quality of service (qos) in ip networks - vsb.czwh.cs.vsb.cz/sps/images/6/6e/qos.pdf• cisco...

© 2005 Petr Grygarek, VSB-TU Ostrava, Routed and Switched Networks 1

Quality of Service Quality of Service (QoS) in IP networks(QoS) in IP networks

Petr GrygPetr Grygáárekrek

2© 2005 Petr Grygarek, VSB-TU Ostrava, Routed and Switched Networks

Quality of Service (QoS)Quality of Service (QoS)

• QoS QoS is the ability of network to support is the ability of network to support application without limiting of it’s function or application without limiting of it’s function or performanceperformance

• ITU-T E.800: ITU-T E.800: Quality of Service is an overall Quality of Service is an overall result of service performance, which determines result of service performance, which determines level of service user satisfactionlevel of service user satisfaction


What influences QoS ?What influences QoS ?


What influences QoS ?What influences QoS ?

Every network component influences QoS:Every network component influences QoS:

• End stations (workstations, servers, …)End stations (workstations, servers, …)• Routers, switchesRouters, switches

• Firewalls, load-balancers, proxies, ...Firewalls, load-balancers, proxies, ...

• Links Links • Both LAN & WANBoth LAN & WAN• including links between routers and stub LANsincluding links between routers and stub LANs


Parameters which constitute QoSParameters which constitute QoS

• bandwidthbandwidth• delaydelay• delay variation (delay variation (jitterjitter))• packet losspacket loss


What contributes to delay ?What contributes to delay ?

• Fixed partFixed part• sserializaerialization delaytion delay• propagation delaypropagation delay

• Variable partVariable part• buffering (buffering (queueingqueueing) ) in routers and switchesin routers and switches


Reasons of delay variationReasons of delay variation

• Non-uniform handling of individual packets of Non-uniform handling of individual packets of the same flow in individual switching devicesthe same flow in individual switching devices

• variable delay caused by unequal times packets wait variable delay caused by unequal times packets wait in queuesin queues

• Load balancing over links with different QoS Load balancing over links with different QoS parametersparameters


Reasons of packet lossReasons of packet loss

• ((physical transmission errorsphysical transmission errors – – receiver receiver discarddiscardss) ) • Overload of switching element’s CPU/switching Overload of switching element’s CPU/switching

fabricfabric• (input drops) (input drops)

• Output queue fullOutput queue full• output dropsoutput drops

• Input queue fullInput queue full• Input dropsInput drops


To keep in mindTo keep in mind

• High available bandwidth comparing to low High available bandwidth comparing to low average utilization does not imply perfect QoSaverage utilization does not imply perfect QoS• bursts may overload input buffersbursts may overload input buffers

• Application may perform considerably worse Application may perform considerably worse over WAN than over LANover WAN than over LAN• much longer delay (even 40 km matters)much longer delay (even 40 km matters)• hard-coded stop-and-wait communication schemehard-coded stop-and-wait communication scheme

• TCP adaptive behaviourTCP adaptive behaviour


QoS Implementation ModelsQoS Implementation Models


QoS implementation modelsQoS implementation models

• Best EffortBest Effort• Integrated Services - Hard QoSIntegrated Services - Hard QoS• Differentiated Services - Soft QoSDifferentiated Services - Soft QoS


Best EffortBest Effort Service Service

• Original Original InternetInternet service service• Makes best effort to transfer packet, but Makes best effort to transfer packet, but

provides no guaranteesprovides no guarantees• In case of congestion, any packet may be In case of congestion, any packet may be

droppeddropped• No priorizationNo priorization


Integrated Services - Hard QoS Integrated Services - Hard QoS (intserv)(intserv)

• ExplicitExplicit reservation of network resources for individual flows reservation of network resources for individual flows before communication startsbefore communication starts• Link capacity, queue memory, CPU of switching elementsLink capacity, queue memory, CPU of switching elements

• Poor scalabilityPoor scalability• Many flows are passing through backbone devices Many flows are passing through backbone devices • Natural in connection-oriented networks (like ATM), problematic in Natural in connection-oriented networks (like ATM), problematic in

packet-switched networkspacket-switched networks• Resource Reservation Protocol (RSVP) Resource Reservation Protocol (RSVP)

• reserves resources for individual flows along it’s pathreserves resources for individual flows along it’s path• Used inside backbone infrastructure today for QoS-aware Used inside backbone infrastructure today for QoS-aware

rerouting techniquererouting technique• RSVP-TE over MPLS in service-provide networks etc.RSVP-TE over MPLS in service-provide networks etc.


