Upload
others
View
14
Download
0
Embed Size (px)
Citation preview
FP7 ICT-SOCRATES
Sensitivity Analysis of the Optimal Parameter Settings of
an LTE Packet Scheduler
I. Fernandez Diaz (TNO)R. Litjens (TNO)
J.L van den Berg (TNO)K. Spaey (IBBT)
D. Dimitrova (UT)
May 18, 2010, VTC ’10 Spring, Taipei, Taiwan
WWW.FP7-SOCRATES.EU
2/∞
CONTEXT
Driven by …– technological complexities– market-oriented perspectives
… there is an on-going trend towardsself-organisation of future mobile networks
– Self-configuration– ‘Plug and play’ installation
of new base stations and features– Self-healing
– Cell outage detection– Cell outage compensation: automatic
minimisation of coverage/capacity loss– Self-optimisation
– Power/tilt optimisation– Load balancing– Self-optimisation of packet
scheduling parameters– Automatic turning off/on sites– … triggered by
incidental events
continuousloop
WWW.FP7-SOCRATES.EU
3/∞
CONTEXT
Driven by …– technological complexities– market-oriented perspectives
… there is an on-going trend towardsself-organisation of future mobile networks
– Self-configuration– ‘Plug and play’ installation
of new base stations and features– Self-healing
– Cell outage detection– Cell outage compensation: automatic
minimisation of coverage/capacity loss– Self-optimisation
– Power/tilt optimisation– Load balancing– Self-optimisation of packet
scheduling parameters– Automatic turning off/on sites– … triggered by
incidental events
continuousloop
WWW.FP7-SOCRATES.EU
4/∞
OUTLINE
ContextObjectivePacket scheduling in LTE networksReference packet schedulerApproach sensitivity analysisNumerical resultsConcluding remarks
WWW.FP7-SOCRATES.EU
5/∞
OBJECTIVE
THIS PAPER:
Assessment of the sensitivity of optimal downlink LTE packet scheduling parameters with respect to a variety of traffic and environment aspects
FOLLOW-UP (IF SIGNIFICANT SENSITIVITY IS FOUND):
Develop self-optimisation algorithms to observe traffic/environment changes and adapt
scheduling parameters
WWW.FP7-SOCRATES.EU
6/∞
PACKET SCHEDULING IN LTE NETWORKS
Task of the packet scheduling algorithm– On a TTI timescale, assign cell’s radio resources to active sessions– Resource granularity in time domain 1 TTI = 1 ms– Resource granularity in frequency domain 1 ‘subchannel’ = 180 kHz
WWW.FP7-SOCRATES.EU
7/∞
REFERENCE PACKET SCHEDULING ALGORITHM
Supports RT and NRT sessions– With a tuneable degree of session-based differentiation– (As opposed to class-based differentiation)
Comprises three key principles– Proportional fairness → Tuneable channel-adaptivity: efficiency vs fairness– Packet urgency → RT packets are characterised by limited delay budgets– Work-conserving → Aim to utilise all resources
WWW.FP7-SOCRATES.EU
8/∞
REFERENCE PACKET SCHEDULING ALGORITHM
For each session i and subchannel c,calculate the priority level
with for each session i
Assign subchannels to sessionsbased on above prioritiesSelect uniform MCS per session
( ) ( )( )
( )( )
ξ
ρ ⎟⎟⎠
⎞⎜⎜⎝
⎛−
+⋅=tWT
tW1tR̂tR
tPii
i
i
c,iservicec,i
channeladaptivity
factor
packeturgencyfactor
( ) ( ) ( ) ( )1-tR1tR̂1tR̂ iii αα +−−=
• = potential bit rate at which session i can be served on subchannel c at TTI t
• = filtered average bit rate at which session i has been served up to TTI t
• = aggregate bit rate at which session i was served in TTI t-1
• = delay of HOL packet of session i experienced up to TTI t
• = maximum allowed packet delay for session i
• = minimum desired bit rate
• = parameter which sets the relative importance of the packet urgency factor
