viability of reverse pricing in cellular networks: a new ... · road to 5g: “how to match supply...
TRANSCRIPT
Radio resource Management and Optimization
Seong-Lyun Kim Professor, School of EEE
Assoc Dean, College of Eng Director, Center for Flexible Radio
Yonsei Univ., Seoul, Korea
Viability of Reverse Pricing in Cellular Networks: A New Outlook on Resource Management
May 8, 20152015 KTH-Yonsei Joint Workshop on 5G research
Joint work with Sang Yeob Jung, Joon Soo Kim, and Jun Ki Hong
2
DB
Server
MAC
0.4 6 GHz
sparse sensingl
Licensed Spectrum
Open Access Spectrum
Licensed Flexible Service over Shared Spectrum:
user-deployed, low cost/complexity, ultra-dense cells
Pain “Sharing”
Protection
FR+
Cognitive Radio (conventional)
Center for Flexible Radio: Practically Limitless User Spectrum (PLUS)
3
Spectrum Res.: 50 kHz Time Res.: 200 ms Spatial Res.: 10 m x 10 m
>10dBm-125dBm
RadarTx
WiFiAP
LTEBS
DTVTower
Sensing Map
frequency
x-position
y-position
RadarTx-Rx
WiFiAP-Rx
LTEBS
DTVTower
LTERx
DTVRx
Opportunity Map 1단계: Spectrum Sensing Map • Receiver Mobility • Interference Prediction 2단계: DSA Opportunity Map
High-Resolution Low-Complexity Sensing at Yonsei
Sensing Map ➥ Opportunity Map
•DTVprimary
•LTE
•WiFi •Radar
•PS-LTE • LAA-LTE • D2D-LTE secondary
Low opportunity (0)
High opportunity (1)
Cloud
DB
ServerMAC
Cloud Sensing/OP Map DB Cloud MAC Interf. Mngmt.
Cloud DB/MAC at SNU
FBMC-based Flexible Duplex
Realtime DSA at Yonsei
Dynamic Spectrum Access via High-Resolution/Low-Complexity Spectrum Sensing
4
Part 1. Introduction
Trends of Mobile Data Traffic Growth
5
12
Nomadic Heavy User Problem (NHP)
1. Total traffic volume increase: 3.04 times (47.66 → 145.05 MB/day). 2. Aggregate walking and nomadic user traffic proportion increase: 2.64 times
(14.74 → 38.93%)3. Nomadic heavy user traffic proportion increase: 2.83 times (10.35 → 29.29%)
[Shim15] T. Shim, J. Park, S.-W. Ko, S.-L. Kim, B. H. Lee, and J. G. Choi, "Traffic Convexity Aware Cellular Networks: A Nomadic Heavy User Perspective," submitted to IEEE Wireless Communications (under revision).
6
11
Mobile Traffic Experimental Environments
Year 2012 2015
Radio access technology WCDMA / LTE WCDMA / LTE / LTE-A
Number of participants82
(male: 61, female: 21)70
(male: 54, female: 16)
Experimental period avg. 1.28 weeks avg. 1.13 weeks
Measurement applications LifeMap My Data Manager, Moves
User speedWalking: up to 10 km/h (avg. 2.58 km/h)
Nomadic: above 10 km/h (avg. 31.9 km/h)
[Shim15] T. Shim, J. Park, S.-W. Ko, S.-L. Kim, B. H. Lee, and J. G. Choi, "Traffic Convexity Aware Cellular Networks: A Nomadic Heavy User Perspective," submitted to IEEE Wireless Communications (under revision).
Visual Example of Our Experimental Data Collection
User Convexity: Convex-Shaped over Velocity
7
[Shim15] T. Shim, J. Park, S.-W. Ko, S.-L. Kim, B. H. Lee, and J. G. Choi, "Traffic Convexity Aware Cellular Networks: A Nomadic Heavy User Perspective," submitted to IEEE Wireless Communications (under revision).
