Automatic Re-planning of Tracking Areas - FP7 SOCRATES · Automatic Re-planning of Tracking Areas Matías Toril Communications Engineering Dept., University of Málaga, Spain ...
Post on 22-Mar-2020
5 Views
Preview:
Transcript
1
Automatic Re-planning of Tracking Areas
Matías TorilCommunications Engineering Dept., University of Málaga, Spain
(mtoril@ic.uma.es)
24/10/2005
Karlsruhe, 22 Feb 2011
FP7 SOCRATES Final Workshop on Self-Organisation in Mobile Networks
(co-located with IWSOS 2011)
2
Outline
1 The tracking area re-planning problem
2 Graph-theoretic formulationp
3 Solution method
4 Performance analysis
5 Conclusions
3
Outline
LAIntroTA
1 The tracking area re-planning problem
Location area planning in legacy networks
LAIntroTA
State of research and technology FOR
2 Graph-theoretic formulation
3 Solution methodSOL
4 Performance analysis
5 C l iANA
5 Conclusions
CON
4
The tracking area planning problem
LAIntroTACellular network structuring
LAIntroTA
PCUPCU
BSCMSC/SGSN
FOR
BSC BSC
LA/RA LA/RA
BSC
PCUPCU
BSC
LA SOL
BSC BSCBSC
PCU PCU PCU PCUPCUPCUPCU
ANA
BTS
Site Site Site Site Site Site
BTS BTSBTSBTS BTSBTSBTSBTSBTSBTS BTSBTSBTS
LA
CON
5
i i ll l k
The tracking area planning problem
LAIntroTALocation management in current cellular networks
Purpose Know location/state of mobiles and direct mobile terminated calls
Algorithm Location update and paging (based on location/paging areas)
LAIntroTA
Algorithm Location update and paging (based on location/paging areas)
Problem Trade-off in location area size» Many small LAs ⇒ more LUs (i.e., DCCH capacity, load in databases)» Few large LAs ⇒ more paging requests (i e PCH capacity)
FOR
» Few large LAs ⇒ more paging requests (i.e., PCH capacity)
Solutions 1) Alternative LU/paging algorithms» LU (time/distance-based, groupal), paging (selective)
SOL
2) Optimise size/shape of LAs» Minimise total #LUs while keeping # paging messages per LA small
LU req.LA #1 LA #2 ANA
BSC MSC[DCCH] BTS
LU req.CN
PG req.[PCH]
Definition of LAsPaging algorithm
CON
VLR HLR
TMSI/LAC+CIBSSMS
[ ]
LA border
6
The tracking area planning problem
LAIntroTAState of research and technology
Current practice
LA plan with BSC (instead of BTS) resolution
LAIntroTA
LA plan with BSC (instead of BTS) resolution
1 LA ≅ 1 BSC ⇒ Many small LAs ⇒ many mobility LUs ⇒ large DCCH traf.
e.g., In GERAN, 50% of SDCCH attempts are LUs
FOR
12% of network capacity reserved for SDCCH
Changes in LA plan only as a result of BSC splitting event
C t i t th t BSC i th LA b l t th MSC
SOL
MSC/SGSN
• Constraint that BSCs in the same LA belong to the same MSC
• Changes in LA plan lead to temporary congestion of DCCH in affected cellsANA
BSC BSCBSC
LA/RA LA/RA
MSC/SGSNBSCBSC
CON
BTS
Site Site Site Site Site Site
PCU PCU PCU PCU
BTS BTSBTSBTS BTSBTSBTSBTSBTSBTS BTSBTSBTS
7
The tracking area planning problem
St t f h d t h l LAIntroTAState of research and technology
New drivers
Changes in vendor equipment LA borders can now cross MSC borders
LAIntroTA
Changes in vendor equipment LA borders can now cross MSC borders
New network algorithms Overlapping TAs [3GPP rel. 7], tracking area list [3GPP rel. 8]Interest on SON NGMN [SON use cases, O&M requirements], 3GPP [Rel. 8/9 LTE]
R l t d k
FOR
Related work
Graph partitioning Local refinement [Plehn 95], integer programming [Tcha 97],genetic algorithm [Gondim 96], simulated annealing [Demirkol 04],linear programming [Bejerano 06] set covering [Lo 04]
SOL
linear programming [Bejerano 06], set covering [Lo 04]
New network algorithms TA list [Modarres 09] , TA overlapping [Varsamopoulos 04]
Dynamic adaptation Trade-off signalling versus reconfiguration cost [Modarres 09]
Adjustment of TA overlapping [Varsamopoulos 04]ANA
Adjustment of TA overlapping [Varsamopoulos 04]
Main contributions
Re-formulation of TA planning as a classical graph partitioning problem
h d f l b d h kCON
Method to optimise TAs frequently based on statistics in the network management
• How often? Which changes? Potential impact on network signalling?
