advanced traffic management (qos) concepts

8/14/2019 Advanced Traffic Management (QOS) Concepts

1/54

Copyright 1998, Cisco Systems, Inc. All rights reserved. Printed in USA.Presentation_ID.scr 1

13191056_05F9_c2 1999, Cisco Systems, Inc. 1 1999, Cisco Systems, Inc.3191056_05F9_c2

2 1999, Cisco Systems, Inc.

319

1056_05F9_c2

Advanced TrafficAdvanced TrafficManagement and QoSManagement and QoS

ConceptsConcepts

Session 319Session 319


2/54


33191056_05F9_c2 1999, Cisco Systems, Inc.

IntroductionIntroduction

Traffic Management

Applications and Transports

So what Are the Issues for

TCP

Voice on IP

Video (Broadcast and Teleconferencing)

4

319


Let's Talk about TrafficLet's Talk about Traffic

ManagementManagement

Why it is a concern

What the guiding principles are

What tools are available

What can be accomplished usingthose tools

What cannot be accomplished


3/54



Why Traffic ManagementWhy Traffic ManagementIs a ConcernIs a Concern

Needs of certain applications

Mail? Web? Transaction processing?

Opportunities with certain transports

6

319


Guiding Principles forGuiding Principles for

Traffic ManagementTraffic Management

We want to achieve

Predictability

Reliability

Availability

In a network that

Keeps intelligenceat the edges

Scales to necessarysizes andbandwidths

Minimizescomplexity

Uses cost-effectivetechnologies


4/54



What Tools Are Available forWhat Tools Are Available forTraffic ManagementTraffic Management

Traffic path control

Queue depth management

Queue rate management

Permission to use a link

8

319


How Well Will TrafficHow Well Will Traffic

Management Do?Management Do?

We know we can do this:

Management of latency

Management of bandwidth

What cannot be accomplished

Creation of bandwidth that otherwisewould not be there


5/54



Source: Gartner Group Study, March 1997

WAN ProtocolBreakdown

WAN ProtocolBreakdown

1994 1996 1998E 2000E 2002E

80%

70%

60%

50%

40%

30%

20%

10%

0%

IP SNA IPX

Others RFC 1490

IP

Primarily a WAN IP TalkPrimarily a WAN IP Talk

IP is the dominantinternet protocol

TCP is the dominantdata transport

95% of Internet trafficuses TCP

Voice is a growingmarket

But beware of hype

Heterogeneouslink layers


319

1056_05F9_c2

Making NetworksMaking NetworksPredictablePredictable

The GrailThe Grail


319

1056_05F9_c2


6/54



This Is what You NeedThis Is what You Needto Understand:to Understand:

TCP-based applications, voice,and video can be managed wellwith a little planning

12

319


Parekh and Gallaghers PaperParekh and Gallaghers Paper

INFOCOMM 93

One must have at most a predictablepredictableamount of traffic in the network

One must have predictablepredictable traffic

delay in each network element Given these, end-to-end delayend-to-end delay of a

host to host message is predictableis predictable


7/54



Definition of PredictableDefinition of Predictable

Does not mean

Fixed, Invariant, or Zero

Means that it has a

Mean value

Statistical distribution

Upper bound

14

319


Predictable Amount of TrafficPredictable Amount of Traffic

in the Networkin the Network

The source must pace trafficinitiation so that standing queuesare bounded

Queues form when arrival rate exceedsdeparture rate

When congestion (too many

messages in one queue) sets in:Sources must not increase their rate

Ideally, sources decrease their rate


8/54



Examples of Source PredictabilityExamples of Source Predictability

TCP will keep at most a certainamount of traffic in flight

We say it is elasticrate isproportional to latency

Voice will send only and exactly as

fast as the coding algorithm permits

We say it is inelastic

16

319


Predictable Packet Treatment inPredictable Packet Treatment in

Routers and SwitchesRouters and Switches

Transit latency must be within limitsacceptable to the application

Variation in transit latency must bewithin limits acceptable to theapplication

No stream may be locked out apartfrom administrative policy

Applicable policy must be observed


9/54



Examples of UnpredictabilityExamples of Unpredictability

Queues change rapidly enough thatthe distribution cannot be described

Discards happen frequently enoughthat there is effectively no upper

bound on delivery time

18

319


Examples of PredictabilityExamples of Predictability

Classes of queues get sufficientservice that ultimate arrival is timelyand normal

