Bandwidth Management and Scheduling in MPLS DiffServ Networks

Bandwidth Management and Scheduling in MPLS DiffServ Networks

Ximing Zeng, Chung-Horng Lung, Changcheng Huang

Department of Systems and Computer Engineering

Carleton University

Ottawa, Canada

Anand SrinivasanEion

Ottawa, Canada

Outline

Diverse QoS Requirements DiffServ Forwarding Classes Cisco Solutions Solutions based on fair queueing Our Solution Performance study

QoS requirements

0%

Packet Loss

Command/ control

(eg Telnet,Interactive

games)

Conversationalvoice and video

Voice/videomessaging

Streamingaudio/video

Transactions(eg E-commerce,Web-browsing, E-

mail access)

Messaging,Downloads

(eg FTP,still image)

Fax

Background(eg Usenet)

5%

100 msec 1 sec 10 sec 100 sec

Zeroloss

Delay

QoS requirements for real time applications

Medium Application Degree of symmetry

Typical Data rates/Amount of Data

Key performance parameters and target values

End-to-end Oneway Delay

Delay Variationwithin a call

InformationLoss**

Audio Conversationalvoice

Two-way 4 - 64 kb/s <150 msecPreferred*

<400 msec limit*

< 1 msec < 3% Packet

Loss Ratio

Video Videophone Two-way 16 -384 kb/s < 150 msecpreferred

<400 msec limitLip-synch : < 100 msec

< 1%Packet

Loss Ratio

Data Telemetry- two-way control

Two-way <28.8 kb/s < 250 msec N.A Zero

Data Interactivegames

Two-way < 1 KB < 250 msec N.A Zero

Data Telnet Two-way(asymmetric)

< 1 KB < 250 msec N.A Zero

QoS requirements for interactive applications

Medium Application Degree of symmetry

Typical data rate/Amount of data

Key performance parameters and target values

One-wayDelay

(response time)

DelayVariation

Informationloss

Audio VoiceMessaging

Primarilyone-way

4-32 kb/s < 1 sec forplayback < 2 sec for

record

< 1 msec < 3%Packet

Loss Ratio

Data Web-browsing- HTML

Primarily one-way

~ 10 kB < 4 sec /page N.A Zero

Data Transactionservices – high

priority e.g. e-commerce,

ATM

Two-way < 10 kB < 4 sec N.A Zero

Data E-mail(server access)

PrimarilyOne-way

< 10 kB < 4 sec N.A Zero

QoS requirements for streaming applications

Medium Application Degree ofsymmetry

Data rate/Amount of data

Key performance parameters and target values

Start-upDelay

Transport delayVariation

Packet loss atsession layer

Audio Speech, mixedspeech and music,medium and high

quality music

Primarily one-way

5-128 kb/s < 10 sec < 1 msec < 1% Packet lossratio

Video Movie clips,surveillance, real-

time video

Primarily one-way

16 -384 kb/s < 10 sec < 1 msec < 1% Packet lossratio

Data Bulk datatransfer/retrieval,

layout andSynchronisation

information

Primarily one-way

10 kB – 10 MB < 10 sec N.A Zero

Data Still image Primarily one-way

< 100 kB < 10 sec N.A Zero

DiffServ Service Classes

Expedited Forwarding (EF) PHB (RFC-2598) Provides a low-loss, low-latency, low-jitter, and assured bandwidth

service. Real-time applications such as voice over IP (VoIP), video, and online trading programs require such a robust network-treatment.

Best Effort ServiceNo service guarantee except for a minimum bandwidth to prevent service starvation.

Assured Forwarding (AFxy) PHB (RFC-2597) Provides certain forwarding assurance by allocating certain bandwidth and buffer space. Applications with certain QoS requirements but not real-time can use AF service. For example: streaming video.

Cisco Solution

LLQ or MDRR

VoIP, Interactive Game…

Video Conferencing…

Video on demand …

E-commerce …

……

http,ftp, email…

High priority EF

Low priorityAF and BE CBWFQ

PQ

EF

AF1x

AF3x

AF4x

BE

AF2x

Total reservable bandwidth is about 75%. BE reservation fixed around 25%.EF traffic is constrained and should not exceed 33%; small queue and packet size.AFs reserve the rest bandwidth.

Cisco Solution

EF<33%BE = 25%

AF1AF2

AF3

AF4

EF is assigned a bandwidth less than 33% of the link speed and is constrained according to the assigned bandwidth.However burst of EF traffic still exists.

BE reserves a certain amount of bandwidth.

The rest of the bandwidth can be allocated to AF services.

AFs and BE may not always get their bandwidth as reserved !

Cisco Solution Advantage:EF packets are guaranteed smallest delay possible

by given them high priority. Tradeoff: AF packets may be delayed due to burst of EF

packets and cannot meet its desired delay bound!

delay

backlog

σ

ρ

r

t

bytes

delay

backlog

t

bytes

To minimize the impact EF brings to the AF classes:EF has small queue size (therefore, close to CBR)EF has small packet size (shorter waiting time for other packets)

Cisco Solution

What if the EF traffic is bursty?

If bandwidth other than the average rate of EF traffic is claimed allocatable, then when EF burst comes, the bandwidth to AF classes cannot be guaranteed. Bad QoS!

Average rate

Peak rateReserved rate

tEF traffic rate

Link rate

If bandwidth other than the peak rate of EF traffic is claimed allocatable, then AFQoS is guaranteed. Low bandwidth utilization!

