Mobile Data Offloading: A Tutorial Jianwei Huang Network Communications and Economics Lab (NCEL) Department of Information Engineering The Chinese University of Hong Kong (CUHK) Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 1 / 147 Slides Available Online Google “Jianwei Huang” http://jianwei.ie.cuhk.edu.hk/Files/MDO-Tutorial.pdf Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 2 / 147 Global Mobile Data Traffic, 2013 to 2018 Global Mobile Data Traffic Overall mobile data traffic is expected to grow to 15.9 exabytes per month by 2018, nearly an 11-fold increase over 2013. Mobile data traffic will grow at a CAGR of 61 percent from 2013 to 2018 (Figure 1). Figure 1. Cisco Forecasts 15.9 Exabytes per Month of Mobile Data Traffic by 2018 Global Mobile Data Traffic Growth Projection (source: Cisco VNI Mobile 2014) The Asia Pacific and North America regions will account for almost two-thirds of global mobile traffic by 2018, as shown in Figure 2. Middle East and Africa will experience the highest CAGR of 70 percent, increasing 14-fold over the forecast period. Central and Eastern Europe will have the second highest CAGR of 68 percent, increasing Annual growth rate ∼over61% 13-fold the forecast period. The emerging market regions of Asia Pacific and Latin America will have CAGRs I I of 67 percent and 66 percent respectively. Expected to reach 15.9 exabytes per month by 2018 A 11-fold increase over 2013 Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 3 / 147 Cellular Mobile Figure Network Capacity 3: Historical Increases in Spectral Efficiency 16 Historical Increases in Spectral Efficiency (source: Femtoforum) If available spectrum is increasing at 8% per year and the number of cell sites is increasing at 7% per year and technology performance is improving at 12% per year then operators can expect their network capacities to increase – on average – at 29% per year (1.08 x 1.07 x 1.12). If network capacity is growing at 29% per year and demand is growing currently at 108%band per year, then there 8% is a significant gap, which begs for I Available spectrum growth: per year further innovation. Annual grow rate ∼ 36% I I Cell site increase: 7% per year What other options exist? One possibility < is architectural What if – the2013) Spectrum efficiency growth: 18% perinnovation. year (2007 definition of a “cell site” were radically changed, in such a way that the number of “sites” dramatically increased and the cost per unit of capacity (after adjusting for the 108%sites) · 107% · 118% = 136% inevitable lower utilisation of smaller significantly decreased? Similar innovation has occurred before in the cellular industry. Decades ago omni-directional sites were sectorised. Operators began adding “down tilt” to their urban site designs. Operators began introducing underlay and overlay sites. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 4 / 147 Background Widening Supply-Demand Gap Network capacity growth vs Data traffic growth 29% vs 66% 36% vs. Slow network capacity growth Network Capacity Lin Gao (NCEL, IE@CUHK) Jianwei Huang (CUHK) 61% vs. Fast data traffic Datagrowth Traffic Mobile Data Offloading Mobile Data Offloading (Tutorial) May 2012 June 2015 1/1 5 / 147 How to Narrow the Gap: “Hard” Approaches Expanding the network capacity through technology innovations I Acquiring new spectrum bands I More efficient interference management through cooperations I Developing high-frequency wireless technology I Upgrading access technology (e.g., WCDMA → LTE → LTE-A) I I Building more pico/micro/macro cell sites ... Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 6 / 147 How to Narrow the Gap: “Hard” Approaches Expanding the network capacity through technology innovations I Acquiring new spectrum bands I More efficient interference management through cooperations I Developing high-frequency wireless technology I Upgrading access technology (e.g., WCDMA → LTE → LTE-A) I I Building more pico/micro/macro cell sites ... Challenges: need to be cost effective and easy to deploy. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 6 / 147 How to Narrow the Gap: “Soft” Approaches Reshaping the demand through economics and software I Tired data pricing I Capped or throttling (e.g., 128kbps if monthly usage >5GB) I Time/Location/Congestion dependent pricing (e.g., delay coupons) I Application specific optimization (e.g., network-friendly implem.) I On device software client (e.g., “fuel gauge” meters) I Content specific control (e.g., two-sided 1-800 pricing) I ... Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 7 / 147 How to Narrow the Gap: “Soft” Approaches Reshaping the demand through economics and software I Tired data pricing I Capped or throttling (e.g., 128kbps if monthly usage >5GB) I Time/Location/Congestion dependent pricing (e.g., delay coupons) I Application specific optimization (e.g., network-friendly implem.) I On device software client (e.g., “fuel gauge” meters) I Content specific control (e.g., two-sided 1-800 pricing) I ... Challenges: need to be user-friendly and network neutral. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 7 / 147 Today’s Focus: Mobile Data Offloading Basic idea: deliver cellular traffic over Wi-Fi or Femtocell. BS2 MU13 BS1 AP1 MU11 MU21 AP2 MU24 AP4 MU14 MU32 AP3 BS3 MU33 MU31 MU11 & MU21 →AP1 , MU24 →AP2 , MU31 & MU33 →AP3 , MU14 & MU32 →AP4 . Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 8 / 147 A Reality Check Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 9 / 147 Global mobile data traffic would grow at a CAGR of 65 percent instead of 61 percent. Offload volume is determined by smartphone penetration, dual-mode share of handsets, percentage of home-based mobile Internet use, and Global Mobile Data Offloading percentage of dual-mode smartphone owners with Wi-Fi fixed Internet access at home. Figure 14. 