Timely is an application concept


10/54



Quality of Service Issues inQuality of Service Issues inTraffic ManagementTraffic Management

Predominantly TCP traffic

Some specific applications

Voice/video traffic


319

1056_05F9_c2

Managing TCP TrafficManaging TCP Traffic

Moving Mountains of DataMoving Mountains of DataWithout Incurring the WorldWithout Incurring the World

Wide WaitWide Wait


319

1056_05F9_c2


11/54



Transport Breakout TCP Applications

Backbone Traffic MixBackbone Traffic Mix

Source: MCI/NSF OC-3MON via http://www.nlanr.net, 1998

22

319


TCP Technology IssuesTCP Technology Issues

Single drops communicate fromnetwork to sending host

You need to slow down

Multiple drops in round trip trigger

time-outs

Something bad happened out here


12/54



N+1

N

N+2N+3

Behavior of a TCP SenderBehavior of a TCP Sender

Sends as much ascredit allows

Starts credit small

Avoid overloadingnetwork queues

Increases creditexponentially

To gauge network capability

24

319


AckN

+1

N+1

N

N+2N+3

AckN

+1

AckN

+1

Behavior of a TCP ReceiverBehavior of a TCP Receiver

When in receipt of nextmessage, schedulesan ACK

When in receipt ofsomething else,

acknowledges all itcan immediately


13/54



AckN

+1

N+1

N

N+2N+3

AckN

+1

AckN

+1

AckN+4

N+1

Sender Response to ACKSender Response to ACK

If ACK acknowledgessomething

Update credit and send

If not, presume it indicatesa lost packet

Send first unacknowledgedmessage right away

Halve current credit

Increase linearly to gaugenetwork throughput

26

319


AckN

+5

N+4

AckN

+1

N+1

N

N+2N+3

AckN

+1

AckN

+1

AckN

+4

N+1

N+4

WorldWorldWideWideWait!Wait!

Multiple Drops in TCPMultiple Drops in TCP

In the event of multiple dropswithin the same session:

Current TCPs wait for time-out

Selective acknowledge maywork around (but seeINFOCOM 98)

New Reno fast retransmitphase takes several RTTsto recover


14/54



Remember Parekh and GallagherRemember Parekh and Gallagher


One must have predictablepredictable trafficdelay in each network element

Given these, end-to-end delayend-to-end delay of ahost to host message is predictableis predictable

28

319


How Can We Make TCP in aHow Can We Make TCP in a

Network Act Predictably?Network Act Predictably?

Predictable amount of trafficin the network:

Well-written TCP implementationsmanage their rates to the availablebandwidth

Router needs toProvide predictable treatment of packets

Queue delay and drop characteristics


15/54



Fundamental FIFO QueueFundamental FIFO QueueManagement TechnologiesManagement Technologies

Tail drop

Network standard behavior

Causes session synchronization whenwaves of traffic experiencecorrelated drops

Random Early Detection (RED)Random drops used to desynchronizeTCP sessions and control rates

30

319


Session SynchronizationSession Synchronization

Sessionsynchronizationresults fromsynchronizedlosses

Tail drop fromwaves of trafficsynchronizeslosses


16/54



RED EnabledRED EnabledCourtesy of Sean Doran, Ebone

Effect ofEffect ofRandom Early DetectionRandom Early Detection

One day, below 100% throughput

Simple FIFO with tail drop

Starting 10:00 second day, 100%throughput

Random Early Detection enabled

32

319


RED EnabledRED EnabledCourtesy of Sean Doran, Ebone

Was that a Fluke?Was that a Fluke?