• = filtering parameter
( )tR̂ i
( )tR ci,
( )1-tRi
( )tWi
iT
ξ
α
serviceρ
WWW.FP7-SOCRATES.EU
9/∞
APPROACH SENSITIVITY ANALYSIS
Sensitivity analysis of the optimum packet scheduling parameter settings with respect to
– Service mix– Average file size (data service)– Coefficient of variation of the file size (data service)– Multipath fading environment– Variability of avg signal strengths between sessions
Reference scenario– Service mix file downloads only– Average file size (data service) 500 kbit– Coefficient of variation of the file size (data service) 1– Multipath fading environment PedestrianA, 3 km/h– Variability of avg signal strengths between sessions spatially uniform user
distribution, σshadowing = 9.4 dB
WWW.FP7-SOCRATES.EU
10/∞
NUMERICAL RESULTS
Reference scenario
Reference scenario
0
200
400
600
800
1000
1200
1400
0 200 400 600 800 1000
Cell load (kbit/s)
10%
Per
cent
ileof
thro
ughp
ut a
t cel
led
ge(k
bit/s
) Max SINRalpha=0.001alpha=0.01alpha=0.1RR
Reference scenario
0
500
1000
1500
2000
2500
3000
3500
4000
0 200 400 600 800 1000
Cell load (kbit/s)
Ave
rage
thro
ughp
ut (k
bit/s
)
Max SINRalpha=0.001alpha=0.01alpha=0.1RR
Reference scenario
0
200
400
600
800
1000
1200
1400
0 200 400 600 800 1000
Cell load (kbit/s)
10%
Per
cent
ileof
thro
ughp
ut (k
bit/s
)
Max SINRalpha=0.001alpha=0.01alpha=0.1RR
Reference scenario
0
200
400
600
800
1000
1200
1400
Alpha
Max
imum
sup
port
edce
lllo
ad(k
bit/s
)
Max SINRalpha=0.001alpha=0.01alpha=0.1RR
Reference scenario
0
200
400
600
800
1000
1200
1400
0 200 400 600 800 1000
Cell load (kbit/s)
10%
Per
cent
ileof
thro
ughp
ut a
t cel
led
ge(k
bit/s
) Max SINRalpha=0.001alpha=0.01alpha=0.1RR
Reference scenario
0
500
1000
1500
2000
Reference scenario
0
200
400
600
800
1000
1200
1400
0 200 400 600 800 1000
Cell load (kbit/s)
10%
Per
cent
ileof
thro
ughp
ut a
t cel
led
ge(k
bit/s
) Max SINRalpha=0.001alpha=0.01alpha=0.1RR
Reference scenario
0
500
1000
1500
2000
2500
3000
3500
4000
0 200 400 600 800 1000
Cell load (kbit/s)
Ave
rage
thro
ughp
ut (k
bit/s
)
Max SINRalpha=0.001alpha=0.01alpha=0.1RR
Reference scenario
0
200
400
600
800
1000
1200
1400
0 200 400 600 800 1000
Cell load (kbit/s)
10%
Per
cent
ileof
thro
ughp
ut (k
bit/s
)
Max SINRalpha=0.001alpha=0.01alpha=0.1RR
Reference scenario
0
200
400
600
800
1000
1200
1400
Alpha
Max
imum
sup
port
edce
lllo
ad(k
bit/s
)
Max SINRalpha=0.001alpha=0.01alpha=0.1RR
500 kb/starget
(FAIRNESS)
WWW.FP7-SOCRATES.EU
11/∞
NUMERICAL RESULTS
Reference scenario
Reference scenario
0
200
400
600
800
1000
1200
1400
0 200 400 600 800 1000
Cell load (kbit/s)
10%
Per
cent
ileof
thro
ughp
ut a
t cel
led
ge(k
bit/s
) Max SINRalpha=0.001alpha=0.01alpha=0.1RR
Reference scenario
0
500
1000
1500
2000
2500
3000
3500
4000
0 200 400 600 800 1000
Cell load (kbit/s)
Ave
rage
thro
ughp
ut (k
bit/s
)
Max SINRalpha=0.001alpha=0.01alpha=0.1RR
Reference scenario
0
200
400
600
800
1000
1200
1400
0 200 400 600 800 1000
Cell load (kbit/s)
10%
Per
cent
ileof
thro
ughp
ut (k
bit/s
)
Max SINRalpha=0.001alpha=0.01alpha=0.1RR
Reference scenario
0
200
400
600
800
1000
1200
1400
Alpha
Max
imum
sup
port
edce
lllo
ad(k
bit/s
)
Max SINRalpha=0.001alpha=0.01alpha=0.1RR
Reference scenario
0
200
400
600
800
1000
1200
1400
0 200 400 600 800 1000
Cell load (kbit/s)
10%
Per
cent
ileof
thro
ughp
ut a
t cel
led
ge(k
bit/s
) Max SINRalpha=0.001alpha=0.01alpha=0.1RR
Reference scenario
0
500
1000
1500
2000
Reference scenario
0
200
400
600
800
1000
1200
1400
