Dynamic Bandwidth Management in QoS aware IP Networks Thesis Yasir Drabu Committee: Dr. Paul Farrell Dr. Javed Khan Dr. Hassan Peyravi (Chair)

Dynamic Bandwidth Management

in QoS aware IP Networks

Thesis Yasir DrabuCommittee:

Dr. Paul FarrellDr. Javed KhanDr. Hassan Peyravi (Chair)

Dynamic Bandwidth Management 2October 31, 2003

Presentation Overview INTRODUCTION

Application, Current Network Problems, QOS, Traffic Problem Definition and Research Goal.

BACKGROUND WORK Traffic Management - Active Queue Management (AQM),

Scheduling Admission Control – Classification, Current Implementation and

limitations. DYNAMIC BANDWIDTH MANAGEMENT (DBM)

Related Work Proposed Model – Analysis and Algorithms

SIMULATION Setup - Topology, traffic models, parameters and scenarios.

RESULTS CONCULSION and FUTURE WORK


Applications – Changing Needs Conventional Apps

Email, FTP, Telnet, etc Loss sensitive, High delay

tolerance, jitter insensitive. The IP Network was designed

for these. New Applications

WWW started a new trend. Video, VoIP, Interactive/

Streaming Video, e-commerce, etc.

Loss tolerant, delay sensitive, jitter sensitive to varying degrees.

The IP Network was not designed for these.

Bandwidth

U T I L I T Y

Bandwidth

U T I L I T Y

Traditional Applications New Real-time Applications


Current Networks - Issues Inherent Problems

Different traffic requirement, similar treatment. Signaling packets, Real-time packets, data packets,

individual packets within a flow, all treated same.

Ill behaved traffic hurts well behaved traffic. Unresponsive UDP flows dominate TCP flows.

Congestion Control limited to end hosts. TCP is predominant means of congestion control.

Changing/ Upgrading Difficult.

Bottom line – Need Quality of service


Quality of Service – QoS

User QoS Highly perceptional, hard to quantify.

Application QoS Applications change so do

requirements. Network QoS

Easy to quantify, well defined. All other QoS can be expressed in

these terms. Metrics well defined.

Availability Delay Delay Variation Throughput (Bandwidth) Packet Loss Rate.

Fig: Core Network QoS metrics


Network QoS – Approaches Best Effort Enhancements (RED, WRED,

ECN) Pros: Implemented over existing

infrastructure. Cons: No QoS guarantee

Integrated Services Pros: High level of QoS Cons: Not Scalable, consistency issues,

implementation a big problem. Differentiated Services

Pros: Scalable, incremental implementation Cons: QoS relative, unable to control flow

misbehavior within aggregate. Constrain Based Routing

Pros: Scalable, better network utilization Cons: Complexity (computational and

space), state information coherence, routing stability.

HybridHard Qos

Best Effort

Soft QoS

Cost

Co

mp

lex

ity

Courtesy Cisco


Today’s IP QoS technology

Technology Description Engineering Aspect

RSVP

DS Byte

Out of Band Signaling

In Band Signaling

Signaling

CAR (committed

access rate)

Classification and policing

(application, protocol , DS Byte)

Policing &

Classification

RED, WRED Weighted Random Early Detection

Service class enforcement

Congestion

Avoidance

WFQ, CBQ Weighted Fair Queuing

Class-Based Queuing

Queuing Policies

Congestion

Management and BW Allocation

MPLS MPLS Diffserv

IP+ATM QoS integration

Leverage

Layer2


Today’s Internet Traffic Conventional Traffic (Exponential, Smooth)

Exponential like Voice Easier to analyze Concrete parameters

arrival rate, queuing delay, etc. Can be simulated using Poisson's Distribution

Actual Traffic (Self Similar, Bursty) Heterogeneous mix of data, voice and multimedia application Difficult to characterize. In some cases not possible to

characterize. Effects of multiplexing

Makes the aggregate more self similar Makes the traffic more exponential (contradicting) Is there a factor, that can be used to decide the actual effect?

Can be simulated using Multiple Pareto distributions.


