estimation of queueing delay for packet scheduling ironet seminar 5 adaptive scheduling conventional...
TRANSCRIPT
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
17.2.2005 Ironet Seminar 2
Outline
Motivation
Contribution
Adaptive Scheduling
Delay Estimators
Simulation Results
Conclusions
Future Work
17.2.2005 Ironet Seminar 3
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
17.2.2005 Ironet Seminar 4
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
17.2.2005 Ironet Seminar 5
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
17.2.2005 Ironet Seminar 6
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
17.2.2005 Ironet Seminar 7
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
17.2.2005 Ironet Seminar 8
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 −+=
17.2.2005 Ironet Seminar 9
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
17.2.2005 Ironet Seminar 10
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
17.2.2005 Ironet Seminar 11
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
�==
17.2.2005 Ironet Seminar 12
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
17.2.2005 Ironet Seminar 13
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)
17.2.2005 Ironet Seminar 14
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
**_=
17.2.2005 Ironet Seminar 15
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
17.2.2005 Ironet Seminar 16
Simulation Scenarios
� Threedifferent traffic mixes wereused:• pure CBR
• a multiplex of Pareto ON-OFF sources
• traffic from ”real” applications
� Only someexamplesof the resultsareshown here
17.2.2005 Ironet Seminar 17
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
17.2.2005 Ironet Seminar 18
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
17.2.2005 Ironet Seminar 19
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)
InstantaneousEstimated
17.2.2005 Ironet Seminar 20
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)
InstantaneousEstimated
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
17.2.2005 Ironet Seminar 21
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)
InstantaneousEstimated
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)
InstantaneousEstimated
17.2.2005 Ironet Seminar 22
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)
InstantaneousEstimated
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)
InstantaneousEstimated
17.2.2005 Ironet Seminar 23
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)
InstantaneousEstimated
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)
InstantaneousEstimated
17.2.2005 Ironet Seminar 24
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)
InstantaneousEstimated
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)
InstantaneousEstimated
17.2.2005 Ironet Seminar 25
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
17.2.2005 Ironet Seminar 26
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