13
Stationary Walking Vagabond0
10
20
30
40
50
60
70
80
90
100
Tra
ffic
Vo
lum
e (
MB
/da
y)
2015
2012
2.35
User Convexity: 3.04
Nomadic
Traffic Volume over User Speed: User Convexity
Def. User convexity: the traffic volume ratio of nomadic users to walking users
User convexity increase: 1.29 times (2.35 → 3.04)
Solution: Traffic Convexity Aware Cell Association
8
[Shim15] T. Shim, J. Park, S.-W. Ko, S.-L. Kim, B. H. Lee, and J. G. Choi, "Traffic Convexity Aware Cellular Networks: A Nomadic Heavy User Perspective," submitted to IEEE Wireless Communications (under revision).
Road to 5G: “How to Match Supply to Demand?”
9
Pricing: Prof. Frank Kelly’s Seminal Work [Kelly98]
• Effective tool for design, operation, and management in cellular networks
[Kelly98] F. Kelly, A. Maulloo, and D. Tan “Rate control for communication networks: Shadow price proportional fairness and stability,” J. Oper. Res. Soc., 1998.[Ha12] S. Ha, S. Sen, C. Joe-Wong, Y. Im, and M. Chiang, “TUBE: Time dependent pricing for mobile data,” Proc. ACM SIGCOMM, 2012.
Mismatch: “Theory” and “Practice” • High spatiotemporal user demand uncertainty
➡ Network demand: 10x in peak-hours vs. off-peak hours [Ha12]
“New congestion pricing framework”
Road to 5G: New Congestion Pricing Framework (1/3)
10
Smart Data Pricing (SDP) [Sen13]
• Broad set of ideas and principles beyond the traditional pricing (e.g., flat-rate or usage-based pricing)
➡ Edge devices as a part of network management
➡ Not just how much to charge, but also when and how to charge
[Sen13] S. Sen, C. Joe-Wong, S. Ha, and M. Chiang, Smart data pricing (sdp): Economic solutions to network congestion. SIGCOMM eBook on Recent Advances in Networking, 2013.
User participation in pricing
Traditional
Flat-rate Usage-based
Operator-driven SDP
Time dependent Pricing
User-driven SDP
Reverse pricing
Time-dependent
pricing
SDP 1. Operator-driven Time-Dependent Pricing (Forward Pricing) [Ha12] • Low prices in less-congested periods, shifting usage from peak to off-peak periods
Road to 5G: New Congestion Pricing Framework (2/3)
11[Ha12] S. Ha, S. Sen, C. Joe-Wong, Y. Im, and M. Chiang, “TUBE: Time dependent pricing for mobile data,” Proc. ACM SIGCOMM, 2012.
Fig. 1. Spectrum usage efficiency over time
➡ CLAIM: Operator still has to face or predict demand uncertainty
Road to 5G: New Congestion Pricing Framework (3/3)
12
SDP 2. Reverse Pricing [Jung15] • Common in travel industry (e.g., airlines and hotels) - priceline.com
• Concept: Data amount/rate purchased, rather than sold
[Jung15] S. Y. Jung and S.-L. Kim, “Viability of reverse pricing in cellular networks: A new outlook on resource management, http://arxiv.org/abs/1504.06395