IntServ: IntServ: Resource Reservation Resource Reservation Protocol (RSVP) Protocol (RSVP)

• Signalling between flow receiver and switching Signalling between flow receiver and switching devices along (reverse) flow pathdevices along (reverse) flow path• also among switching devicesalso among switching devices• signalling goes against dataflow directionsignalling goes against dataflow direction

• DDistinct aistinct andnd shared reservation shared reservation • Requires support of end-stations SW (in end-to-Requires support of end-stations SW (in end-to-

end solution)end solution)


Differentiated Services - Soft QoS Differentiated Services - Soft QoS (diffserv)(diffserv)

• Traffic classification and markingTraffic classification and marking• Defined Defined Per-Hop Behavior (PHB) Per-Hop Behavior (PHB) of switching of switching

devices (routers, switches) for every traffic class devices (routers, switches) for every traffic class • Limited number of classes – better scalabilityLimited number of classes – better scalability

• In practice approx. 5 classes are enough in enterprise In practice approx. 5 classes are enough in enterprise networksnetworks

• Only prefers some data classes, no delivery time Only prefers some data classes, no delivery time guaranteeguarantee• only relative preferencesonly relative preferences• network has to be designed with sufficient capacity an network has to be designed with sufficient capacity an

input admission controlinput admission control


Traffic ClassificationTraffic Classification• Classification on the network boundaryClassification on the network boundary

• or at QoS-domain/trust domain boundaryor at QoS-domain/trust domain boundary• Marking at layer Marking at layer 22:: 802.1pq 802.1pq

• 33--bit Class of Service (CoS) bit Class of Service (CoS) fieldfield• Marking at layer Marking at layer 3: 3: 8-bit TOS/DSCP field in IP header8-bit TOS/DSCP field in IP header

• Before: Before: Type of Service (ToS), 8 Type of Service (ToS), 8 IP precedence classesIP precedence classes• Now: Now: Differentiated Service Code PointDifferentiated Service Code Point (DSCP) (DSCP)

• Routers and L3 switches needs to map between L2Routers and L3 switches needs to map between L2 CoS CoS aand L3nd L3 ToS/DSC ToS/DSCPP• QoS has to be ensured end-to-endQoS has to be ensured end-to-end• Mostly preconfigured DSCP<->CoS mapsMostly preconfigured DSCP<->CoS maps


Standard Diffserv Per-hop BehaviorsStandard Diffserv Per-hop Behaviors• Expedited Forwarding Expedited Forwarding (EF)(EF)

• ““VirtuVirtual leased line”al leased line”• Little loss, small but variable delay, guaranteed bandwidthLittle loss, small but variable delay, guaranteed bandwidth

• Assured Forwarding Assured Forwarding (AF) (AF)• 4 4 classes specified byclasses specified by Class Selector Class Selector ( (1-4)1-4)• 3 drop preference3 drop preferences in every classs in every class

• (1=(1=low,low, 2= 2=medium,medium,3=3=high).high).

• Denoted as Denoted as AFxy, xAFxy, x==class selector, yclass selector, y==drop preferencedrop preference• x,y x,y coded in coded in DSCP. DSCP.

