*100Kbps = 2.2Kb
(c) (d) (e)
slope r
b bits b*R/(R-r) 3Kb
2.4Kb 0.6Kb
slope R
<= R bps
T = 4ms : 3Kb packet arrives T = 10ms : T = 16ms : packet
regulator time transmitted
39 40

10
 Token Bucket Specs - Example Overview

• What is QoS?
BW Flow B
2
Flow A: r = 1 MBps, B=1 byte
• Queuing discipline and scheduling
1 Flow A Flow B: r = 1 MBps, B=1MB

1 2 3 Time • Traffic Enforcement

• Integrated services
42
41

Integrated Services Guarantee Proven by Parekh

Traffic Classes
• Guaranteed service • Given:
– For hard real-time applications – Flow i shaped with token bucket and leaky bucket rate
– Fixed guarantee, network meets commitment if clients send control (depth b and rate r)
at agreed-upon rate – Network nodes do WFQ
– Admissions control limits number of flows
• Predicted service
– For delay-adaptive applications • Cumulative queuing delay Di suffered by flow i has
– Two components upper bound
• If conditions do not change, commit to current service – Di < b/r, (where r may be much larger than average rate)
• If conditions change, take steps to deliver consistent – Assumes that r < link speed at any router
performance (help apps minimize playback delay)
• Implicit assumption – network does not change much over time
– All sources limiting themselves to r will result in no network
queuing
• Datagram/best effort service • Basis for the IETF integrated services standard
43 44

11
 Lessons

• What type of applications are there? 

Elastic, adaptive real-time , and hard
real-time.
• Why do we need admission control 
to maximize utility
• How do token buckets + WFQ provide
QoS guarantees?

45

12