0 200 400 600 800 1000
Cell load (kbit/s)
10%
Per
cent
ileof
thro
ughp
ut a
t cel
led
ge(k
bit/s
) Max SINRalpha=0.001alpha=0.01alpha=0.1RR
Reference scenario
0
500
1000
1500
2000
2500
3000
3500
4000
0 200 400 600 800 1000
Cell load (kbit/s)
Ave
rage
thro
ughp
ut (k
bit/s
)
Max SINRalpha=0.001alpha=0.01alpha=0.1RR
Reference scenario
0
200
400
600
800
1000
1200
1400
0 200 400 600 800 1000
Cell load (kbit/s)
10%
Per
cent
ileof
thro
ughp
ut (k
bit/s
)
Max SINRalpha=0.001alpha=0.01alpha=0.1RR
Reference scenario
0
200
400
600
800
1000
1200
1400
Alpha
Max
imum
sup
port
edce
lllo
ad(k
bit/s
)
Max SINRalpha=0.001alpha=0.01alpha=0.1RR
500 kb/sTarget
(FAIRNESS)
WWW.FP7-SOCRATES.EU
12/∞
NUMERICAL RESULTS
Sensitivity w.r.t. average file size
Larger files allow a lower α (higher spectrum efficiency; larger ‘fairness window’)
WWW.FP7-SOCRATES.EU
13/∞
NUMERICAL RESULTS
Sensitivity w.r.t. coefficient of variation of file size
A larger CoV means more small files for which a higher α is optimal (see previous slide)
WWW.FP7-SOCRATES.EU
14/∞
NUMERICAL RESULTS
Sensitivity w.r.t. multipath fading environment
WWW.FP7-SOCRATES.EU
15/∞
NUMERICAL RESULTS
Sensitivity w.r.t. variability of avg signal strengths between sessions
low = dense hot spot, no shadowingmedium = reference scenariohigh = uniform spatial user distribution, σshadowing = 14 dB
Under very low variability, fairness is established even with a pure channel-aware scheduler.The higher the variability, a high α is needed to lift up the cell edge sessions.
WWW.FP7-SOCRATES.EU
16/∞
CONCLUDING REMARKS
Summary sensitivity analysis in data only scenarios– α = 0.01 is the optimal parameter setting in most scenarios, except when
– the average file size is very large (α 0 optimal)– the coefficient of variation of file size is very high (α 1 optimal)– the differences in average signal strengths are large (α 1 optimal)
– Comparison with a self-optimised scheduling algorithm– Assume that a SON function perfectly recognises the situation and
always chooses the correspondingly optimal scheduling parameter– The gain of self-optimisation in the studied scenarios is not very large,
viz. 3.3% on average over the considered scenarios and 16.6% maximum– Considering that a SON implementation is imperfect and the scenarios with the
highest observed gains are not very likely, we conclude that the potential for self-optimisation is not very significant
Sensitivity analysis for video/data scenarios– Refer to the paper for the detailed results and discussion– Although the potential for self-optimisation is somewhat larger, it is still not
overly significant
WWW.FP7-SOCRATES.EU
17/∞
CONCLUDING REMARKS
Some open issues– Impact of integration of ICIC and PS on the sensitivity analysis– Possibility of other reference packet schedulers
– Less intelligent packet schedulers have more SON potential– Note that the presented reference packet scheduler already inherently
possesses some adaptivivity properties– Fairness aim, regardless of individual channel qualities– Automatic shift of resources to RT traffic if RT load increases– Automatic response to congestion– …
– Consideration of the LTE uplink– Integration of SON in the heart of the packet scheduler (at the ms timescale),
as an alternative to the addition of a slowly adaptive SON layer atop the PS layer
– Respond to instantaneous (n) rather than average (λ) load– Respond to instantaneous (nRT:nNRT) rather than average (λRT:λNRT) service mix– …
WWW.FP7-SOCRATES.EU
18/∞