8
Graph-theoretic formulation
TA1 The tracking area re-planning problem in GERAN
2 Graph-theoretic formulation
TA
Naïve formulation
Adapted advanced formulation
ALGFOR
3 Proposed methodSOL
4 Performance analysis
5 ConclusionsANA
5 Conclusions
CON
9
Graph-theoretic formulation
TANaïve formulation
PCU 1 PCU 2 LA 1 LA 2
TA
Cell 1
1ω
5ω 2ω12γ15γ
ALGFOR
Cell 2
Cell 5
14γ
3ω45γ
4ω23γNetwork
model
SOL
Cell 3
Cell 4
34γmodel
( )
• Network area optimised:
Traditionally 1 MSC/VLRANA
(TAP) Minimise
subject toOptimisationd l
Traditionally 1 MSC/VLR
Currently 1 NMS
CONmodel
10
Graph-theoretic formulation
TAAdapted advanced formulation
TA
1) Paging cost in objective function
2) Paging cost term π paging constraint term ALGFOR
3) Time dependence
LU-to-HO ratio
Paging-to-LU cost ratio
'γ
SOL
(TAP) Minimise1 1( , ) ( ,..., ) ( , ) ( ,..., )
( )k k
ij i j ii j V V i j V V i
r cδ δ
γ ω ω ω∈ ∉
⎛ ⎞⎜ ⎟⎜ ⎟⎝ ⎠
+ + +∑ ∑ ∑( )( , )
(1 ) ( ) ij ij i j ij iii j
r S S ccγ ω ω ω− + ++ ∑∑ ( )( )( , )( , )
( ) 1 ( ) ij i j ij i j ii j ii j
r Sc cγ ω ω ω ω ω⎛ ⎞
− + − + +⎜ ⎟⎝ ⎠
+ ∑ ∑∑ijγ
( )( )( , )
( ) 1 ij i j iji j E
r Scγ ω ω∈
− + −∑ ( )( )( ) ( ) ( )
( , )( ) 1 s s s
ij i j iji j E
r Scγ ω ω∈
− + −∑ ( )( )( ) ( ) ( )
( , )[ ] [ ] [ ] [ ] [ ]( ) 1
ij
s s si j ij
t i j Et t t t tr Scγ ω ω
∈− + −∑ ∑ ANA
subject to( )
n
pki aw
i VBω
∈<∑ ( )[ ]
n
pki aw
i Vt Bω
∈<∑
CON
11
The assignment of PCUs in GERAN
TA1 The location area re-planning problem in GERAN
2 Graph-theoretic formulation
TA
3 Solution method
Proposed methodology
FOR
Proposed methodology
Classical graph partitioning algorithms
Graph resolutionMODSOL
4 Performance analysis ANA
5 Conclusions
CON
12
Solution method
TAGoals 1) Keep the number of TA re-plans as small as possible
2) Minimise impact of changing the TA plan
3) Minimise network signalling cost when re planning TAs
TA
3) Minimise network signalling cost when re-planning TAs
Proposed methodology
FOR
1) Define time granularity for measurements ⇒ hour, day, week
2) Collect network stats in several periods ⇒ HO, LU, CS traffic, total/peak pagingSOL
( ) ( ) ( )k3) Build network graphs ⇒
4) Compute graph correlation between periods ⇒
) d f l d d Cl l h ( k )
ANA
( ) ( ) ( ), ,ij
s s pki iγ ω ω
( , )u vρ
5) Identify correlated measurement periods ⇒ Clustering algorithm (e.g., k-means)
6) Compute TA plan for correlated periods ⇒ Classical graph partitioning algorithmin a row from past periods (e.g., ML refinement) CON
7) Select re-configuration instant ⇒ Low impact on control channels (e.g., night)
8) Estimate users affected by changes ⇒ (e.g., from traffic distribution)( )ciω
13
Solution method