No, heres what happened that week

Session synchronization reducedthroughput until RED enabled


17/54



0

50

100

150

200

250

300

350

400

Elapsed Time

NsRTT

Mean RTT Min RTT Max RTT STD DEV

MeanMeanLatencyLatencyCorrelatesCorrelates

withwithMaximumMaximum

QueueQueueDepthDepth

FIFO Traffic TimingsFIFO Traffic Timings

34

319


0

50

100

150

200

250

300

350

400

Elapsed Time

MsRTT

Mean RTT Min RTT Max RTT STD DEV

MeanMeanLatencyLatencyCorrelatesCorrelates

withwith

MinimumMinimumDropDropThresholdThreshold

AdditionalAdditional

CapacityCapacityto Absorbto AbsorbBurstsBursts

RED Traffic TimingsRED Traffic Timings


18/54



ThereforeTCP QoS Definition:ThereforeTCP QoS Definition:

Normally at most one drop perround trip

Mean variation in latency bounded

by predictable network

36

319


TCP Flow StatisticsTCP Flow Statistics

>90% of sessions have ten packetseach way or less

Transaction mode (mail, small web page)

>80% of all TCP traffic results from


19/54



An Interesting CommonAn Interesting CommonFallacy about RED:Fallacy about RED:

RED means you will have more drops

Statement derives from observed statistics

RED means that you will have

Closer to 100% utilization of your line

Less average delay per packet

But queuing theory?As a line approaches 100% utilization, drops willincrease, even though served load increaseseven though served load increases

38

319


TCP Traffic Management IssuesTCP Traffic Management Issues

Applications

Often have site-specific policyassociated with them

Traffic often identifiable by port numbers

SitesGenerally identifiable by address prefixor interface traffic is received on


20/54



TCP Bandwidth PolicyTCP Bandwidth PolicyQuestions to AnswerQuestions to Answer

Particular site or application wants atleastleast a certain bandwidth

Particular site or application wants atmostmost a certain bandwidth

Particular site or application wants toaverageaverage about a certain bandwidth

40

319


ClassifierClassifier

Queues

InterfaceInterface

This Is Where ClassesThis Is Where Classes

Come inCome in

Classes can be for:

Voice

Importantapplication/site

Unimportant

application/siteAssuring at least a rate

Limiting to a rate


21/54



Managed Link

Left Right

ICU

UC Me

U Betcha

Some Class of Traffic Wants atSome Class of Traffic Wants atLeastLeast a Certain Bandwidtha Certain Bandwidth

Example:

Several organizations share cost of link

Distribute bandwidth proportional tofiscal responsibility

42

319


Class of Traffic Wants atClass of Traffic Wants at MostMost

a Certain Bandwidtha Certain Bandwidth

Traffic shaping

Similar queuing technology to class-based weighted fair queuing

Rate assigned to

Interface or sub-interfaceFrame Relay circuit

ATM virtual channel (in hardware)


22/54



T-1

64

KBPS

Examples of Rate ControlExamples of Rate Control

Intranet exposure

Limit rate of web surfingoutside the company

Frame Relay network

Access rate exceedsPVC ratelimit rate torate of PVC

44

319


Some Class of Traffic Wants toSome Class of Traffic Wants to

AverageAverage a Certain Bandwidtha Certain Bandwidth

Service provider or largeenterprise model

Designed for

Cost containment

Managed response toconflicting demands


23/54



Marking TCP Traffic at EdgeMarking TCP Traffic at Edge

A useful technique:

Mark traffic at a network edge withsimple classifier

This allows network to

Do the right thing without having to fullyclassify everywhere

Use more effective markings

46

319


Serving TCP Traffic with theServing TCP Traffic with the

Assured ServiceAssured Service

Presumes service level agreement

Flat rate for traffic meeting a rate/burst profile

Usage charging for traffic out of profile

Drop management (weighted RED)

All traffic subject to loss

Traffic out of profile much more subject to loss

Enhances ISP traffic engineering

(Good for service provider and consumer)


24/54



Assured Service

0%

10%

20%

30%

40%

50%

60%

70%

010

20

30

40

50

60

70

80

90

Time

Usage

Structure of Presumed ServiceStructure of Presumed ServiceLevel AgreementLevel Agreement

Up to raterate over intervalintervalis in profile

Traffic within profilegets some guarantees

Traffic out of profilehas no guarantees

Potentially dropped byWRED at bottleneck

Usage pricingUsage pricing of overage

48

319


Line Congested?Drop at Some Rate!

