Design and Implementation of Measurement-Based Resource

Design and Implementation ofMeasurement-Based Resource AllocationSchemes Using the Realtime Traffic Flow

Measurement Architecture

Robert D. Callaway†, Michael Devetsikiotis†, and Chao Kan‡

†Department of Electrical and Computer Engineering ‡Alcatel Research and Innovation Center

North Carolina State University Alcatel USA, Inc.

Raleigh, NC 27695-7911 Plano, TX 45045

{rdcallaw,mdevets}@eos.ncsu.edu [email protected]

June 22, 2004

IEEE International Conference on Communications 2004: Paris, France

Presentation Outline

• Motivation and Background

• Effective Bandwidth Estimators

• Overview of Realtime Traffic Flow Measurement Architecture

• Modifications to Realtime Traffic Flow Measurement Architecture

• Emulation Setup and Tests

• Results and Conclusions

Callaway, Devetsikiotis, & Kan 1

Motivation and Background

• Goals of Self-Sizing Networks

? Optimize network utilization while ensuring QoS

? Ensure QoS of network traffic

? Adaptively change to network conditions while meeting the above criterion

• Benefits of Measurement-Based Resource Allocation

? Not dependent on a priori assumptions

? Able to track some transient behavior in traffic (non-abrupt changes)


Our Contribution

• Review effective bandwidth proposals in literature

• Implement effective bandwidth algorithms in IETF-standardized environment

• Verify and validate the implementation

• Monitor allocations and QoS of traffic to measure algorithm accuracy

• Demonstrate by emulation that algorithms are implementable in real networks


Effective Bandwidth Estimators

• A generic formula for effective bandwidth was proposed by Kelly as:

eb(s, t) =1

stlog E

[e

sX[0,t]]

• The s parameter in the general definition cannot be directly estimated from

measurements; therefore, the direct application of this formula in an online measurement

resource allocation scheme is not practical.

• Three algorithms were chosen for further analysis because of their computational

complexity, performance, and memory requirements.

? Gaussian Approximation

? Courcoubetis Approximation

? Norros Approximation


Gaussian Approximation

Guerin, et. al defined the Gaussian Approximation as:

CEB = µ + σ√−2 ln ε− ln 2π

where µ is the mean arrival rate of the traffic, σ is the standard deviation of the arriving

traffic, and ε is the QoS parameter (packet loss probability).

• Assumes a bufferless link

• Serves as an upper bound


Courcoubetis Approximation

Courcoubetis, et. al defined the following approximation for effective bandwidth:

CEB = µ +IDs

2B

where µ is the mean arrival rate of the traffic, ID is the index of dispersion, s is the

space parameter, and B is the buffer size of the queue.

• The index of dispersion is defined as:

ID = limn→∞

1

nE

( n∑i=1

Xi

)2

• The s parameter is calculated from an asymptotically exponential decrease assumption.

• This approximation does not address long range dependent traffic.


Norros Approximation

Norros defined the following approximation for effective bandwidth:

CEB = µ +[B

H−1κ (H)

√−2aµ ln ε

] 1H

where κ (H) = HH(1−H)1−H, µ is the mean arrival rate of the traffic, B is the buffer

size of the queue, H is the Hurst parameter of the traffic , a is the coefficient of variation

of the traffic, and ε is the QoS parameter (packet loss probability) of the traffic flow.

• The coefficient of variation is approximated by the index of dispersion; this

approximation is only valid when the arriving traffic is short range dependent.

• The Norros Approximation is the only formula we considered that uses the Hurst

parameter in its calculations; therefore, it is the only formula that takes self-similarity

into consideration.

• It is also the only formula that addresses long range dependent traffic.


Overview of Realtime Traffic Flow MeasurementArchitecture

The 3 main components within the RTFM architecture are the meter, reader, and

manager.

• The meter serves to collect statistics on network flows that pass through links that are

connected to it.

• The reader retrieves the statistics from the meter at a regular interval via SNMP.


Implementation within RTFM Architecture

• We are interested in the number of arriving bytes in a given time period (tslot) for a

particular traffic flow; RFC 2722 provides a byte counting statistic called toOctets.

• We utilize a sliding window system with a size of N slots in our online implementation.

Initialization ofEffective

BandwidthThread

Delay for tslotseconds

Input toOctetsfrom the last tslot

into slidingwindow system

RecomputeMean

RecomputeVariance

UsingCourcoubetis or

Norros ?

RecomputeIndex of

DispersionYes

time_to_realloc=0

RecomputeEffective

Bandwidth

Changeservice rateon queue

Yes

time_to_realloc=N time_to_realloc--

No

No

The mean, variance, and index of dispersion are recalculated after every tslot. Each

network flow (or class) is filtered into its own queue, so after N slots, the service rate of

the queue is dynamically changed to the measured effective bandwidth.