• Best Effort Best Effort


QoS Implementation MechanismsQoS Implementation Mechanisms


Mechanisms of QoS enforcementMechanisms of QoS enforcement

• Applied only when network congestion arisesApplied only when network congestion arises• If network is not congested, QoS mechanisms would If network is not congested, QoS mechanisms would

only cause unnecessary overhead and commonly are only cause unnecessary overhead and commonly are not appliednot applied

• The goal is to guarantee maximum bandwidth, The goal is to guarantee maximum bandwidth, minimum delay and minimum jitterminimum delay and minimum jitter


Input filtering, classification and Input filtering, classification and markingmarking


Traffic Policing Traffic Policing

• Limits input rate according to Commited Limits input rate according to Commited Information Rate (CIR)Information Rate (CIR)• Sometimes takes EIR into account alsoSometimes takes EIR into account also

• Mostly uses Token Bucket AlgorithmMostly uses Token Bucket Algorithm


Input Traffic ClassificationInput Traffic Classification• Classification based onClassification based on L L11-L7 -L7 informationinformation

• Incoming port, MAC address, Incoming port, MAC address, protoprotocolcol, IP, IP addresses, TCP/UDP portaddresses, TCP/UDP port, URL, MIME typ, URL, MIME typee, ...) , ...)

• Classified by sender or by network boundary Classified by sender or by network boundary devicedevice• Trust boundary – clients normally untrusted, servers Trust boundary – clients normally untrusted, servers

trustedtrusted• NBAR: Network-Based ApplicationNBAR: Network-Based Application RecognitionRecognition

• Backbone devices trust classification performed Backbone devices trust classification performed at the network boundaryat the network boundary


Traffic Flow TypesTraffic Flow Types

• Elastic - adaptive to current throughputElastic - adaptive to current throughput• Aggressive – don't react on congestion Aggressive – don't react on congestion

notifications (explicit/implicit)notifications (explicit/implicit)

• Sensitive – sensitive to delay/jitter/dropsSensitive – sensitive to delay/jitter/drops


Congestion ManagementCongestion Management


Congestion Management PrincipleCongestion Management Principle

• Traffic of various classes sorted into queuesTraffic of various classes sorted into queues• ExplicitExplicit (marking-based) assignment to queue (marking-based) assignment to queue• Default queueDefault queue

• Intelligent dropping of packets of aggressive or Intelligent dropping of packets of aggressive or elastic flowselastic flows

• Queues serviced by various algorithmsQueues serviced by various algorithms• Selection of next packet to send out of output lineSelection of next packet to send out of output line


Typical Queueing AlgorithmsTypical Queueing Algorithms• FIFO FIFO • Priority Queuing (PQ) Priority Queuing (PQ)

• absolute prioritiesabsolute priorities• Custom Queuing (CQ) Custom Queuing (CQ)

• proportional queue processingproportional queue processing• Weighted Fair Queuing (WFQWeighted Fair Queuing (WFQ))

• automatic traffic classification into flowsautomatic traffic classification into flows, fair handling of , fair handling of flowsflows

• Prefers shorter packets (interactive applications)Prefers shorter packets (interactive applications)• Class-Based Weighted Fair Queuing (CBWFQ)Class-Based Weighted Fair Queuing (CBWFQ)

• Manual classification into classes, FIFO in every class, WFQ Manual classification into classes, FIFO in every class, WFQ between classesbetween classes

• Low-Latency Queuing (LLQ) Low-Latency Queuing (LLQ) • WFQ + absolute priority queueWFQ + absolute priority queue


Priority QueueingPriority Queueing


Custom QueueingCustom Queueing


Weighted Fair QueueingWeighted Fair Queueing


Congestion AvoidanceCongestion Avoidance

• TCP Slow Start behaviourTCP Slow Start behaviour• When the queue fills-up, router starts to discards When the queue fills-up, router starts to discards

all packet (tail drop)all packet (tail drop)• Many TCP flows start to lose ACKs andMany TCP flows start to lose ACKs and perform perform

slow start at the same timeslow start at the same time• The result is global TCP flow synchronizationThe result is global TCP flow synchronization,,

causingcausing network load oscillation network load oscillation

• Random Early Discard technique is used to Random Early Discard technique is used to combat against itcombat against it


(Weighted) Random Early Discard(Weighted) Random Early Discard

• As queue becomes full, router will start to randomly As queue becomes full, router will start to randomly discard packets with increasing probabilitydiscard packets with increasing probability