Best Effort Service in SimpleBest Effort Service in Simple

IP NetworksIP Networks


25/54



Line Congestedand Packet Out ofand Packet Out of

ProfileProfile? Drop atSome Rate!

Assured Service in SimpleAssured Service in SimpleIP NetworksIP Networks

50

319


Best effort Service in anBest effort Service in an

ATM-Based NetworkATM-Based Network


26/54



Assured Service in anAssured Service in anATM-Based NetworkATM-Based Network

52

319


So, for TCPSo, for TCP

Traffic can be contained to a rate in amanner consistent with good qualityof service

Traffic can be managed well with alittle foresight and planning


27/54


53 1999, Cisco Systems, Inc.3191056_05F9_c2

Convergence with VoiceConvergence with VoiceNetworksNetworks

Its about Internet Telephony!Its about Internet Telephony!


54

319


Again, the Premise:Again, the Premise:

TCP-based applications,voicevoice, and video can bemanaged well with a little

planning


28/54



0

20

40

60

80

100

1996 1998 2000

Multimedia

Dynamic WWW

Static WWW

FTP and Telnet

Email and News

Other

Changing Corporate NetworkChanging Corporate NetworkApplication PredominanceApplication Predominance

2%2%

7%7%

27%27%

17%17%

39%39%

8%8%

2%2%

7%7%

27%27%

17%17%

39%39%

8%8%

13%13%

28%28%

15%15%

17%17%

12%12%

14%14%

Numbers

in Percent

Source: The Yankee Group, 1996

56

319


Email

Informationsearch/access

Subscriptionservices/Push

Conferencing/multimedia

Video/imaging

250

200

150

100

50

1997 1998 1999 2000 2001

Traffic Projectionsfor Voice and Data

Rel. BitVolume

Circuit Switched Voice

Data(IP)

From 2000 on, 80% of ServiceFrom 2000 on, 80% of ServiceProvider Profits Will Be DerivedProvider Profits Will Be Derivedfrom IP-Based Services.from IP-Based Services.Source: CIMI Corp.Source: CIMI Corp.

Growth of IP TrafficGrowth of IP Traffic

Source: Multiple IXC Projections


29/54



IP orIP or

VoiceVoice

ATMATM

OpticalOptical

B-ISDNB-ISDNIP or VoiceIP or Voiceover ATMover ATM

IP overIP overSONET/SDHSONET/SDH

IP overIP overOpticalOptical

IPIP

SONET/SDHSONET/SDH

OpticalOptical

ATMATM

SONET/SDHSONET/SDH

IP orIP orVoiceVoice

OpticalOptical

VoiceVoice

IPIP

OpticalOptical

High End IP Transport AlternativesHigh End IP Transport Alternatives

Lower Cost, Complexity and OverheadLower Cost, Complexity and Overhead

Multiplexing, Protection and Management at Every LayerMultiplexing, Protection and Management at Every Layer

58

319


H.323 Voice/VideoH.323 Voice/Video

Voice

Constant bit rate when sending

Relatively small messages (44-170 bytes)

Video

Generally high variable bit rateControlled by codec efficiency on picture

Message size is generally the MTU


30/54



Delta Frames

KeyFrame

KeyFrame

Video: Traffic PatternVideo: Traffic Pattern

60

319


Video: Effect of DelayVideo: Effect of Delay

Delta Frames

KeyFrame

KeyFrame


31/54



Distribution of Deliveries in Time

TransmissionTime

PlaybackPoint

Unless itsToo Late

TypicalDelivery

Application Buffers Data

to Ensure Consistency

Preferred DeliveryPreferred DeliveryIntervalInterval

Video: Playback PointVideo: Playback Point

62

319


Synchronization ofSynchronization of

Voice and VideoVoice and Video

McGurk effect: voice can soundgarbled to human ear when notsynchronized with video

Therefore, we have tosynchronize these


32/54



QoS Definition for Voice:QoS Definition for Voice:

Low loss rate

Low absolute delay in two-waysituations

Broadcast voice doesnt have

this problem Low variation in delay

64

319


Key Issue for Voice QoS:Key Issue for Voice QoS:

Silent periods must not be randomlyinserted or removed so as to makeother sounds unintelligible

End to end delay must becomprehended by human listener


33/54



QoS Definition for Video:QoS Definition for Video:

Low loss rate

Low absolute delay in two-waysituations

Low variation in delay

66

319


Key Issues for Video:Key Issues for Video:

All packets that comprise a videoframe must arrive during the sameframe interval

OK if its the last millisecond of thatinterval

Audio and video must besynchronized when shown to user


34/54



How Can We Make Internet VoiceHow Can We Make Internet VoiceAct Predictably?Act Predictably?