QoS – Bandwidth Allocation Problem Fixed Bandwidth allocated for QoS guarantee Admission Control

Requires a priori information about traffic characteristics. Traffic model does not accurately describe statistical behavior. User defined parameters may not accurately represent the

actual traffic. LDR traffic compounds the issue. Some work in Call Admission Control – MBACs, little to no work

in Aggregate Flow control. Weighted Scheduling – Weights associated based on

admission rates. Bottom Line – Bandwidth Allocation is inefficient.


Problem Definition and Goal

“To design and develop a dynamic bandwidth management model for efficient utilization of a shared link in a QoS aware IP

Network.” Design Guidelines:

1. Optimize bandwidth Utilization Better Delay Vs Utilization trade off.

2. Lower Loss rate Minimizes bad drop/mark decisions.

3. Fairness Non responsive UDP type flows can be controlled.

4. Coexistence with Best effort Prevent starvation of BE traffic

5. Scalable and Easily deployable The model has to be scalable on a huge network and be incrementally

deployed.


Background Work Three major components:

Active Queue Management (AQM) Random Early Detect Exponential Weighted Measured

Average (EWMA) Admission Control – Arrivals Scheduling – Service Allocation

AQM

SchedulingAdmission

Control

Network ProvisioningNetwork Provisioning

Performance ManagementPerformance Management

Admission ControlAdmission Control

SchedulerScheduler

QoS Re-routingBandwidth Reallocation

QoS Re-routingBandwidth Reallocation

Time ScaleShorter

Longer

Milliseconds

Seconds

Minutes

Hours

Days

Active Queue ManagementActive Queue Management

Microseconds


Typical QoS Aware Interface

Usually a QoS Enabled Edge Router has – Classifier, Admission controller, per class queue,

scheduler. Optional Metering unit.

Classifier S OUT

Admission Control

IN

Me

te

rin

g


Classification – Traffic Management

TrafficManagement

Proactive

Active Queue Management

Policing

Scheduling

HostCentric

RouterCentric

Reactive

Reservation Based

Non-Reservation Based

Static

Adaptive

FIFO

FQ

WFQWRR/CBQDWRR

PQ

ECN

Back Pressure

AWRR/ACBQ


Classification – Admission Control

Extensive research on Call Admission Control. MBAC for more

efficient bandwidth allocation.

Packet Admission Control, limited work as compared to CAC.

AdmissionControl

ParameterBased

Host centric

Call Admission Control

Packet Admission Control

Non-statistical

Statistical

MeasurementBased

Router centricCAR

AAC

Simple Sum

Measured SumHoeffding Bounds

Slow Start

Early Rejection

Direct Probing

Policy Based Admission Control


Dynamic Bandwidth Management To over come the inefficient bandwidth allocation we

use dynamic approach. Measure traffic conditions “online” and make admission

and allocation decisions. Three approaches – closed loop, open loop and hybrid.

Dynamic Bandwidth Management

Closed Loop

Hybrid

Open Loop

Queue Length

Loss

Delay

Rate

Closed loop State information used as

feedback

Open Loop Prediction based on Past

observations.


DBM – Proposed Architecture On QoS Enabled Router Interface:

Introduce a controller Feedback from Token Bucket - Starvation Rate tells us rate of packet arrival Feedback from Queue – Average Queue Size tells us rate of packet

departure. The Control has two components

Adaptive Admission Control (AAC) Adaptive Class Based Queuing (ACBQ)

S

r3 r2 r1 r0

B

0

B

1

B

2

B

3

Per Class Queue

w0

w1

w2

w3

Controller

Adjusted Token Rates

Adjusted Scheduler Weights


The Controller Components

Main Function: Monitors Arrival Rate Decides Packet

admission Design Parameters:

Bucket Size Token Rate

Decision Parameter: Bucket Starvation Rate

Main Function: Monitor Queue State Decides BW allocation

Design Parameters: Service Weight Queue Thresholds

Decision Parameter: Average Queue Size

Dynamic Bandwidth Controller

Adaptive Admission Control Adaptive Scheduler

)( iB

)( ir

)( iw

)( hiT

)( is )ˆ( i


Adaptive Admission Control - Analysis Bucket Starvation Rate

This indicates how many tokens a flow needs.

Fuller buckets mean lesser requirement and vice versa.