• Slow start of some TCP connectionSlow start of some TCP connection((ss)) happens earlier happens earlier comparing with standardcomparing with standard tail-drop tail-drop• Avoids global TCP flow synchronizationAvoids global TCP flow synchronization

• WWeighted RED – packet discard probability influenced eighted RED – packet discard probability influenced also by packetalso by packet’s priority’s priority

• Note: Note: RED RED can regulate only TCP flowscan regulate only TCP flows• There is no feedback in UDP protocol itselfThere is no feedback in UDP protocol itself• Anyway, WRED is often used for input buffering limitation for Anyway, WRED is often used for input buffering limitation for

all kinds of trafficall kinds of traffic


WRED ConfigurationWRED ConfigurationFor every traffic class we defineFor every traffic class we define • Constant to calculate floating average queue length Constant to calculate floating average queue length

(based on current and last queue length)(based on current and last queue length)• Minimum queue length to begin random discardMinimum queue length to begin random discard• Constant to calculate drop probability based on Constant to calculate drop probability based on

current (averaged) queue lengthcurrent (averaged) queue length• Applies only when queue length reaches minimum length Applies only when queue length reaches minimum length

for random discardfor random discard• Maximum queue lengthMaximum queue length

• When a limit is reached, all packets of the class are When a limit is reached, all packets of the class are dropped with max drop probability until queue length dropped with max drop probability until queue length decreasesdecreases• Some kind of hysteresis neededSome kind of hysteresis needed


RED and non-TCP flowsRED and non-TCP flows

• RED is also useful for flows that lacks feedback RED is also useful for flows that lacks feedback mechanismmechanism

• Allows to limit queue tail drops of well-behaved Allows to limit queue tail drops of well-behaved flows caused by aggressive flowsflows caused by aggressive flows• input bufferinginput buffering


Traffic ShapingTraffic Shaping• Protects against sending of packets to ISP faster Protects against sending of packets to ISP faster

than ISP agrees to accept themthan ISP agrees to accept them• ISP applies traffic policing so sending more traffic ISP applies traffic policing so sending more traffic

only wastes bandwidthonly wastes bandwidth• The goal is to adapt traffic rate to the commited rateThe goal is to adapt traffic rate to the commited rate

• Important when physical line access rate is Important when physical line access rate is higher than CIR of some virtual circuithigher than CIR of some virtual circuit

• Mostly implemented using Leaky Bucket Mostly implemented using Leaky Bucket algorithm algorithm • not to be confused with Token Bucketnot to be confused with Token Bucket

Notice the difference between Traffic Shaping and Notice the difference between Traffic Shaping and Traffic PolicingTraffic Policing


Fragmentation and InterleavingFragmentation and Interleaving• Used on very slow (WAN) links with large serialization delayUsed on very slow (WAN) links with large serialization delay• Ensures reasonable delay and jitter of interactive applicationsEnsures reasonable delay and jitter of interactive applications

• packets of interactive application have to be delivered on timely packets of interactive application have to be delivered on timely mannermanner

• sending of long packet on the same line can take more time than is sending of long packet on the same line can take more time than is the maximum allowed interval between two interactive application’s the maximum allowed interval between two interactive application’s packetspackets

• Long packets are fragmented and fragments are interleaved Long packets are fragmented and fragments are interleaved with packets of interactive applicationswith packets of interactive applications


Policy routing and QoSPolicy routing and QoS

• Policy routing may be based onPolicy routing may be based on• TOS/DSCP valueTOS/DSCP value• Packet sizePacket size


QoS-based RoutingQoS-based Routing


QoS-based Routing:QoS-based Routing:Principles and ProblemsPrinciples and Problems

• Many routing protocols can propagate Many routing protocols can propagate currentcurrent QoS QoS parameters of individual linksparameters of individual links• Cisco (E)IGRP, IS-IS, BGPCisco (E)IGRP, IS-IS, BGP

• Applying DSCP marking based on BGP policies (community attribute) Applying DSCP marking based on BGP policies (community attribute) - QPPB- QPPB

• Problem with routing oscillationProblem with routing oscillation• and thus unwanted jitter of transferred packetsand thus unwanted jitter of transferred packets