TAGraph correlation
Definitions
TA
G(2)
FORG(1)
[ ] ( , )
[ ] [ ] ( )
ij
i
i j E
i V
γ γ
ω ω
= ∀ ∈
= ∀ ∈Ω
SOL
G(0)
| |
[ ] ( , ) ,
( ) ( ) ( ) ( )1( )E
s s s s
i j E i V
u u v vu v
γ ω
γ γ γ γρ
Ω = ∀ ∈ ∀ ∈
⎛ ⎞⎛ ⎞− −= ⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟∑
30
0.99
1
ANA
1 ( ) ( )
| |
( , )| |
( ) ( ) ( ) ( )1( , )| |
s u v
Vs s s s
u vE
u u v vu vV
γγ γ
ω
ρσ σ
ω γ ω ωρσ σ
=
= ⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠⎝ ⎠⎛ ⎞⎛ ⎞− −
= ⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠⎝ ⎠
∑
∑20
25
0.97
0.98
CON
1 ( ) ( )| |
1( , )
s u vV
u v
ω ωσ σ
ρ
=
Ω
⎜ ⎟⎜ ⎟⎝ ⎠⎝ ⎠
=| | | |
1 ( ) ( )| | | |
( ) ( ) ( ) ( )E Vs s s s
s u vE V
u u v vσ σ
+
= Ω Ω+
⎛ ⎞⎛ ⎞Ω −Ω Ω −Ω⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠⎝ ⎠
∑5
10
15
0.94
0.95
0.96
Graph correlation coefficientand clusters
5 10 15 20 25 30 0.93
14
Solution method
TAClassical graph partitioning algorithms
Refinement algorithms
G d l ith (GR)
Multi-level refinement
C i l ith
TA
Greedy algorithm (GR)
Kernighan-Lin algorithm (KL)
Fiduccia-Mattheyses algorithm (FM)
Coarsening algorithm
Initial partitioning
Uncoarsening algorithm
FOR
* Example: k=2, Baw=9 MODSOL
ANA
G(0)
G(1) G(1)
G(0)-3 -1 -1 -3
-2 -3 -3 -2
-1 +1 -3 -3
-2 -1 -3 -2
-1 +1 -3 -3
-2 -1 -3 -2
+1 -1 -3 -3
0 +1 -3 -2
+1 -1 -3 -3
0 +1 -3 -2
-1 -3 -3 -3
+2 -1 -5 -2
-1 -3 -3 -3
+2 -1 -3 -2
-3 -3 -3 -3
-2 -3 -3 -2
-3 -3 -3 -3
-2 -3 -3 -2
CON
Coarsening G(2)
G(m)
UncoarseningG(2)
-3 -1 -1 -3
-2 -3 -3 -2
-3 -1 -1 -3
-2 -3 -3 -2
-3 -3 +1 -1
-2 -3 -1 -2
-3 -3 +1 -1
-2 -3 -1 -2
-3 -3 -1 +1
-2 -3 +1 0
-3 -3 -1 +1
-2 -3 +1 0
-3 -3 -3 -1
-2 -3 -1 +2
-3 -3 -3 -1
-2 -3 -1 +2
-3 -3 -3 -3
-2 -3 -3 -2
G(m)
Initialpartitioning
Step 0
Fiduccia-Mattheyses
Step 1Step 2Step 3
Step 4
Step 5Step 6
Step 7
Step 8
15
Solution method
TAClassical graph partitioning algorithms
Adaptive multi-start ⇒ Multi-level evolutionary biasing (EB)
TA
Edg
e-cu
t
FOR
Greedy Graph Growing Partitioning Clustered Adaptive Multi-StartRandom Greedy Graph Growing Partitioning
E
Distance from global optimumMODSOL
y p g g
ut
p y p g g
600000
700000
t
Minimum valueOptimal value
ut ANA600000
700000
t
Minimum valueOptimal value
Edg
e-cu
Floyd-Warshal G d G h G i P titi i 300000
400000
500000
Edg
e-cu
Edg
e-cu
CON300000
400000
500000
Edg
e-cu
t
Nbr. of attempts
Adaptive Multi-StartNaive Multi-StartSingle attempt
Greedy Graph Growing Partitioning 300000
1 10000 500 1000
Nbr. of attempts
300000
1 10000 500 1000
16
Solution method
TAGraph resolution
BSC vs BTS
TA
BSC-level graph
FOR
Sorted Heavy
Edge Matching
graph
MODSOL
Site
Site-level graph
ANA
Matching
Cell-levelgraph CONg p
17
Performance analysis