Resource recommendation rule: 10 GB + ? GB
User 1
User 2
Available capacity: 20 GB
Operator
Price: 1 $/GB
Resource recommendation rule: 5 GB + ? GB
Hidden bid-acceptance threshold: ? $/GB
Name user 1’s price: ? $/GB
Name user 2’s price: ? $/GB
Demand: 10 GB
Payment: 10 $
Demand: 5 GB
Payment: 5 $
Objective
13
Q. How to design reverse pricing w/ forward pricing to achieve “triple-win” solutions?
1. Resource utilization efficiency
Realistic Issues - Incomplete Information1. Demand uncertainty: user demand expectation but not variance
2. User heterogeneity: willingness to pay, bidding profiles
2. Users’ total payoffs
3. Operator’s total revenue
14
Part 2. Teaser
15
Viability of Reverse Pricing (1/7)
Fig. 2. Toal user demand
Resource Utilization Efficiency
16
Viability of Reverse Pricing (2/7)
Users’ Total Payoffs
Fig. 3. Total user payoff
17
Viability of Reverse Pricing (3/7)
Fig. 4. Operator’s revenue
Operator’s Total Revenue
18
Q. Is proposed reverse pricing is Real, Implementable? ➡ We Architected & Prototyped a fully functional reverse pricing system via App. & Server ➡ Through experiment study, we will confirm viability of reverse pricing or compare theory and practice
Main
My Information
Name your own price
Results
Data usage pattern
Logout
‘Trafficbid'
App Server
‘AWS’
Android ver.
http://ramoyonsei.angelworks.co.kr/ta_admin/index.php?/home
Hello. This website is for reverse pricing
Bid Profiles
User Information
Settings
Can download to search “Trafficbid”
Viability of Reverse Pricing (4/7)
19
Viability of Reverse Pricing (5/7)
App. Prototype
20
Viability of Reverse Pricing (6/7)
Results
Number Time Winning bid Success / Fail
1 00:00 ~06:00 0.2 $/ MB O
2 06:00 ~ 12:00 0.3 $/ MB X
3 12:00 ~ 18:00 0.4 $/ MB O
4 18:00 ~ 24:00 0.7 $/ MB O
Data usage pattern
0
10
20
30
40
0 ~ 6 6 ~ 12 12 ~ 18 18 ~ 24
Dat
a am
ount
(MB)
Time (hour)
App. Prototype
21
Viability of Reverse Pricing (7/7)
Server• Set minimum participation price, winning bid, user-specific recommended data amount
22
Part 3. System Model
System Model (1/3)
23
Resource-Constrained & Time-Slotted Network• A total amount of limited resource (e.g., data amount, rate, bandwidth, etc)
• A single operator
• A set of users
• A set of time slots (e.g., peak, normal, off-peak demand slots)
Operator
user1 userI
…
Q
I = {1, · · · , I}
H = {1, · · · , H}
User
1’s
dem
and
0
25
50
75
100
Resource scheduling horizon
0 4 8 12 16 20 24
Fig. 5. Resource-constrained & time-slotted networkUs
er I’s
dem
and
0
25
50
75
100
Resource scheduling horizon
0 4 8 12 16 20 24
System Model (2/3)
24
ui(✓hi , s
hi , p
h) = ✓hi ln(1 + shi )� phshi
User’s Payoff Function• The payoff function for user at time slot
➡ Logarithmic utility function is commonly used to model the proportionally fair resource allocation [Basar02]
➡ is the maximum price per unit resource for user , implying changing necessities of resource over time
i h
time-varying willingness to pay resource ex-ante price
[Basar02] T. Basar and R. Srikant, “Revenue-maximizing pricing and capacity expansion in many-users regime,” Proc. IEEE INFOCOM, 2002.
✓hi i
System Model (3/3)
25
Interaction Between Operator and Users
Fig. 6. Timing of the game that characterizes the interaction between operator and users
Stage I. Pricing based on demand predictions
Stage II. Use or report resources as a price taker
Stage IV. Name your own price
Stage III. Resource allocation rule & Hidden bid-acceptance threshold
Analysis: Backward Induction
Timing of the game
Assumption
26
Part 4. Analysis
Stage IV. Name Your Own Price (1/2)
27
Users’ Participation Decisions in Stage IV• Operator publicly specify a minimum participation unit price below which is not sold
• User decides whether to take part in the pricing process, i.e.,
• From (1), define the set of potential users that would name strictly positive prices at time slot , i.e.,
0 phmin < phx
hi
i
✓
hi ln(1 + x
hi )� p
hminx
hi � ✓
hi ln(1 + s
hi )� p
hx
hi (1)
I+(phmin) =
(i 2 I : phmin
✓
hi ln
�(1 + x
hi )/(1 + s
hi )�+ p
hs
hi
x
hi
)
h
(2)
recommended resource by operator in Stage III
user1 userI
…useri
…Group 1: name own price! Group 2: Do not name own price!
