estimation of queueing delay for packet scheduling ironet seminar 5 adaptive scheduling conventional...

17.2.2005 Ironet Seminar 1

Estimation of Queueing Delay for Packet Scheduling

Johanna Antila and Marko Luoma

Networking Laboratory, Helsinki University of Technology

{jmantti3, mluoma}@netlab.hut.fi


Outline

Motivation

Contribution

Adaptive Scheduling

Delay Estimators

Simulation Results

Conclusions

Future Work


Motivation

Adaptivity hasbecome a key word in network provisioning• self-configurablenetworks

Scheduling is a crucial component in adaptiveprovisioning• adjust the class resources dynamically

- useof measurements


ContributionStarting point:• Service differentiation is based on DiffServ architecture

Westudy delay-based adaptiveprovisioning• Delay-bounded HPD (DBHPD scheduler)

We develop estimator algorithms for DBHPD’sdelay estimation problemBy simulations we evaluate• Filtering properties of the proposed estimators

- Several traffic mixes


AdaptiveSchedulingConventional scheduling approaches:• divide the link capacity statically between the service

classes, e.g. WFQ, DRR

Webase the scheduling decisions on measurements of short term and long term packetdelays

delay-bounded HPD (DBHPD) algorithm• provides absoluteand proportional delay differentiation


Notations

Let us use the following notations• i = differentiation parameter• di(m) = queuing delay of the m’ th packet in class i• wi(m) = normalized head waiting timeof class i when

m packets havedeparted• g = constant• i = filtering coefficient for class i• si = mean packet sizeof class i• qi = maximum queuesize for class i• C = link capacity


Delay-bounded HPD delay-bounded HPD scheduler• aims to provideproportional delay differentiation

and a delay bound.

• the algorithm first checks whether a packet in the best class is about to violate its deadline with the following equation:

safecurrin ttdt +<+ max


Delay-bounded HPD� If delay violation is not occurring, the algorithm

tries to maintain the delay ratios betweenconsecutive classes

� this is achieved by selecting for transmission a packet from a backlogged class j with maximumnormalized hybrid delay:

)/)()1(/)(max(arg iiii mwgmdgj δδ −+=


Delay-bounded HPD

� We havealready evaluated the basic version of the DBHPD algorithm• comparisons with static DRR and adaptive

DRR in a simplesetup

• comparison with static DRR in a network setup

� In recent work, the focushas been on queueing delay estimation for DBHPD


Delay Estimators� Estimation of the long term delay is a crucial part

of the DBHPD algorithm• the head-of-linepacket delay and thisestimated delay

together determine the time-scaleat which DBHPD operates

� Possibleestimators• SimpleSum

• EWMA variants

• Kalman filters and filtersbased on neural networks


Delay Estimators� SimpleSum estimator:

• sum of the queueing delays divided by the departedpackets

• disadvantages- impossible to implement

- ” infinite” history

)(

)()(

)(

1

mD

mdmd

i

mD

mi

i

i

�==


Delay Estimators� SimpleEWMA estimator

• update long term delay with exponential smoothing- eliminate the overflow problem

• traffic characteristics of the classesarevery different- useof separate filtering coefficient for each class

)(md)((m)d(m)d iiiii 11 −−+=

)ln(**

1)(

ii

iiqqN

q =γ


Delay Estimators� The gamma-function behaves as follows

(assuming four traffic classes):

100 200 300 400 500 600 700 800 900 10000

0.005

0.01

0.015

0.02

0.025

0.03

0.035ga

mm

a (q

)

Queue length q (packets)


Delay Estimators� EWMA estimator with restart (EWMA-r)

• EWMA filter is restarted as the queuebecomes activeafter an idleperiod when a certain threshold of packetshavearrived in the queue