A trade off has to be made! The actual bandwidth reserved to EF class should close to the peak rate to minimize the service impact.

Cisco Solution

BE=25%

AF1AF2

AF3

AF4

EF average

EF wasted

Under LLQ, to minimize the service impact to AF service Classes, EF bandwidth isOver-provisioned.

Other solutions

Assign each class certain bandwidthVoIP, Interactive Game…

Video Conferencing…

Video on demand …

E-commerce …

……

http,ftp, email…

WFQ/DWRR

EF

AF1x

AF3x

AF4x

BE

AF2x

Other solutions (WFQ)

Use weighted fair queueing to assign bandwidth to EF, AF and BE classes.

Advantage: Service to AF packets will not be affected by EF traffic, they always get their reserved bandwidth

Disadvantage: Over-provisioning is still needed to guarantee small delay to EF classes.

σ

ρ

delay

rIf EF traffic is bursty, to have a small delay, a large bandwidth reservation is needed, which causes the same problem of wasted bandwidth.

delay

r

Other solutions (DWRR)

Dynamically adjust the bandwidth to EF class according to the backlog of EF traffic. The unused bandwidth can be used by BE traffic.

Advantage: the EF still gets a relatively low delay. Practical problems: 1)How often should we adjust the bandwidth allocation? 2)If it is WFQ, how can we adjust the virtual finish time for all the

backlogged packets on line? 3)The bandwidth unused by EF can be used by BE, but is there

any guaranteed minimum bandwidth? Or how can we assign the unused bandwidth?

Our Solution

Proposed scheduler architectureVoIP, Interactive Game…

Video Conferencing…

Video on demand …

E-commerce …

……

http,ftp, email…

CBWFQ

EF

AF1x

AF3x

AF4x

BE

AF2x

PQ

High priority

Low priority

Our Solution

EF and BE share the bandwidth:

EF traffic rate

Average rate

Peak rateReserved rate

t

),max( EFres

BEEFavg rrrr

EFavgr

EFresr

Link rate

BE traffic rate

Our Solution

EF and BE share the bandwidth:

),max( EFres

BEEFavg rrrr

EF traffic ratet

Link rate

BE traffic rate Reserved ratePeak rate

Average rate EFavgr

EFresr

Our Solution

rBE = 25%

AF4

AF3AF2

AF1

The pie under WFQor Cisco LLQ

EFavgr

EFavgrEF

resr -AF4

AF3AF2

AF1

EF + BE = rBE +

Now free!

The pie under our solution.You can have an extra slice!

EFavgr

Our Solution

Advantages: EF is given no less (if not more) bandwidth than in WFQ.

Performance is guaranteed. AFs are guaranteed the same bandwidth, the same

performance can be expected. Bandwidth can be allocated to EF and AF users more efficiently!

Tradeoff: BE traffic may experience a longer delay due to EF bursts.

However, they are not delay sensitive and their average minimum throughput is still guaranteed.

Simulation result

Src 0: EF traffic: 7 on-off voice sources369 packets/sec in average.

Src 1: BE traffic: 800 packets/sce, Exponential

Src 2/3: AF traffic: 400 packets/sce,Exponential

Link speed: 2000 packets/sce.

Average load: 98.45%

Simulation result

EF delay under LLQ, DWRR and WPRR

LLQ provides the smallestDelay to EF class.

WPRR provides delay which isComparable tp LLQ

DWRR provides a much longerDelay.

Simulation result

AF delay under LLQ, DWRR and WPRR

The same delay bound is Guaranteed under both DWRRAnd WPRR.

Under LLQ, the AF delay is longerDue to the burstness of EF traffic.6% of the packets violate the delayBound.

Simulation result

BE throughput

Simulation result

BE throughput

Conclusion

We developed a new scheduler for DiffServ routers with the following advantage:

High bandwidth utilization Guaranteed QoS Guarantee small delay and loss for EF. Provide QoS guarantee to AF by reserving the bandwidth. Guarantee the minimum throughput of BE.

References [1] S.Blake, D.Black, M.Carlson, E.Davies, Z.Zhang, W.Weiss, “An Architecture for Differentiated Services.”

IETF RFC 2475. Dec 1998.

  [2] V. Jacobson, K. Nichols, K. Poduri, “An Expedited Forwarding PHB.” IETF RFC 2598. June 1999.

[3] J. Heinanen, F. Baker, W. Weiss, J.Wroclawski, “Assured Forwarding PHB Group.” IETF RFC 2597. June

1999.

  [4] J. Mao, W.M. Moh. B Wei, “PQWRR scheduling algorithm in supporting of DiffServ” 2001. ICC 2001. IEEE

International Conference on Communications,Volume: 3 , Pages:679 – 684.11-14 June 2001.

  [5] A.K.Parekh, R.G.Gallager, “A Generalized Processor Sharing Approach to Flow Control in Integrated

Service Networks: The single node case,” IEEE/ACM Transactions on Networking, Pages:344 - 357. June 1993.

  [6] H.Wang, C.Shen, K.G.Shin, “Adaptive-weighted packet scheduling for premium service” Communications,

2001. ICC 2001. IEEE International Conference on , Volume: 6 ,  Pages:1846 – 1850. 11-14 June 2001.

[7] F. Le Faucheur, L. Wu, B. Davie, S. Davari, P. Vaananen, R.Krishnan, P. Cheval, J. Heinanen, “Multi-Protocol

Label Switching (MPLS) Support of Differentiated Services.” IETF RFC 3270 May 2002.  

Thank You !