)()1(ˆ)1()(ˆ

:

tptptp

RatenStrarvatioAveragedThe

iii

iii Bbts

RatenStartvatio

/1)(

:

m

jjii mitststp

onDistributiTokenNormalize

1

1),(/)()(

:

)(ˆ tpr

RateTokenNew

ii

ClassiofSizeBucketB

ClassiofBuckettheinTokensb

ClassesofNumberm

CapacityLink

thi

thi

:

:

:

:


AAC: Algorithm


Adaptive Scheduler (ACBQ) - Analysis Each Queue has

TH – Upper Threshold

Average Queue Size

For Lesser Delay lower TH .

Average Queue Size indicates w.r.t to TH tells us how much service the queue class requires.

ihiii Tttq

RateStarvationServiceAbsolute

/)(ˆ)(

:

n

jjii tqtqtw

RateStravationServiceNormalized

1

)(/)()(

:

)()(

:

twt

RateServiceAverage

ii


ACBQ – Algorithm

The parameters:002.0tosettypicallyWq

001.0__ timeontransmissitypical


Simulation Setup


Simulation ScenariosNo. Scenario Name Traffic Load AC Scheduling

1 Base Line Exp 0.3-0.95 None Static CBQ

2 Static Allocation Exp 0.3-0.95 CAR Static CBQ

3 Partial Adaptive Exp 0.3-0.95 AAC Static CBQ

4 Fully Adaptive Exp 0.3-0.95 AAC Adaptive CBQ

5 Base Line Pareto 0.3-0.95 None Static CBQ

6 Static Allocation Pareto 0.3-0.95 Car Static CBQ

7 Partial Adaptive Pareto 0.3-0.95 AAC Static CBQ

8 Fully Adaptive Pareto 0.3-0.95 AAC Adaptive CBQ


Data Collections

Each simulation scenario was run 3 times with different seeds For a duration of 70 seconds

Data was collected between 10-70 seconds, assuming the simulator too 10 seconds to reach steady state.

The instantaneous values per simulation scenario were averaged over the duration of the simulation.

Then again the averages were averaged for the three runs.


Delay Performance – Exponential


Delay Performance – Pareto (LRD)


Jitter Performance - Exponential


Jitter Performance - Pareto


Loss Rate – Exp and LRD


Offered Load Vs Throughput

Average: 11% BW savingAt 80% Load – 30% BW saving

Average: 5% BW savingAt 80% Load – 14.5% BW saving


Simulation Results Summary

Higher Efficiency Lower Packet Drop - Prevented bad admission

decisions. Increased Throughput

Bursty traffic showed better gain.

Tradeoffs for Higher Throughput Increased Delay Increased Computation to O(N), Where N is the

number of QoS classes. N is always small.


Conclusion

CONTRIBUTION New approach to Bandwidth Management Our approach performs better than commercial

implementing in terms of bandwidth utilization. FUTURE WORK

Define an accurate relation between the QOS metrics and Control Parameter.

What is the best time scale of operation? How does it behave with RED and its Variants? End to End QOS is still not addressed.


Thank you.


View On Fundamental Limitations

Application Utility as a function of Network Performance This is undefined, and needs to be well defined.

QOS tries to provide better than BE How about worse than BE, Scavenger service, “nice”

in UNIX process resource sharing. Elevated services, non-elevated services (Internet2).

Deployment is a bigger challenge that most people think.


Opnet

Router under study

IP P

rocess Mode

ip_dispatch Process Model

2 Ethernet and 8 Slip Interface Node Model

Router N

ode Model

Child Process

IP Node Model:The router implements the complete networking stack. We are interested in the IP node. We modify this node to allow us to probe some of the statistics that we collect in relation to our proposed model. We also modify the underlying process models by implementing our adaptive algorithms.

IP Process Model:The IP process model has several child process models like ip_icmp, ip_vpn etc. Each child process model implements a specific feature of IP.

IP Child Process Model:The child process model, ip_output_iface implements all the Scheduling and AQM algorithms like Class Based Queuing, RED, WRED, etc.

Dynamic Bandwidth Management in QoS aware IP Networks Thesis Yasir Drabu Committee: Dr. Paul Farrell Dr. Javed Khan Dr. Hassan Peyravi (Chair)

Documents

qos relative

application qos applications

network qos easy

core network qos metrics

high level of qos cons

current network problems

future work slide

traffic models