• filter is restarted if estimator has been idle for a time

C

sqfactorabscycle iii

i

**_=


Delay Estimators� EWMA estimator based on proportional error of

the estimate (EWMA-pe)• also the EWMA filtering coefficient is adapted

- adaptation is based on the proportional error of the estimate

)1()*1()(*)( −−+= mdnmdnmd iiiii γγ

=)(

)(

md

mdfn

i

i


Simulation Scenarios

� Threedifferent traffic mixes wereused:• pure CBR

• a multiplex of Pareto ON-OFF sources

• traffic from ”real” applications

� Only someexamplesof the resultsareshown here


Simulation Scenarios� CBR loads (total load and the loads in different

traffic classes) used in traffic mix 1

0 20 40 60 80 100 120 140 1600.5

0.6

0.7

0.8

0.9

1

1.1

1.2

1.3

Time (s)

Tot

al lo

ad

0 20 40 60 80 100 120 140 1600

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Load

Time (s)

Class 0Class 1Class 2Class 3


Simulation Scenarios� Simple topology was used

• complex topology not required when the filteringpropertiesare tested

100 Mbps2 ms

100 Mbps2 ms

sources

sources

sources

sources

sinks

sinks

sinks

sinks

Class 0

Class1

Class2

class3

Class 0

Class1

Class2

class3

40 ms10 Mbps


Simulation Results� Simple Sum estimator with pure CBR-traffic

0 20 40 60 80 100 120 140 160 1800

0.02

0.04

0.06

0.08

0.1

0.12

Que

uein

g D

elay

(s)

Time (s)

InstantaneousEstimated

0 20 40 60 80 100 120 140 160 1800

0.1

0.2

0.3

0.4

0.5

Que

uein

g D

elay

(s)

Time (s)



Simulation Results� EWMA-r estimator with pure CBR-traffic

0 20 40 60 80 100 120 140 160 1800

0.005

0.01

0.015

0.02

0.025

Que

uein

g D

elay

(s)

Time (s)


0 20 40 60 80 100 120 140 160 1800

0.02

0.04

0.06

0.08

0.1

0.12

Que

uein

g D

elay

(s)

Time (s)



Simulation Results� EWMA-r estimator with Pareto ON OFF traffic

405 405.5 406 406.5 407 407.5 408 408.5 409 409.5 4100

0.1

0.2

0.3

0.4

0.5

0.6

Que

uein

g D

elay

(s)

Time (s)


400 405 410 415 420 425 430 435 4400

0.1

0.2

0.3

0.4

0.5

0.6

Que

uein

g D

elay

(s)

Time (s)



Simulation Results� EWMA-pe estimator with Pareto ON OFF traffic

405 405.5 406 406.5 407 407.5 408 408.5 409 409.5 4100

0.1

0.2

0.3

0.4

0.5

0.6

Que

uein

g D

elay

(s)

Time (s)


400 405 410 415 420 425 430 435 4400

0.1

0.2

0.3

0.4

0.5

0.6

Que

uein

g D

elay

(s)

Time (s)



Simulation Results� Simple Sum and EWMA-r estimator with FTP

traffic

870 880 890 900 910 920 9300

0.1

0.2

0.3

0.4

0.5

Que

uein

g D

elay

(s)

Time (s)


870 880 890 900 910 920 9300

0.1

0.2

0.3

0.4

0.5

Que

uein

g D

elay

(s)

Time (s)



Simulation Results� EWMA-pe estimator with HTTP and FTP traffic

870 880 890 900 910 920 9300

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

Que

uein

g D

elay

(s)

Time (s)


870 880 890 900 910 920 9300

0.1

0.2

0.3

0.4

0.5

Que

uein

g D

elay

(s)

Time (s)



Conclusions� Lessons learned from the simulations

• SimpleSum and SimpleEWMA estimatorslead to falsescheduling decisions

- not suitable for the delay estimation problem

• EWMA-r and EWMA-peseem to bepromising- good filtering properties- simple to implement

• Especially EWMA-peoperated well with verydifferent types of traffic mixes


FutureWork� Network-wideperformance analysis of the

DBHPD algorithm with the new estimators• investigation of e.g. packet loss patterns and possible

oscillatory behavior of TCP

• refinement of the estimator functions if required

� Implementation of the estimators in an existingprototypeversion of DBHPD• measurements

estimation of queueing delay for packet scheduling ironet seminar 5 adaptive scheduling conventional...

Documents