TA1 The tracking area re-planning problem
2 Graph-theoretic formulation
TA
3 Solution methodFOR
4 Performance analysis
Analysis set-up
A l i lt
SOL
Analysis results
5 ConclusionsANA
CON
18
Performance analysis
TAAnalysis set-up
Goal Check time correlation of graphs in a real GERAN network
TA
Estimate benefit of different LA re-plan approaches in a real network
Check number of changes and population ratio affected by LA changes
Scenario 1 NMS (5498 BTSs, 54 BSCs, 50 LAs)
FOR
( )
Methodology 0) Read NMS data of 4 weeks ⇒ 2 weeks + 2 weeks one month later
1) Build BSC-level graphs ⇒ HO [γij], paging/CS/LU [ ]SOL
( ) ( ) ( ),,s pk ci i iω ω ω
2) Compute graph correlation ⇒
3) Define periods of high correlation ⇒ k-means clustering
4) Compute LA plans ⇒ ML evolutionary biasing, Baw=400000ANA
( , )u vρ
) Co pute p a s ⇒ e o ut o a y b as g, aw 00000
• Initial operator solution (k=50)• Overall, daily, periodic (perfect estimation, imperfect estimation,
imperfect estimation with local optimisation) CON
Criteria Total edge cut (⇒ Overall network signalling cost)
Total number/weight of changes (⇒ Nbr. of BSCs/users changing LA)( )ciω
19
Performance analysis
TAAnalysis set-up
Network area
TA
FOR
SOL
ANA
CON
Cell-level graph BSC-level graph
20
Performance analysis
TAAnalysis results
Graph correlation: BTS level
TA
Vertex weight Edge weight
25
1
FOR
25
1
200.9
0.95
SOL20
0.9
0.95
10
15
0 8
0.85
ANA10
15
0.85
5 10 15 20 25
5
0.75
0.8
CON5 10 15 20 25
5
0.8
5 10 15 20 25
| |
1 ( ) ( )
( ) ( ) ( ) ( )1( , )| |
Es s s s
s u v
u u v vu vEγ
γ γ
γ γ γ γρσ σ=
⎛ ⎞⎛ ⎞− −= ⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠∑
| |
1 ( ) ( )
( ) ( ) ( ) ( )1( , )| |
Vs s s s
s u v
u u v vu vVω
ω ω
ω γ ω ωρσ σ=
⎛ ⎞⎛ ⎞− −= ⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠∑
5 10 15 20 25
21
Performance analysis
TAAnalysis results
Graph correlation: BSC level
TA
Vertex weight Edge weight FOR
25
0.95
1
25
0.995
1
SOL20
0.85
0.9 20
0.98
0.985
0.99
ANA10
15
0 75
0.8
0.85
10
15
0.965
0.97
0.975
CON5 10 15 20 25
5
0.7
0.75
5 10 15 20 25
5
0.95
0.955
0.96
5 10 15 20 25
| |
1 ( ) ( )
( ) ( ) ( ) ( )1( , )| |
Es s s s
s u v
u u v vu vEγ
γ γ
γ γ γ γρσ σ=
⎛ ⎞⎛ ⎞− −= ⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠∑
| |
1 ( ) ( )
( ) ( ) ( ) ( )1( , )| |
Vs s s s
s u v
u u v vu vVω
ω ω
ω γ ω ωρσ σ=
⎛ ⎞⎛ ⎞− −= ⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠∑
5 10 15 20 25
22
A l i lt
Performance analysis
TAAnalysis results
Automatic clustering of measurement periods
TA
2K
∑ ∑ 2K
∑ ∑K-means ⇒ FOR
K 1 K=2
1
2arg min ( , )ss
sC C
d μ= Ω∈
Ω∑ ∑1
2arg min (1 ( , ))ss
sC C
ρ μ= Ω∈
− Ω∑ ∑
SOL
K=1
K=3
K=2