• Some routing protocols (e.g. IS-IS) may calculate multiple Some routing protocols (e.g. IS-IS) may calculate multiple shortest paths optimized for various metric (multi-topology shortest paths optimized for various metric (multi-topology routing)routing)• Packet can be forwarded according to routing table selected based Packet can be forwarded according to routing table selected based

on it’s TOS field valueon it’s TOS field value• Consumes lot of resources (CPU, memory, bandwidth)Consumes lot of resources (CPU, memory, bandwidth)


QoS support in operating systemsQoS support in operating systems

• Windows – 2000 and higher – Diffserv (QoS API)Windows – 2000 and higher – Diffserv (QoS API)• RSVP canceledRSVP canceled

• Linux – wide variety of QoS mechanisms for QoS-Linux – wide variety of QoS mechanisms for QoS-aware packet handlingaware packet handling

• Specialized router and switch softwareSpecialized router and switch software• SW-based and HW-based implementationsSW-based and HW-based implementations• Note that not every SOHO vendor “really” supports Note that not every SOHO vendor “really” supports

QoSQoS• Always make sure whether configured QoS really works ;-)Always make sure whether configured QoS really works ;-)


QoS in networking technologiesQoS in networking technologies

QoS is built-in into some media access control QoS is built-in into some media access control methods themselvesmethods themselves

• Old technologies: Token Ring, 100VG-AnyLanOld technologies: Token Ring, 100VG-AnyLan• WiFi: IEEE 802.11eWiFi: IEEE 802.11e• WiMax (802.16)WiMax (802.16)


Implementation of QoS mechanismsImplementation of QoS mechanisms

1.1. Determine applications used in your network Determine applications used in your network and their requirementsand their requirements

• Monitor traffic for representative usage periodMonitor traffic for representative usage period2.2. Suggest and implement QoS mechanismsSuggest and implement QoS mechanisms3.3. Verify behavior of your networkVerify behavior of your network4.4. GOTO 1 ;-) GOTO 1 ;-)

Don't turn on “magic” automatic QoS features nor enable QoS Don't turn on “magic” automatic QoS features nor enable QoS globally without understanding and tuning QoS parametersglobally without understanding and tuning QoS parameters(default settings do not have to be reasonable) – you may impact (default settings do not have to be reasonable) – you may impact network performance in very negative way.network performance in very negative way.


QoS and IPv6QoS and IPv6• Adheres the same principlesAdheres the same principles

• Diffserv: DSCP and 802.1p fields treated the same wayDiffserv: DSCP and 802.1p fields treated the same way• RSVP: support for IPv6 neededRSVP: support for IPv6 needed

• Handling IPv6 prefixes, transport over IPv6Handling IPv6 prefixes, transport over IPv6

• Not implemented on all network device's platformsNot implemented on all network device's platforms• Classification based on IPv6 headers, L4 headers and deep Classification based on IPv6 headers, L4 headers and deep

packet inspectionpacket inspection

• Multiple options how to treat the same application on Multiple options how to treat the same application on IPv4 and IPv6 arisesIPv4 and IPv6 arises• Same application => same treatmentSame application => same treatment• De-priorize IPv6 as the majority of users use IPv4De-priorize IPv6 as the majority of users use IPv4


ReferencesReferences

• Congestion Avoidance Overview:Congestion Avoidance Overview:http://www.cisco.com/en/US/docs/ios/12_1/qhttp://www.cisco.com/en/US/docs/ios/12_1/qos/configuration/guide/qcdconav.htmlos/configuration/guide/qcdconav.html

• Congestion Management Overview:Congestion Management Overview:http://www.cisco.com/en/US/docs/ios/12_2/qhttp://www.cisco.com/en/US/docs/ios/12_2/qos/configuration/guide/qcfconmg_ps1835_TSDos/configuration/guide/qcfconmg_ps1835_TSD_Products_Configuration_Guide_Chapter.html_Products_Configuration_Guide_Chapter.html

quality of service (qos) in ip networks - vsb.czwh.cs.vsb.cz/sps/images/6/6e/qos.pdf• cisco...

Documents