Stage IV. Name Your Own Price (2/2)
28
Users’ Bidding Strategies in Stage IV• Hidden bid-acceptance threshold (price):⌧h ⇠ U [phmin, p
h]
max
phminbhi ph
ui(✓hi , x
hi , b
hi )P(b
hi � ⌧
h) + ui(✓
hi , s
hi , p
h)P(b
hi < ⌧
h), 8i 2 I+
(p
hmin), 8h 2 HP1: (3)
Proposition 1. For each user , the optimal solution for Problem P1 at each
time slot is given by
i 2 I
h 2 H
(4)
(1
2xhi
h✓
h
i
ln⇣
1+x
hi
1+s
hi
⌘+ s
h
i
p
h + x
h
i
p
h
min
i, 8i 2 I+(ph
min
),
0, 8i 2 I\I+(phmin
).bh
⇤
i =
Stage III. Resource Recommendation Rule & Hidden Bid-Acceptance Threshold (1/4)
29
Issue 1. Resource Allocation Rule • Q. What is fairness criterion among users for resource allocation?
➡Constraint 1. Demand Capacity
➡Constraint 2. Revenue via reverse pricing Revenue via forward pricing
User 1
User 2
Operator
10 GB
5 GB
20 GB
1 $/GB
�
‘user payment residual resource’
Stage III. Resource Recommendation Rule & Hidden Bid-Acceptance Threshold (1/4)
30
Issue 1. Resource Allocation Rule (Cont’d)• Q. What is fairness criterion among users for resource allocation?
➡ A. Proportional residual resource recommendation rule
/
User 1
User 2
Operator10 GB + (2/3) 5 GB
5 GB + (1/3) 5 GB
20 GB
⇥
⇥1 $/GB
Stage III. Resource Recommendation Rule & Hidden Bid-Acceptance Threshold (2/4)
31
Issue 2. Minimum Participation Unit Price• Q. How to publicly specify a minimum participation unit price?
➡A. At least the revenue earned by forward pricing
User 1
User 2
Operator10 GB + (2/3) 5 GB
5 GB + (1/3) 5 GB 20 GB
⇥
⇥
Lemma 1. The minimum participation unit price that the operator sets should satisfy the following condition
ph > phmin � phP
i2I shiQ
, 8h 2 H.
1 $/GB, 0.75 $/GB
(5)
Stage III. Resource Recommendation Rule & Hidden Bid-Acceptance Threshold (3/4)
32
Issue 2. Minimum Participation Unit Price (Cont’d)• Q. How to set minimum participation unit price for revenue maximization?
➡A. As a strategic variable, taking the following tradeoff into account
0
75
150
225
300
User 1 User 2 User 3 User 4
Low Highphmin phminbh
⇤
i
Fig. 7. The comparison of bidding profiles under different minimum participation unit prices
With high ,
do not name prices
phmin
With high ,
tend to name higher prices
phmin
Stage III. Resource Recommendation Rule & Hidden Bid-Acceptance Threshold (4/4)
33
Issue 3. Hidden Bid-Acceptance Threshold• Q. How to set hidden bid-acceptance threshold given a specific minimum participation unit price?
Proposition 2. For any minimum participation unit price subject to (5), the hidden bid acceptance threshold is given by
phmin
⌧h⇤= phmin (6)
➡A. Do not forgo potentially profitable trades via reverse pricing
Stage I and II. Forward Pricing and Users’ Desired Resources
34
Stage II. Users’ Desired Resources
P2: 8i 2 I, 8h 2 Hmax
shi
ui(✓hi , s
hi , p
h) = ✓hi ln(1 + shi )� phshi , (7)
Stage I. Forward Pricing based on Demand Predictions
max
p�0
X
h2H
X
i2IE✓h
i
�✓hi � ph
�P3:
s.t.X
i2I
✓✓hiph
� 1
◆ Q, 8h 2 H.