Emulation Setup

• We added the effective bandwidth algorithms into the meter component of the RTFM

architecture.

• We installed NeTraMeT onto several Linux PC’s in order to validate and verify the

integrity of our environment.

• Traffic used in the emulation tests was generated using the Sup-FRP method proposed

by Byu & Rowen.

Effective

Bandwidth

Incoming

Traffic Outgoing

Traffic

C

C

C

RTFM meter

RTFM meter reader / manager

SNMP

EB Algorithms

ingress 172.16.0.1

10/100 Mbps Switch

carolina 172.16.0.2

ncstate 172.16.0.3

wolfpack 172.16.0.25

core1 10.0.1.2

Logical Diagram Network Diagram


Emulation Cases

We present the results from three cases of our emulation tests:

• Case I: The performance of each algorithm is tested against the same traffic trace.

• Case II: The scalability of the implementation is tested when multiple flows are sent

simultaneously through the measurement architecture.

• Case III: The ability of the implementation to track abrupt transient behavior in the

traffic characteristics (mean arrival rate)


Emulation Results: Case I

0 100 200 300 400 500 600 700 800 900 10000

0.5

1

1.5

2

2.5

3x 105 Plot of Traffic Trace vs. Estimated Effective Bandwidths − Meter Implementation: 1 Stream

Time (sec)

Thro

ughp

ut (b

ytes

/sec

)

Actual TrafficGaussian MethodCourcoubetis MethodNorros Method

0 1 2 3 4 5 6 7 8 9

x 104

10−4

10−3

10−2Packet Loss Probability vs. Target PLP: 10−3 − Meter Implementation: 1 Stream

Packet Number

Pac

ket L

oss

Pro

babi

lity

Target PLPGaussian PLPCourcoubetis PLPNorros PLP

From these graphs, we can see that each of the EB algorithms can provide the requested

QoS while providing significant bandwidth savings over peak-rate allocation.


Emulation Results: Case II

0 100 200 300 400 500 600 700 800 900 10000

0.5

1

1.5

2

2.5

3x 105Plot of Traffic Trace vs. Estimated Effective Bandwidths − Gaussian Method − Meter Implementation: 3 Streams

Time (sec)

Thro

ughp

ut (b

ytes

/sec

)

Actual TrafficStream 1Stream 2Stream 3

0 1 2 3 4 5 6 7 8 9

x 104

10−4

10−3

10−2

10−1Packet Loss Probability vs. Target PLP: 10−3 − Gaussian Method − Meter Implementation: 3 Stream

Packet Number

Pac

ket L

oss

Pro

babi

lity

Target PLPStream 1 PLPStream 2 PLPStream 3 PLP

These graphs illustrate the robustness of the implementation to track multiple flows

simultaneously and still provide the QoS for each flow.


Emulation Results: Case III

0 100 200 300 400 500 600 700 800 900 10000

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5x 105 Plot of Traffic Trace vs. Estimated Effective Bandwidths − Meter Implementation: 1 Stream

Time (sec)

Thro

ughp

ut (b

ytes

/sec

)

Actual TrafficGaussian MethodCourcoubetis MethodNorros Method

0 0.5 1 1.5 2 2.5

x 105

10−4

10−3

10−2

10−1Packet Loss Probability vs. Target PLP: 10−3 − Meter Implementation: 1 Stream

Packet Number

Pac

ket L

oss

Pro

babi

lity

Target PLPGaussian PLPCourcoubetis PLPNorros PLP

These graphs show that two of the algorithms are unable to provide the requested QoS

when there is an abrupt increase in the mean arrival rate of the traffic.


Conclusions & Summary of Our Contribution

• Implemented EB algorithms in open-source implementation of RTFM environment

• Verified and validated the implementation

• Showed the robustness and scalability of the system

? The measurement time scale is relative to the traffic characteristics; therefore, a

static tslot value fails to accurately capture the characteristics of non-stationary

traffic.

? Additional work has shown that dynamically changing the length of tslot at the

completion of N window slots allows the system to accurately track abrupt changes

in the characteristics of the traffic.

? With a dynamic tslot, QoS constraints can be met even when dramatic changes in

the traffic characteristics are observed.

• Demonstrated by emulation that algorithms are feasible to be implemented in real

networks


Acknowledgements

• This research was partly supported by the Center for Advanced Computing and

Communication - North Carolina State University, as a Core Project. The authors

thank Fatih Hacıomeroglu for his assistance and suggestions.


Design and Implementation of Measurement-Based Resource

Documents