Predictable amount of traffic inthe network

Predictable treatment of packets inrouters and switches

Planning to support these aspectsresults in a predictable network

68

319


Voice/Video TrafficVoice/Video Traffic

Management IssuesManagement Issues

The fundamental problems withVoice/video traffic are

It doesnt slow down in response todelay or loss

It requires minimal variation in delay


35/54



Predictable Amount ofPredictable Amount ofTraffic in the NetworkTraffic in the Network

The implication is that we have tocontrol used capacity

Capacity that individual calls consume

If you experience poor quality, use amore compact encoding or a lowerframe rate

Capacity that total call volume canconsume

If there isnt capacity, refusenew calls

70

319


Predictable Treatment of PacketsPredictable Treatment of Packets

in Routers and Switchesin Routers and Switches

We have to place voice in queuesthat give it high priority

Maintain tight delay budgets

Application of class-based WFQ


36/54



Planning for a Predictable NetworkPlanning for a Predictable Network

Enable CBWFQ on all relevant links

Configure voice queue with more bandwidththan traffic will need, or

For low bandwidth, priority queue [12.0(6)T]

Low speed links should use

Link Fragmentation or FRF.12

RTP compression for voice

Enable RSVP call negotiation

Refuse excess calls

72

319


FRF.12, and Link FragmentationFRF.12, and Link Fragmentation

and Interleavingand Interleaving

Premise:

Reducing voice packet size reducessession requirements on network

So compress out IP, UDP, and RTP

headers as much as possible

Limits jitter on lower bandwidth links


37/54



Resource ReservationResource Reservation

Current deployment

Current extensions

Extensions being developed

74

319


Current DeploymentCurrent Deployment

RSVP version 1

Call control for individual sessions

Deployed

Cisco 11.2

Microsoft Windows 98 (service pack)

Microsoft Windows NT 2000

Appropriate to edge networks


38/54



Current ExtensionsCurrent Extensions

Policy management via COPS

LAN management via subnetbandwidth manager

76

319


Policy Management Via COPSPolicy Management Via COPS

Local or centralpolicy server canauthorize decisions

Local policy:

Simple policies

Central policyserver:

Certificates,

Complex policies


39/54



LAN Management Via SubnetLAN Management Via SubnetBandwidth ManagerBandwidth Manager

Subnet bandwidthmanager is RSVPin a switch

Controls aggregatereservations ona LAN

78

319


Extensions Being DevelopedExtensions Being Developed

Rapid deployment of calls

Aggregate classification inedge networks

Aggregate classification and

admission in service providernetworks


40/54



Rapid Deployment of CallsRapid Deployment of Calls

Problem: need acknowledgedreservation installation

Solution: acknowledge it

80

319


PSTN

PSTN

Aggregate Classification inAggregate Classification in

Edge NetworksEdge Networks

Use differentiatedservices codepoints to identifytraffic

Rather thanspecific flows


41/54



Aggregate Classification inAggregate Classification inEdge NetworksEdge Networks

Reservation requestedby host in the usualway (RFC 2205)

Flow classificationand policing at firsthop router

Flow admission alongend to end path

Aggregate classification

and policing at

subsequent routers

82

319


Voice/video calls

Placed across aggregationdomain boundary

Aggregate ClassificationAggregate Classification and Admissionand Admission

Across Service Provider NetworksAcross Service Provider Networks

Why?