K=4
ANAK=5 K=6
CONK=7 K=8
5 10 15 20 25 5 10 15 20 25
Separate plan for business days and weekends
23
Performance analysis
TAAnalysis results
Comparison of methods ⇒ operator vs overall optimised solution
TA
FOR
4,00
5,00opsolution
overall
SOL
2,00
3,00
nalli
ng c
ost
ANA0,00
1,00
Sign
CON
, Sun
Mon
TueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
un
Time
Si lli t th h l d b i LASignalling cost more than halved by merging LAs
24
Performance analysis
TAAnalysis results
Comparison of methods ⇒ overall vs daily optimised solution
TA
1,80overall
FOR
1,50
1,60
1,70ng
cos
t
overall
daily
SOL
1 20
1,30
1,40
Sign
allin
ANA
1,20 Sun
Mon
TueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
un
40
50
overall
daily
[BSC
s]
CON20
30
Nbr
. of c
hang
es
Too many changesin the network
[
0
10
Sun
Mon
TueW
edThuFriS
atS
unM
onTueW
edThuFriS
atM
onM
onTueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
un
N
Time
25
Performance analysis
TAAnalysis results
Comparison of methods ⇒ overall vs daily optimised solution
TA
1,80overall
FOR
1,50
1,60
1,70ng
cos
t
overall
daily
SOL
1 20
1,30
1,40
Sign
allin
ANA0,8
1
o
overall
daily
1,20 Sun
Mon
TueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
un
CON0 2
0,4
0,6
Popu
latio
n ra
tio
Too many usersaffected by
changes frequently
0
0,2
Sun
Mon
TueW
edThuFriS
atS
unM
onTueW
edThuFriS
atM
onM
onTueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
un
Time
g q y
26
Performance analysis
TAAnalysis results
Comparison of methods ⇒ daily vs period1,80
overall
TA
1,50
1,60
1,70ng
cos
toverall
daily
period
FOR
Period based method
1 20
1,30
1,40
Sign
allin
SOL
Period-based methodachieves near-optimal
performance …
1,20 Sun
Mon
TueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
un
ANA40
50overall
daily
period[BSC
s]
CON10
20
30
Nbr
. of c
hang
es
period
… with less changesin the network
0
10
Sun
Mon
TueW
edThuFriS
atS
unM
onTueW
edThuFriS
atM
onM
onTueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
un
N
Time
27
Performance analysis
TAAnalysis results
Comparison of methods ⇒ perfect vs imperfect estimation1,80
overall
TA
1,50
1,60
1,70in
g co
stoverall
daily
period
period est
FOR
Estimation errors
1,20
1,30
1,40
Sign
alli
SOL
Estimation errorsmight lead to
forbidden solutions
1 week is not enough, Sun
Mon
TueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
un
40
50overall
daily
period
ANA
gfor predicting
[BSC
s]
10
20
30
Nbr
. of c
hang
es
period
period est
CON
0
10
Sun
Mon
TueW
edThuFriS
atS
unM
onTueW
edThuFriS
atM
onM
onTueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
un
Time
28
Performance analysis
TAAnalysis results