(8)
(9)
Proposition 3. Given , the optimal solution to Problem P3 is given by Q >
Pi2I ✓̃hi + ✏hi✓̃hI � ✏hI
� I
ph⇤=
Pi2I ✓̃hi + ✏hiQ+ I
, 8h 2 H. (10)
Demand uncertainty stemming from time-varying willingness to pay
35
Part 5. Numerical Results
36
Viability of Reverse Pricing (1/3)
Fig. 8. Toal user demand
Resource Utilization Efficiency • Forward pricing: still fluctuating residual capacity due to intrinsic demand uncertainty
• Reverse pricing w/ forward pricing: no capacity left over resource scheduling horizon (only in this example)
37
Viability of Reverse Pricing (2/3)
Users’ Total payoffs • The proposed pricing scheme always outperforms the forward pricing only
• Increased flexibility for pricing
Fig. 9. Total user payoff
➡ total user payoff
38
Viability of Reverse Pricing (3/3)
Fig. 10. Operator’s revenue
Operator’s Total Revenue • Perhaps surprisingly, operator can extract more revenue by reverse pricing
• Operator should re-examine the idea of involving users in pricing process
39
Goal of Maximizing Operator’s Revenue (1/3)
Fig. 11. Operator’s revenue under
Operator’s Total Revenue • At a particular time slot 1, should be chosen as 1.515, achieving maximum revenue gain (e.g., 28%)
• However, it is impossible to set it in reality due to some uncertainty of bidding profiles
p1min
p1min
40
Goal of Maximizing Operator’s Revenue (2/3)
Fig. 12. Total user payoff under
Users’ Total Payoffs • Total user payoff is non-increasing over
• This is due to the fact that users are induced to bid with higher prices as increases
p1min
p1min
p1min
41
Goal of Maximizing Operator’s Revenue (2/3)
Fig. 13. Resource utilization efficiency under
Resource Utilization Efficiency • Resource utilization efficiency is non-increasing over
• When , only some users name their own prices, decreasing resource utilization efficiency
• When , no users name their own prices, degenerating forward pricing only case.
p1min
p1min
1.515 < p1min < 1.519
1.519 p1min p1⇤= 1.5616
42
Part 6. Conclusion
43
The idea of involving users in pricing decisions seems counterproductive. But, it may be the time to ‘re-examine’ this assumption.
Q. How to design reverse pricing w/ forward pricing to achieve “triple-win” solutions?
1. Operator’s revenue - YES
2. Resource utilization efficiency - YES
3. Net utility summed over all users - YES
Concluding Remarks (1/2)
Radio resource Management and Optimization
Seong-Lyun Kim Professor, School of EEE
Assoc Dean, College of Eng Director, Center for Flexible Radio
Yonsei Univ., Seoul, Korea
Thank You
May 8, 20152015 KTH-Yonsei Joint Workshop on 5G research
Joint work with Sang Yeob Jung, Joon Soo Kim, and Jun Ki Hong
Reference
45
[Shim15] T. Shim, J. Park, S.-W. Ko, S.-L. Kim, B. H. Lee, and J. G. Choi, "Traffic Convexity Aware Cellular
Networks: A Nomadic Heavy User Perspective," submitted to IEEE Wireless Communications (under revision).
[Kelly98] F. Kelly, A. Maulloo, and D. Tan “Rate control for communication networks: Shadow price proportional
fairness and stability,” J. Oper. Res. Soc., 1998.
[Ha12] S. Ha, S. Sen, C. Joe-Wong, Y. Im, and M. Chiang, “TUBE: Time dependent pricing for mobile data,”
Proc. ACM SIGCOMM, 2012.
[Sen13] S. Sen, C. Joe-Wong, S. Ha, and M. Chiang, Smart data pricing (sdp): Economic solutions to network
congestion. SIGCOMM eBook on Recent Advances in Networking, 2013.
[Jung15] S. Y. Jung and S.-L. Kim, “Viability of reverse pricing in cellular networks: A new outlook on resource
management, http://arxiv.org/abs/1504.06395
[Basar02] T. Basar and R. Srikant, “Revenue-maximizing pricing and capacity expansion in many-users regime,”
Proc. IEEE INFOCOM, 2002.