Otherwise, you dontknow that bandwidthexists on a path

Aggregate reservations

Placed from ingressto egress for DSCP used

Use expeditedforwarding service

Limited rate of change


42/54



Solving Voice/Video Issues UsingSolving Voice/Video Issues Usingthe Expedited Forwarding Servicethe Expedited Forwarding Service

Rate control

Application at source

Reservation in network

Jitter control

WFQs priority queue (low speed)Statistically empty queue (CB-WFQ)

84

319


The Implications forThe Implications for

Voice and VideoVoice and Video

We can control call volume

And therefore traffic volume

We can scalably prioritize traffic inthe system

And therefore deliver on latency issues So, voice and video can be managed

well with a little planning


43/54



Traffic Path ControlTraffic Path Control

What if IP Routing Isnt QuiteWhat if IP Routing Isnt Quite

Good Enough for Your Traffic?Good Enough for Your Traffic?


86

319


Traffic EngineeringTraffic Engineering

Historical approaches

Load sharing

Routing metrics

A new oneLabel switching


44/54



Load SharingLoad Sharing

Multipath routing

Equal and unequal cost

Multilink PPP

88

319


RoutingRouting

Administrative metrics

Designed to move traffic to statisticallylow volume links

Load sensitive metrics

Designed to move data away fromcongested links

Tendency towards oscillation


45/54


46/54



Principles of Label SwitchingPrinciples of Label Switching

Labeled paths:

Multiple enumerated point to pointrelationships between pairs of routers

Sets of pair-wise relationships create alabeled tunnel

Conceptually similar to ATM VCs orFrame Relay DLCs, but

Interface independentUsed to model network layer constructs

Variable length packets

92

319


Notice: TwoLabels on One

Interface,Distinguishing

Routes

Network Layer ConstructsNetwork Layer Constructs

Types of traffic streams

Destination routes

Source-destination routes

AS pairs

BGP community pairs

Tunnels can create

Any routing that meetsengineering needs


47/54



Using Labeled Tunnels to CreateUsing Labeled Tunnels to CreateVirtual Private NetworksVirtual Private Networks

Imagine edgenetwork with privateaddress space

Stretch labeledtunnels across thenetwork

Now, do it again

Disjoint networks

Same address space

Separate routing

94

319


MPLS Traffic EngineeringMPLS Traffic Engineering

Same technology can dragspecific routes around

Several less-used paths vs afew denser paths

Initially seen as

off-line engineering Can use either

LDP or RSVP to install routes


48/54



CoS in MPLS NetworksCoS in MPLS Networks

CClass oof SService

Roughly similar to diff-serv code point

Eight values, not sixty-four

Implements similar drop/delaymanagement within labeled tunnels

Therefore, MPLS networks havefundamental TCP QoS support

96

319


The Obvious HoleThe Obvious Hole

Wouldnt it be nice if engineeredlabeled tunnels could

Have specific bandwidths guaranteed?

Recover from network events quickly

and automatically using reasonable ifnot optimal routes?


49/54



MPLS Routing for ResourceMPLS Routing for ResourceReservationReservation

Use OSPF/IS-IS to distributebandwidth availability information

Edge router does SPF calculationwhen needed

RSVP used to install labeled tunnel

while checking for race events CoS field used to identify traffic for

queued rate support

98

319


Automated Reinstallation ofAutomated Reinstallation of

Labeled TunnelsLabeled Tunnels

RSVP tears downaffected tunnels

Edge devicesrecalculate routes

RSVP used to

re-install tunnels Bandwidth checks

result in retry


50/54


51/54


52/54



Assured Forwarding ServiceAssured Forwarding Service

Designed for TCP

Classes control rates for SLAs

Drop controls trace effects backto sources

Implement using

Committed access rate,Weighted Random Early Detection,

Class-based weighted fair queuing

104

319


Expedited Forwarding ServiceExpedited Forwarding Service

Appropriate to voice/video

Requires

Under-subscribed traffic classes

Reservation of bandwidthPolicing


53/54



Guiding PrinciplesGuiding Principlesfor Predictabilityfor Predictability


One must have predictablepredictable trafficdelay in each network element

Given these, end-to-end delayend-to-end delay of ahost to host message is predictableis predictable

106

319


In Your NetworkIn Your Network

TCP-based applications, voice,and videoand yourbandwidthcan be managedwell with a little planning


54/54


Please Complete YourPlease Complete YourEvaluation FormEvaluation Form

Session 319Session 319


advanced traffic management (qos) concepts

Documents