Comparison of methods ⇒ perfect vs imperfect estimation with overload factor
TA
1,80overall
'( ) ( )pk pkω ωFOR
1,50
1,60
1,70
ng c
ost
overall
daily
period
period est (r=1.05)
( ) ( )pk pki irω ω=
SOL
1,20
1,30
1,40
Sign
alli
Building estimates fromseveral week is better
than using overload factor
40
50overall
daily
period
ANA
1,20 Sun
Mon
TueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
un
[BSC
s]
10
20
30
Nbr
. of c
hang
es
period
period est (r=1.05)
CON
0
10
Sun
Mon
TueW
edThuFriS
atS
unM
onTueW
edThuFriS
atM
onM
onTueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
un
N
Time
29
Performance analysis
TAAnalysis results
Comparison of methods ⇒ local optimisation process
TA
1,80
FOR
1,50
1,60
1,70
ng c
ost
overall
daily
period
period est (r=1.05)
SOL
1 20
1,30
1,40
,
Sign
allin
period est opt (r=1.05)
Some benefit fromdisplacing changes…
ANA
1,20 Sun
Mon
TueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
un
40
50overall
daily
displacing changes…
[BSC
s]
CON20
30
Nbr
. of c
hang
es period est (r=1.05)
period est opt (r=1.05)
but increasing the
[
0
10
Sun
Mon
TueW
edThuFriS
atS
unM
onTueW
edThuFriS
atM
onM
onTueW
edThuFriS
atS
unM
onTueW
edThuFriS
atS
un
N
Time
… but increasing thefrequency of changes.
30
Performance analysis
TAAnalysis results
Comparison of methods
TA
1,7
overall
daily
period
1,7overall
daily
period
FOR
1,6
gnal
ling
cost
period
period est
period est (r=1.05)
period est opt1,6
gnal
ling
cost period est
period est (r=1.05)
period est optSOL
Avg.
sig
Avg.
sig
ANA
1,50 0,1 0,2 0,3
Avg. population ratio affected by changes
1,50 5 10 15
Avg. nbr. of changes CON[BSCs]
Period-based TA optimisation has the best trade-off between signalling cost and number of changes
31
Conclusions
TA1 The tracking area re-planning problem
2 Graph-theoretic formulation
TA
3 Solution method
4 Performance analysis
FOR
4 Performance analysis
5 Conclusions
SOL
5 Conclusions
Main results
Open issues ANA
CON
32
Conclusions
TAMain results
Problem formulation
TA
Possible to use commercial partitioning packages for TA planning problem
Graph correlation
N t k h h hi h l ti b t b i d k d
FOR
Network graphs show high correlation between business days or week-ends
Correlation becomes smaller as time goes by
Graph correlation coefficient can be used to detect need for re-planning
SOL
Solution method
Most of the benefit of TA re-planning is obtained by changing plan twice a week ANA
Need for averaging measurements over several weeks to build reliable graphs
Open issues CON
New TA concepts ⇒ TA list, overlapping TAs
Dynamic approaches ⇒ Reactive (e.g., problem-triggered) method
top related