52 Percent of Total Mobile Data Traffic Will Be Offloaded by 2018 Mobile Traffic Offloading Prediction (source: Cisco VNI Mobile 2014) The amount of traffic offloaded from smartphones will be 51 percent by 2018, and the amount of traffic offloaded from tablets will be 69 percent by 2018. Mobile offloading will increase from 45% in 2013 to 52% in 2018 A supporting trend is the growth of cellular connectivity for devices such as tablets which in their earlier generation were limited to Wi-Fi connectivity only. With increased desire for mobility and mobile carriers offer of data plans catering to multi-device owners, we find that the cellular connectivity is on a rise albeit cautiously as the end users are testing the waters. As a point in case, we estimate that by 2018, 42 percent of all tablets will have a cellular in 2013 (Figure 15). Jianwei Huangconnection (CUHK)up from 34 percent Mobile Data Offloading (Tutorial) June 2015 10 / 147 Offloading Increases with Technology Figure 16. Mobile Data Traffic and Offload Traffic, 2018 Mobile and Offloaded Traffic from Mobile-Connected Devices (source: Cisco VNI Mobile 2014) Trend 6: Comparing Mobile Network Speeds Globally, the average mobile network connection speed in 2013 was 1,387 Kbps. The average speed will grow at a compound annualattract growth rate of 13 percent, and willdevices. exceed 2.5 Mbps by 2018. Smartphone speeds, generally 4G networks will high-usage third-generation (3G) and higher, are currently almost three times higher than the overall average. Smartphone The offloading ratio 4Greaching will 7be speeds will nearly double on by 2018, Mbps.the highest. There is anecdotal evidence to support the idea that usage increases when speed increases, although there is often a delay between the increase in speed and the increased usage, which can range from a few months to several years. The Cisco VNI Forecast relates application bit rates to the average speeds in each country. Many Jianwei Huangof the (CUHK) Mobile Data can Offloading June 2015rates for 11 / 147 trends in the resulting traffic forecast be seen (Tutorial) in the speed forecast, such as the high growth Complementary Cellular Small Cells and Wi-Fi Cellular small cells provide a uniform and reliable capacity layer and better coverage. WiFi provides a more powerful capacity boost. Technology co-location reduces Capex and Opex for offloading. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 12 / 147 Case Study: AT&T in US 2.7B 1.23B 382.1M 85.5M 2009 2010 2011 2012 AT&T Annual Wi-Fi Connections (source: AT&T) Statistics of 2012 I I I 32, 000 Wi-Fi hotspots 2.7 billion Wi-Fi connections (80% from AT&T’s mobiles) 12.9 billion MB Wi-Fi data (5.2 billon MB from mobile devices) Seamless offloading I I Auto-login of AT&T Wi-Fi hotspots Auto-roaming to Fon Wi-Fi hotspots supported by Hotspot 2.0 Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 13 / 147 AT&T Small Cell Strategy AT&T Small Cell Strategies (source: AT&, Senza Fili) Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 14 / 147 Case Study: Sprint in US Spring Boingo WiFi collaboration announced April 2015 (source: Internet) Sprint customers access Boingo hotspots for free in 35 major US airports 40 out of 56 millions of Sprint’s customer devices can auto-authenticate with Boingo Wi-Fi hotspot connections Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 15 / 147 Case Study: China Mobile China Mobile Cellular and WiFi Traffic (source: China Mobile, Senza Fili) 4.2 millions Wi-Fi APs deployed in 2012 Reach 6 millions in the next three years Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 16 / 147 Nanocell Architecture China Mobile NanoCell Strategy Access to Cellular Core network and WLAN Core Network MME HSS LTE Nanocell Backhaul Nanocell GW P-GW S-GW Internet WLAN WLAN Portal LTE and WLAN Inference Coordination EPC Nanocell OAM Nanocell SoC Chip Architecture Flexible Backhaul AC AAA Trusted Secured Environment Unified Authentication (source: China Mobile) 6 Combine LTE femtocell and carrier-grade Wi-Fi in the same box Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 17 / 147 Our Focus on Wi-Fi White paper Carrier Wi-Fi® for mobile operators fic for en ey do ccess scriber om, orks, f Wi-Fi. e less hey rely Wi-Fi fload – scriber Figure 3. Cellular and Wi-Fi traffic from mobile devices. FourSenza Types Source: Fili of Wi-Fi (source: Senza Fili) coffee highly ective uch Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 18 / 147 Cost Benefit of Wi-Fi White paper Carrier Wi-Fi® for mobile operators ct he MHz The ed our ue sts h is , it Per-bit TCO of different technology choices (source: Senza Fili) er Hz Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 19 / 147 technology of choice. Operator Benefits The Wi-Fi offload business opportunity map for mobile carriers Defensive: Cut costs Seamless & automatic: Value of everyday convenience International roaming: Lifestyle & lower costs + Opportunistic: Boost revenues Relieve congestion Lower network CAPEX/OPEX Retain customers Reduce churn Additional services Attract new customers Mobile operator’s business opportunities due to Wi-Fi offload (source: Aptilo) NEW MASS-MARKET WI-FI OFFLOAD SERVICE TYPES service provisioning point of view Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 20 / 147 User Benefits Faster connections Lower battery drain (when close to AP) Easier to use Reduced TCP handshake Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 21 / 147 Different Offloading Types Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 22 / 147 Two Types of Data Offloading User-initiated offloading I User decides when and how to offload I When automatic offloading is not possible or users’ judgements needed Network-initiated offloading I Mobile operator makes the offloading decision I Seamless Wi-Fi Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 23 / 147 Seamless Wi-Fi Near-term: user transparence I Automatic handover from cellular to Wi-Fi I Automatic authentication by Wi-Fi I Traffic reroute to local Internet Long-term: carrier-grade Wi-Fi I Large bandwidth and high throughout based on latest Wi-Fi standards I Tight integration with cellular network through new standards I Traffic reroute to cellular operator’s core network I Cellular operator has control over quality and service experiences I Goes beyond data offload Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 24 / 147 Challenges of Wi-Fi Data Offloading Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 25 / 147 Challenges of Wi-Fi Data Offloading Maturity of Wi-Fi-cellular integrations Pricing of cellular and Wi-Fi services Quality of Wi-Fi experiences Deployment of Wi-Fi hotspots Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 26 / 147 Challenge 1: Maturity of Wi-Fi-Cellular Integrations Manual Wi-Fi network selection and input of username/password Tedious, time-consuming, and inconvenient Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 27 / 147 Solutions Operator-specific on-device configurations (AT&T) Standards: HotSpot 2.0 (a video), NGH, ANDSF in 3GPP I I Automatic Wi-Fi selection and login with strong security Already supported by Apple iOS Wi-Fi Sessions the network of a commercially network. Measurements show of Wi-Fi sessions in the course sessions grows steeply. Increase of SIM-based Wi-Fi Offloading (source: Aptilo) 3 Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 28 / 147 Challenge 2: Pricing of Cellular and Wi-Fi Services Is Wi-Fi free? Flat-fee? Usage-based? How is cellular charged? Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 29 / 147 Challenge 2: Pricing of Cellular and Wi-Fi Services Is Wi-Fi free? Flat-fee? Usage-based? How is cellular charged? Several possibilities I Combined cellular and Wi-Fi service with a total monthly data cap I Volume-capped cellular service with unlimited free Wi-Fi access I Low cost Wi-Fi data service for customers without a cellular data plan I Session-based or subscription-based Wi-Fi service only Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 29 / 147 Challenge 2: Pricing of Cellular and Wi-Fi Services Is Wi-Fi free? Flat-fee? Usage-based? How is cellular charged? Several possibilities I Combined cellular and Wi-Fi service with a total monthly data cap I Volume-capped cellular service with unlimited free Wi-Fi access I Low cost Wi-Fi data service for customers without a cellular data plan I Session-based or subscription-based Wi-Fi service only Key Questions: how should the cellular operator I Jointly design cellular and Wi-Fi pricing plans? I Balance additional revenue and offloading benefits of Wi-Fi? Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 29 / 147 Challenge 3: Quality of Wi-Fi Experiences Not all Wi-Fi hotspots are created equal I 802.11b (11 Mbps) vs. 802.11n (150 Mbps) Neither are the cellular networks I I 3G (EDGE Evolution, 1.6 Mbps) vs. 4G (HSPA/LTE, 300 Mbps) Half of North American mobile connections on HSPA/LTE (2013) Real world data rates vary based on time and location. I I Cellular can be more predicable or even sometimes faster than Wi-Fi Also need to consider delay, consistency of delay, etc. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 30 / 147 Challenge 3: Quality of Wi-Fi Experiences Not all Wi-Fi hotspots are created equal I 802.11b (11 Mbps) vs. 802.11n (150 Mbps) Neither are the cellular networks I I 3G (EDGE Evolution, 1.6 Mbps) vs. 4G (HSPA/LTE, 300 Mbps) Half of North American mobile connections on HSPA/LTE (2013) Real world data rates vary based on time and location. I I Cellular can be more predicable or even sometimes faster than Wi-Fi Also need to consider delay, consistency of delay, etc. Key Question: When and how to offload traffic to Wi-Fi considering network conditions and application QoS requirements? Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 30 / 147 Challenge 4: Wi-Fi Hotspots Develoyment Historically mobile operators do not own large Wi-Fi hotspots Approach 1: Direct deployment I I I AT&T (in US), China Mobile, PCCW Pros: Easier to control and integrate Cons: Costly and time consuming to deploy, difficulty in finding deployment locations, providing backhaul, and managing multiple service provisions. Approach 2: Collaborations with Wi-Fi operators I I I Sprint and Boingo, T-Mobile and iPass, DT/BT and FON Pros: Fast and flexible Cons: Complicated to manage integration and revenue sharing Approach 3: Dynamic network sharing and expansion Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 31 / 147 Challenge 4: Wi-Fi Hotspots Develoyment Historically mobile operators do not own large Wi-Fi hotspots Approach 1: Direct deployment I I I AT&T (in US), China Mobile, PCCW Pros: Easier to control and integrate Cons: Costly and time consuming to deploy, difficulty in finding deployment locations, providing backhaul, and managing multiple service provisions. Approach 2: Collaborations with Wi-Fi operators I I I Sprint and Boingo, T-Mobile and iPass, DT/BT and FON Pros: Fast and flexible Cons: Complicated to manage integration and revenue sharing Approach 3: Dynamic network sharing and expansion Key Question: How multiple cellular operators interact in a competitive offloading market? Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 31 / 147 Recent Results Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 32 / 147 Recent Results Technology issues: I Delayed-aware offloading I Congestion-aware offloading I Predictive offloading Economics issues: I Operator bargaining I Offloading market I User-centric offloading and onloading Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 33 / 147 Technology Issues for Wi-Fi Offloading Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 34 / 147 Delayed Wi-Fi Offloading Wi-Fi have small coverages, and may not always be available. Delayed Wi-Fi offloading: delay a data transfer until the user enters a Wi-Fi hotspot. Delay-tolerant applications (e.g., movie download and software update): Tolerate certain delays without sacrificing user satisfactions. Exploiting human mobility and traffic diversity Save network cost and device energy Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 35 / 147 How Much Can Wi-Fi Offload? A recent study in Korean (LLYRC, “Mobile Data Offloading: How Much Can WiFi Deliver?” ToN’13) Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 36 / 147 How Much Can Wi-Fi Offload? A recent study in Korean (LLYRC, “Mobile Data Offloading: How Much Can WiFi Deliver?” ToN’13) On-the-spot offloading without delay I I Save 65% of network capacity Save 55% of energy Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 36 / 147 How Much Can Wi-Fi Offload? A recent study in Korean (LLYRC, “Mobile Data Offloading: How Much Can WiFi Deliver?” ToN’13) On-the-spot offloading without delay I I Save 65% of network capacity Save 55% of energy Offload with 100 secs delay (insignificant) I I Save additional 2-3% of network capacity Save additional 3% energy Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 36 / 147 How Much Can Wi-Fi Offload? A recent study in Korean (LLYRC, “Mobile Data Offloading: How Much Can WiFi Deliver?” ToN’13) On-the-spot offloading without delay I I Save 65% of network capacity Save 55% of energy Offload with 100 secs delay (insignificant) I I Save additional 2-3% of network capacity Save additional 3% energy Offload with 1 hour delay (significant) I I I Save additional 29% of network capacity Save additional 20% energy Suitable scenarios: software update, large file transfer, ... Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 36 / 147 How Much Can Wi-Fi Offload? A recent study in Korean (LLYRC, “Mobile Data Offloading: How Much Can WiFi Deliver?” ToN’13) On-the-spot offloading without delay I I Save 65% of network capacity Save 55% of energy Offload with 100 secs delay (insignificant) I I Save additional 2-3% of network capacity Save additional 3% energy Offload with 1 hour delay (significant) I I I Save additional 29% of network capacity Save additional 20% energy Suitable scenarios: software update, large file transfer, ... Key Question: How to optimize the delayed-based offloading? Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 36 / 147 Delay Optimal WiFi Offloading Joint work with Man Hon Cheung (CUHK) IEEE WiOpt 2013, IEEE JSAC 2015 Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 37 / 147 System Model ϭ Ϯ ϱ ϵ ϯ ϲ ĞůůƵůĂƌ^ ϭϬ ϰ ϴ ϳ ϭϭ ϭϮ tŝͲ&ŝW ϭϯ ϭϰ ϭϱ ϭϲ DŽďŝůĞhƐĞƌ A total of L = {1, . . . , L} locations I Cellular is available at all locations Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 38 / 147 System Model ϭ Ϯ ϱ ϵ ϯ ϲ ĞůůƵůĂƌ^ ϭϬ ϰ ϴ ϳ ϭϭ ϭϮ tŝͲ&ŝW ϭϯ ϭϰ ϭϱ ϭϲ DŽďŝůĞhƐĞƌ A total of L = {1, . . . , L} locations I Cellular is available at all locations Wi-Fi availability is location-dependent: I L(1) = {4, 11, 13, 16}, L(0) = L\L(1) . Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 38 / 147 System Model ϭ Ϯ ϱ ϵ ϯ ϲ ĞůůƵůĂƌ^ ϭϬ ϰ ϴ ϳ ϭϭ ϭϮ tŝͲ&ŝW ϭϯ ϭϰ ϭϱ ϭϲ DŽďŝůĞhƐĞƌ A total of L = {1, . . . , L} locations I Cellular is available at all locations Wi-Fi availability is location-dependent: I L(1) = {4, 11, 13, 16}, L(0) = L\L(1) . Mobility pattern: User moves from location l to l 0 with prob p(l 0 | l). Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 38 / 147 System Model ϭ Ϯ ϱ ϵ ϯ ϲ ĞůůƵůĂƌ^ ϭϬ ϰ ϴ ϳ ϭϭ ϭϮ tŝͲ&ŝW ϭϯ ϭϰ ϭϱ ϭϲ DŽďŝůĞhƐĞƌ A total of L = {1, . . . , L} locations I Cellular is available at all locations Wi-Fi availability is location-dependent: I L(1) = {4, 11, 13, 16}, L(0) = L\L(1) . Mobility pattern: User moves from location l to l 0 with prob p(l 0 | l). Deadline: A file of K bits must be sent by time T . Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 38 / 147 Tradeoff Objective: To achieve a good tradeoff between I Reducing cellular usage: Wait till entering WiFi hotspots. I Satisfying user’s QoS requirement: Transmit through cellular now if not meeting WiFi soon. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 39 / 147 Tradeoff Objective: To achieve a good tradeoff between I Reducing cellular usage: Wait till entering WiFi hotspots. I Satisfying user’s QoS requirement: Transmit through cellular now if not meeting WiFi soon. Question: Given the user’s mobility pattern and the WiFi availability, should the user remain idle, use cellular, or use Wi-Fi (if available) in each time slot? Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 39 / 147 Markov Decision Process Decision epochs: t ∈ T = {1, . . . , T }. State: s = (k, l) I I k: remaining file size l: location index Action: a I a = 0 (idle), 1 (cellular), 2 (Wi-Fi). I a ∈ A(l) = {0, 1}, I if l ∈ L(0) (Wi-Fi is not available). a ∈ A(l) = {0, 1, 2}, if l ∈ L(1) (Wi-Fi is available). Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 40 / 147 Markov Decision Process Cellular usage cost at time t ∈ T ( 1, if a = 1 (using cellular), ct (k, l, a) = 0, otherwise. Penalty for incomplete file transfer at T + 1: ĉT +1 (k, l) = h(k). I Nondecreasing in k with h(0) = 0. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 41 / 147 Markov Decision Process State transition probability: p (k 0 , l 0 ) | (k, l), a p (k 0 , l 0 ) | (k, l), a = p(l 0 | l) p k 0 | (k, l), a , I Probability ( 0 p k | (k, l), a = I I 1, if k 0 = [k−µ(l, a)]+ and a ∈ A(l) , 0, otherwise. µ(l, a): Data rate at location l with action a. p(l 0 | l): Obtain based on the mobility pattern of the MU. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 42 / 147 Markov Decision Process Optimization problem: To minimize the expected total cellular usage plus the penalty for incomplete file transfer " T # X π π π π min Es ct (s t , δt (s t )) + ĉT +1 (s T +1 ) . π∈Π I t=1 Policy π = (δt (k, l), ∀ k ∈ K, l ∈ L, t ∈ T ): Decision rules at all the states and time slots. We solve the problem using finite-horizon dynamic programming. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 43 / 147 Optimal Algorithm We propose an optimal delayed Wi-Fi offloading algorithm. It does not have closed-form in general. Difficult to derive engineering insights. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 44 / 147 Special Case: Single Threshold Optimal Policy Assume I I h(k) is a convex and nondecreasing, The cellular and Wi-Fi data rates are location independent. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 45 / 147 Special Case: Single Threshold Optimal Policy Assume I I h(k) is a convex and nondecreasing, The cellular and Wi-Fi data rates are location independent. Theorem The optimal policy π ∗ has a single location-dependent threshold in k: Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 45 / 147 Special Case: Single Threshold Optimal Policy Assume I I h(k) is a convex and nondecreasing, The cellular and Wi-Fi data rates are location independent. Theorem The optimal policy π ∗ has a single location-dependent threshold in k: I At a location without WiFi: transmit using cellular if k ≥ kt∗ (l), otherwise remain idle . Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 45 / 147 Special Case: Single Threshold Optimal Policy Assume I I h(k) is a convex and nondecreasing, The cellular and Wi-Fi data rates are location independent. Theorem The optimal policy π ∗ has a single location-dependent threshold in k: I I At a location without WiFi: transmit using cellular if k ≥ kt∗ (l), otherwise remain idle . At a location with WiFi: Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 45 / 147 Special Case: Single Threshold Optimal Policy Assume I I h(k) is a convex and nondecreasing, The cellular and Wi-Fi data rates are location independent. Theorem The optimal policy π ∗ has a single location-dependent threshold in k: I I At a location without WiFi: transmit using cellular if k ≥ kt∗ (l), otherwise remain idle . At a location with WiFi: F If Wi-Fi rate is smaller than cellular rate: transmit using cellular if k ≥ kt∗ (l), otherwise transmit using WiFi. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 45 / 147 Special Case: Single Threshold Optimal Policy Assume I I h(k) is a convex and nondecreasing, The cellular and Wi-Fi data rates are location independent. Theorem The optimal policy π ∗ has a single location-dependent threshold in k: I I At a location without WiFi: transmit using cellular if k ≥ kt∗ (l), otherwise remain idle . At a location with WiFi: F If Wi-Fi rate is smaller than cellular rate: transmit using cellular if k ≥ kt∗ (l), otherwise transmit using WiFi. F If Wi-Fi rate is larger than cellular rate: always transmit using WiFi. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 45 / 147 Special Case: Single Threshold Optimal Policy 20 20 18 18 Remaining File Size k (Mbits) Remaining File Size k (Mbits) Convex penalty and location-independent data rates 16 14 12 10 8 6 4 2 0 16 14 12 10 8 6 4 2 2 4 6 8 10 12 14 16 18 20 0 2 4 6 Time slot t 8 10 12 14 16 18 20 Time slot t Idle (◦), cellular (•), Wi-Fi (+). Left: Location without Wi-Fi: µ1 = 2 Mbps (cellular). Right: Location with Wi-Fi: µ1 = 2 Mbps (cellular), µ2 = 1 Mbps (WiFi). Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 46 / 147 General: Multi-threshold Optimal Policy 20 20 18 18 Remaining File Size k (Mbits) Remaining File Size k (Mbits) Step penalty and location-dependent data rates 16 14 12 10 8 6 4 2 0 16 14 12 10 8 6 4 2 2 4 6 8 10 12 14 16 18 20 0 2 4 6 Time slot t 8 10 12 14 16 18 20 Time slot t Idle (◦), cellular (•), Wi-Fi (+). Left: Location without Wi-Fi: µ(l, 1) = 2.1 Mbps. Right: Location with Wi-Fi: µ(l, 1) = 3.1 Mbps > µ(l, 2) = 2.1 Mbps. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 47 / 147 Performance Evaluations Compare three schemes: 1 ODWO: Our proposed Optimal delayed Wi-Fi offloading. 2 On-the-spot offloading: Offload to the Wi-Fi network whenever Wi-Fi is available. 3 Wiffler: prediction-based offloading [Balasubramanian MobiSys’10]. Setting: I Cellular and Wi-Fi data rate: random with mean = 3 Mbps and standard deviation = 1 Mbps. I Probability that Wi-Fi is available = 0.7. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 48 / 147 Performance Evaluations Setting: Deadline T = 3 min, penalty h(k) = k 2 , ∀ k ∈ K. 120 2.5 ODWO Wiffler On−the−spot On−the−spot Wiffler ODWO File Transfer Efficiency 100 Total Cost 80 60 40 1.5 1 0.5 20 0 20 2 25 30 35 40 45 50 55 60 0 20 File Size K (Mbytes) 30 40 50 60 70 File Size K (Mbytes) Define File transfer efficiency = probability of completing file transfer average number of cellular time slots used ODWO achieves the minimal cost and the highest file transfer efficiency. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 49 / 147 Summary Algorithm design: General and threshold-based ODWO. Analysis: Threshold optimal policy with convex penalty function and location-independent data rates. Performance evaluation: ODWO achieves the minimal total cost and the highest file transfer efficiency. Next Step: What about multiple users making decisions? Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 50 / 147 Congestion-Aware Network Selection and Data Offloading Joint work with Man Hon Cheung and Richard Southwell (CUHK) CISS 2014 Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 51 / 147 A Much More General Model Multiple users Wi-Fi availability: I I I Location-dependent: Limited Wi-Fi coverage. User-dependent: Different subscriptions and plans (e.g., Skype Wi-Fi). Time-dependent: open or closed access mode at different time. Network-dependent switching time and switching cost: I I Switching time: Delay during handoff. Switching cost: Additional power consumption and QoS disruption. Usage-based pricing Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 52 / 147 System Model ϭ Ϯ ϯ ϰ DhϮ ϱ ϵ ϭϯ ϲ ϭϬ ϭϰ ϭϭ ϭϱ Dhϭ ϴ ϳ EĞƚǁŽƌŬϭ ;ĞůůƵůĂƌͿ ϭϮ ϭϲ EĞƚǁŽƌŬϮ ;tŝͲ&ŝͿ Users I = {1, . . . , I }, networks N = {1, . . . , N}. Locations L = {1, . . . , L}, time slots T = {1, . . . , T }. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 53 / 147 System Model ϭ Ϯ ϯ ϰ DhϮ ϱ ϵ ϭϯ ϲ ϭϬ ϭϰ ϴ ϳ EĞƚǁŽƌŬϭ ;ĞůůƵůĂƌͿ ϭϭ ϭϱ Dhϭ ϭϮ ϭϲ EĞƚǁŽƌŬϮ ;tŝͲ&ŝͿ Users I = {1, . . . , I }, networks N = {1, . . . , N}. Locations L = {1, . . . , L}, time slots T = {1, . . . , T }. Cellular network is always available. Wi-Fi availability is user/location/time dependent: M(i, l, t) ⊆ N . Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 53 / 147 Resource Block MU 1 ϭ Ϯ ϯ ϰ DhϮ ϱ ϵ ϭϯ ϲ ϭϬ ϭϰ ϴ ϳ EĞƚǁŽƌŬϭ ;ĞůůƵůĂƌͿ ϭϭ ϭϱ Dhϭ 1 2 3 4 2 3 4 Time Network 1 Network 2 ϭϮ ϭϲ MU 2 EĞƚǁŽƌŬϮ ;tŝͲ&ŝͿ 1 Time Network 1 Network 2 Resource block: a network available at a particular time. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 54 / 147 Resource Block MU 1 ϭ Ϯ ϯ ϰ DhϮ ϱ ϵ ϭϯ ϲ ϭϬ ϭϰ ϴ ϳ EĞƚǁŽƌŬϭ ;ĞůůƵůĂƌͿ ϭϭ ϭϱ Dhϭ 1 2 3 4 2 3 4 Time Network 1 Network 2 ϭϮ ϭϲ MU 2 EĞƚǁŽƌŬϮ ;tŝͲ&ŝͿ 1 Time Network 1 Network 2 Resource block: a network available at a particular time. Trajectory of a user determines its resource blocks (unshaded ones). Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 54 / 147 Resource Block MU 1 ϭ Ϯ ϯ ϰ DhϮ ϱ ϲ ϭϬ ϵ ϭϰ ϭϯ ϴ ϳ EĞƚǁŽƌŬϭ ;ĞůůƵůĂƌͿ ϭϭ ϭϱ Dhϭ 1 2 3 4 2 3 4 Time Network 1 Network 2 ϭϮ ϭϲ MU 2 EĞƚǁŽƌŬϮ ;tŝͲ&ŝͿ 1 Time Network 1 Network 2 Resource block: a network available at a particular time. Trajectory of a user determines its resource blocks (unshaded ones). Example: I I User 1’s trajectory: (14, 15, 16, 16) ⇒ six resource blocks. User 2’s trajectory: (4, 8, 12, 16) ⇒ five resource blocks. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 54 / 147 Selection of Resource Blocks 1 MU 1 2 3 4 1 Network 2 2 1 MU 2 2 3 4 Time 2 1 Time Network 1 1 Network 2 2 4 Time t Route 1 MU 1 Route 2 Route 3 network n 1 Network 1 3 2 3 4 Time t Route 1 Route 2 MU 2 network n Selection of resource blocks ⇒ feasible route in the graph. I Resource block: vertex in the graph I Network selection between two time slots: edge in the graph Examples: switching time = 1. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 55 / 147 User Suplus = Utility - Payment Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 56 / 147 User Suplus = Utility - Payment (n) µ Utility = a m+x (n,t) (network throughput). I µ(n) : Capacity of network n ∈ N . I m: Congestion level. I x (n,t) : Background traffic of network n ∈ N at time t ∈ T . I a: Scaling weight. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 56 / 147 User Suplus = Utility - Payment (n) µ Utility = a m+x (n,t) (network throughput). I µ(n) : Capacity of network n ∈ N . I m: Congestion level. I x (n,t) : Background traffic of network n ∈ N at time t ∈ T . I a: Scaling weight. (n) µ Payment = γ(n) m+x (n,t) ∆t (usage-based pricing). I γ(n): Unit price of network n ∈ N . Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 56 / 147 User Suplus = Utility - Payment (n) µ Utility = a m+x (n,t) (network throughput). I µ(n) : Capacity of network n ∈ N . I m: Congestion level. I x (n,t) : Background traffic of network n ∈ N at time t ∈ T . I a: Scaling weight. (n) µ Payment = γ(n) m+x (n,t) ∆t (usage-based pricing). I γ(n): Unit price of network n ∈ N . Surplus = utility - payment (in a single time slot): σ (n,t) (m) = (a − γ(n)∆t) I µ(n) . m + x (n,t) This depends on the resource block (n, t) ⇒ one vertex v in route ri Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 56 / 147 User Payoff = Total Surplus - Total Switching Cost Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 57 / 147 User Payoff = Total Surplus - Total Switching Cost Switching Cost: computed based on edges in the route ri I I No channel switching: zero cost Channel switching: positive cost Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 57 / 147 User Payoff = Total Surplus - Total Switching Cost Switching Cost: computed based on edges in the route ri I I No channel switching: zero cost Channel switching: positive cost Payoff = total surplus - total switching cost: X X ρi (r ) = σ v mv (r ) − ge. v ∈V(ri ) e∈E(ri ) I Surplus σ v (mv ): depends on vertex v , coupled across users in mv . I Switching cost g e : depends on edge e, decoupled across users. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 57 / 147 Network Selection Game 2 1 3 1 2 network n 1 4 Time t Route 1 MU 1 Route 2 Route 3 2 1 3 4 Time t Route 1 Route 2 MU 2 2 network n Players: users. Strategies: feasible routes Each player chooses a route to maximize its payoff Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 58 / 147 Key Questions How should a user make a route choice? Will users’ choices converge to a network equilibrium? How fast does convergence happen? Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 59 / 147 How Should a User Choose its Route? Better response update: a user chooses a new route to improve his payoff, assuming that other users’ route choices are fixed. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 60 / 147 How Should a User Choose its Route? Better response update: a user chooses a new route to improve his payoff, assuming that other users’ route choices are fixed. Result: The complexity of computing a better response is polynomial in terms of number of users (O(I 2 )). I Key idea: finding a better response update is equivalent of computing a shortest path in a graph. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 60 / 147 Will Users’ Greedy Choices Converge? We want to achieve the Pure Nash equilibrium (NE): I A route choice profile r = {ri , ∀i ∈ I} where no user can perform a better response update Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 61 / 147 Will Users’ Greedy Choices Converge? We want to achieve the Pure Nash equilibrium (NE): I A route choice profile r = {ri , ∀i ∈ I} where no user can perform a better response update It would be nice to have the Finite improvement property (FIP), where better response updates always converge to a pure NE. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 61 / 147 Will Users’ Greedy Choices Converge? We want to achieve the Pure Nash equilibrium (NE): I A route choice profile r = {ri , ∀i ∈ I} where no user can perform a better response update It would be nice to have the Finite improvement property (FIP), where better response updates always converge to a pure NE. Result: every network selection game has the FIP. I Key idea: show that the game is equivalent to a congestion game. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 61 / 147 Total Number of Better Response Updates How Fast Does Convergence Happen? 160 140 120 100 80 60 40 20 0 5 10 15 20 25 30 Total Number of MUs I The convergence scales well with the number of users. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 62 / 147 Performance Evaluations Compare three schemes: I NSG: Network selection game. I On-the-spot offloading (OTSO): Offload to the Wi-Fi network whenever Wi-Fi is available. I Cellular-only: Use the cellular network all the time. Setting: I Grid topology with L = 16 possible locations. I Cellular data rate: 300 Mbps (shared among L locations). I Wi-Fi data rate: 54 Mbps (for one location). I Cellular price > 0 and Wi-Fi price = 0. I Switching time = 1. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 63 / 147 Impact of the Number of Users Setting: Switching cost c switch = 1 and cellular price = US $3/Gbyte. 80 NSG OTSO Cellular−Only Average Payoff per MU 70 60 50 40 30 20 10 5 10 15 20 25 30 Total Number of MUs I Average payoff decreases with the level of contention. Cellular-only performs well under low traffic load. OTSO performs well under high traffic load and low c switch . Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 64 / 147 Impact of the Wi-Fi Availability Setting: I = 30 MUs and cellular price = US $6/Gbyte. 600 Average Payoff per MU 500 Class 1 Class 2 Class 3 400 300 200 100 0 50 250 450 650 850 Wi−Fi Data Rate (Mbps) Class 1: Can access all the networks all the time. Class 2: Can access the cellular network all the time, and a Wi-Fi hotspot 50% of the time. Class 3: Can only access the cellular network. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 65 / 147 Summary Network selection and data offloading: First study on the interactions of multiple heterogeneous MUs. Explicit modelling of user mobility, Wi-Fi availability, switching time and cost, and pricing. Network Selection Game Analysis: FIP ⇒ Convergence to a pure NE. Key Question: what about incomplete network information? Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 66 / 147 Delay-Aware Predictive Network Selection in Data Offloading Joint work with Haoran Yu, Man Hon Cheung (CUHK), Longbo Huang (Tsinghua) IEEE GLOBECOM 2014 Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 67 / 147 Network Selection with Cost-Delay Tradeoff An operator’s perspective: how to dynamically select networks for users to balance the long-term operation cost and traffic delay? I Carrier-grade WiFi Challenge: limited information on system randomness We consider two cases: 1 Only having current slot information 2 Having both current and predicted future information Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 68 / 147 System Model 6 5 4 7 3 2 1 8 9 Multiple networks, locations, and users (similar as before) I I Network availability is location-dependent Users randomly move across the locations with random traffic arrivals Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 69 / 147 System Model 6 5 4 7 3 2 1 8 9 Multiple networks, locations, and users (similar as before) I I Network availability is location-dependent Users randomly move across the locations with random traffic arrivals Ql (t): the amount of user l’ unserved traffic (queue length) Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 69 / 147 Case 1: Only Knowing Current Network Information Operator only knows current user locations and queue lengths I No statistic knowledge of users’ mobilities and traffic patterns Queue length based optimization based on Lyapunov method Intuition: I When Ql (t) is small, suspending service does not lead to severe delay. Strategy: wait till enter Wi-Fi area I When Ql (t) is large, suspending service incurs severe delay. Strategy: serve user l immediately even with a high operational cost Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 70 / 147 Delay-Aware Network Selection (DNS) Delay-Aware Network Selection (DNS) Algorithm At each time slot t, the operator: Chooses the network selection vector α (t) that solves L i h X Ql (t)rl (α (t)) + Vc (α (t)) minimize − l=1 variables αl (t) ∈ NSl (t) ∪ {0} , ∀l ∈ L. Updates the queueing vector Q (t + 1) accordingly. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 71 / 147 Performance of DNS Achieve [O (1/V ) , O (V )] cost-delay tradeoff (V : control parameter) Conclusion: The operation cost can be pushed arbitrarily close to the optimal value, but at the expense of an increase in the traffic delay. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 72 / 147 Case 2: Predictive Network Selection How to improve DNS with predicted future information? I When we know the statistics of user mobility and traffic patterns Solution: propose a novel frame-based Lyapunov optimization technique and design the GP-DNS algorithm Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 73 / 147 230 ion vecαm (τ ), 0.4 0.6 GP-DNS Algorithm 0.8 1 1.2 1.4 1.6 Average Queue Length 1.8 2 2.2 2.4 4 x 10 Fig. 3: Cost-Delay Tradeoff of DNS and GP-DNS. 16000 ctor that he same m) ; 39.0% 14000 DNS GP−DNS (T=5) GP−DNS (T=10) GP−DNS (T=25) Average Queue Length 12000 10000 40.5% 8000 40.3% 6000 40.4% 4000 2000 0 240 250 260 Average Operation Cost 270 easible Fig. Cost-Delay 4: Delay Tradeoff Reduction through Prediction. of DNS and GP-DNS This is updating rule and the fact that the number of feasible network grows selection vectorsimproves is finite. the Thecost-delay details aretradeoff given in [17]. Future information T . To I ropose We operator observe pursues that thean complexity of ofAlgorithm 1 iswith onlyT = 25 If the operation cost 250, GP-DNS P-DNS polynomial T . In size) particular, a concrete computation-friendly (prediction in window saves 40.5% traffic delay over DNS. exity is way of solving line 6 is also given in [17]. Jianwei Huang (CUHK) Data Offloading (Tutorial) V. Mobile N UMERICAL R ESULTS June 2015 74 / 147 Summary Prediction can significantly improve network performance. Need to carefully tradeoff algorithm complexity and network performance. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 75 / 147 Economics Issues for Wi-Fi Offloading Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 76 / 147 Bargaining-based Mobile Data Offloading Joint work with Lin Gao & Duozhe Li (CUHK) George Iosifidis & Leandros Tassiulas (Yale University) IEEE JSAC 2014 Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 77 / 147 Mobile Data Offloading BS2 MU13 BS1 AP1 MU11 MU21 AP2 MU24 AP4 MU14 MU32 AP3 BS3 MU33 MU31 One mobile network operator (MNO) offloads to multiple Access Point (APs). Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 78 / 147 Key Problems Efficiency: How to offload traffic efficiently? Fairness: How to share the benefit among the MNO and APs fairly? Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 79 / 147 Bargaining-based Solution Bargaining is useful for resolving situations where I Players have a common desirable to reach a mutual agreement. I Players have individual payoffs. I Allowing disagreement: no agreement may be forced on any player. I There is a conflict of interest among players about the agreement. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 80 / 147 An Illustrative Example Scenario: Player 1 sells a book to Player 2 at a price p I Problem: Two players bargain for the price p. Players’ payoffs: u1 = p, u2 = 1 − p. I I Assumption: the book is worth 1 to player 2. The set of feasible agreements: U = {(u1 , u2 )|u1 + u2 = 1} The disagreement: D = (d1 , d2 ) = (0, 0) I Assumption: the book is worth 0 to player 1. A bargaining solution is an outcome (v1 , v2 ) ∈ U ∪ D What will be a proper bargaining solution? Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 81 / 147 Nash Bargaining Theory Nash bargaining theory: An axiom-based theory I Pareto Efficiency I Symmetry I Invariant to Affine Transformations I Independence of Irrelevant Alternatives Nash bargaining solution I Unique solution that satisfies the Nash’s 4 axioms. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 82 / 147 Nash Bargaining Solution Nash Bargaining Solution (NBS) Nash bargaining solution solves the optimization problem: max (v1 − d1 ) · (v2 − d2 ) v1 ,v2 subject to (v1 , v2 ) ∈ U ∪ D v1 ≥ d1 , v2 ≥ d2 Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 83 / 147 Nash Bargaining Solution Nash Bargaining Solution (NBS) Nash bargaining solution solves the optimization problem: max (v1 − d1 ) · (v2 − d2 ) v1 ,v2 subject to (v1 , v2 ) ∈ U ∪ D v1 ≥ d1 , v2 ≥ d2 When (d1 , d2 ) = (0, 0): NBS is (v1 , v2 ) = (0.5, 0.5); When (d1 , d2 ) = (0, 0.4): NBS is (v1 , v2 ) = (0.3, 0.7); Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 83 / 147 System Model One Mobile Network Operator (MNO) I I Operating multiple macrocell base stations; Serving many mobile users (MUs); N Access Point (APs) I APs are geographically non-overlapping with each other; BS2 MU13 BS1 AP1 MU11 MU21 AP2 MU24 AP4 MU14 MU32 AP3 BS3 MU33 MU31 Example: N = 4 APs. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 84 / 147 Decisions Variables Variables: I I The traffic offloaded to each AP; The payment to each AP; Traffic Offloading Profile: x = (x1 , . . . , xN ) I xn : the traffic offloaded to AP n; Payment Profile: z = (z1 , . . . , zN ) I zn : the payment to AP n; Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 85 / 147 Payoffs MNO’s Payoff: net cost reduction U(x; z) = J(x) − I I N X zn n=1 J(x): the MNO’s benefit in terms of operational cost reduction PN n=1 zn : the MNO’s total payment to APs AP’s Payoff: net profit increase Vn (xn ; zn ) = −Qn (xn ) + zn I I Qn (xn ): the AP n’s cost due to offoloading zn : the AP n’s profit from serving the MNO Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 86 / 147 Social Welfare Social Welfare: sum of the MNO’s and all APs’ payoffs Ψ(x) = J(x) − I N X Qn (xn ) n=1 The payment between the MNO and each AP is internal transfer and is canceled out. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 87 / 147 Social Welfare Social Welfare: sum of the MNO’s and all APs’ payoffs Ψ(x) = J(x) − I N X Qn (xn ) n=1 The payment between the MNO and each AP is internal transfer and is canceled out. An efficient offloading decision maximizes the social welfare. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 87 / 147 Starting Point: One-to-One Bargaining One MNO and one AP One-to-One Bargaining Problem max U(x; z) · Vn (x; z) (x,z) s.t. U(x; z) ≥ U0 , V(x; z) ≥ V0 I I U0 = 0: the disagreement of the MNO; V0 = 0: the disagreement of the AP; Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 88 / 147 An Equivalent Formulation Social Welfare = MNO’s Payoff + AP’s Payoff Ψ(x) = U(x; z) + V(x; z) Define AP’s payoff as π = V(x; z) Then MNO’s payoff U(x; z) = Ψ(x) − π An Equivalent Bargaining max (Ψ(x) − π) · π (x,π) s.t. Ψ(x) − π ≥ 0, π ≥ 0 Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 89 / 147 An Equivalent Formulation Social Welfare = MNO’s Payoff + AP’s Payoff Ψ(x) = U(x; z) + V(x; z) Define AP’s payoff as π = V(x; z) Then MNO’s payoff U(x; z) = Ψ(x) − π An Equivalent Bargaining max (Ψ(x) − π) · π (x,π) s.t. Ψ(x) − π ≥ 0, π ≥ 0 Solving this allows us to see the relationship between AP’s payoff (π) and the social welfare (Ψ(x)) clearly. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 89 / 147 NBS for One-to-One Bargaining One-to-One NBS The NBS (xn∗ , πn∗ ) for the one-to-one bargaining is xn∗ = xno = arg maxxn Ψ(xn ), and πn∗ = 1 2 · Ψ(xno ) xno = arg maxxn Ψ(xn ): Bargaining solution maximizes social welfare; πn∗ = 1 2 · Ψ(xno ): AP gets half of the generated social welfare; U = Ψ(xno ) − πn∗ = social welfare; Jianwei Huang (CUHK) 1 2 · Ψ(xno ): the MNO gets half of the generated Mobile Data Offloading (Tutorial) June 2015 90 / 147 General One-to-Many Bargaining Sequential Bargaining: MNO bargains with all APs sequentially (in a predefined order). Concurrent Bargaining: MNO bargains with all APs concurrently. Completed Bargaining On-going Bargaining Future Bargaining MNO MNO AP 1 AP 5 AP 2 AP 3 (a) Sequential Bargaining Jianwei Huang (CUHK) AP 1 AP 5 AP 2 AP 4 AP 3 AP 4 (b) Concurrent Bargaining Mobile Data Offloading (Tutorial) June 2015 91 / 147 AP Grouping APs may form groups to bargain with the MNO jointly. How will grouping affect the payoffs of MNO and APs’ ? Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 92 / 147 Sequential Bargaining Bargain from user 1 to user N Sequential Nash Bargaining Solution (NBS) {x ∗ , π ∗ } = {(xn∗ , πn∗ )}n∈N Sequential NBS The NBS {x ∗ , π ∗ } under the sequential bargaining is xn∗ = xno , πn∗ = I I ¯n ∆ , ∀n = 1, ..., N 2 x o = arg maxx Ψ(x): bargaining solution maximizes social welfare; ¯ n : the virtual marginal social welfare generated by AP n; ∆ Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 93 / 147 Virtual Marginal Social Welfare Virtual Marginal Social Welfare generated by AP n ¯n = ∆ 1 X In+1 =0 1 X ∆n (In+1 ; ...; IN ) ... 2N−n IN =0 Represents average marginal social welfare by AP n, assuming I I MNO has reached agreements with all APs 1, ...., n − 1 (before n); MNO will reach agreement with each AP in {n + 1, ..., N} (after n) with a probability of 0.5. Consider all possibilities (through the values of In+1 to IN ) Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 94 / 147 Impact of AP Ordering in Sequential Bargaining AP: Early-Mover Advantage Under the sequential bargaining, an AP will obtain a higher payoff if it bargains with the MNO earlier. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 95 / 147 Impact of AP Ordering in Sequential Bargaining AP: Early-Mover Advantage Under the sequential bargaining, an AP will obtain a higher payoff if it bargains with the MNO earlier. MNO: Invariance to the Order Under the sequential bargaining, the bargaining order of APs does not affect the MNO’s payoff. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 95 / 147 Impact of Grouping in Sequential Bargaining Intra-Grouping Benefit Under the sequential bargaining, group bargaining always benefits the APs in the group. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 96 / 147 Impact of Grouping in Sequential Bargaining Intra-Grouping Benefit Under the sequential bargaining, group bargaining always benefits the APs in the group. Inter-Grouping Benefit Under the sequential bargaining, group bargaining Improves the payoffs of all APs bargaining before the group, No impact on the APs bargaining after the group. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 96 / 147 Concurrent Bargaining Bargain concurrently with user 1 to user N Concurrent Nash Bargaining Solution (NBS) {x ∗ , π ∗ } = {(xn∗ , πn∗ )}n∈N Concurrent NBS The NBS {x ∗ , π ∗ } under the concurrent bargaining is xn∗ = xno , πn∗ = I I en ∆ , ∀n = 1, ..., N 2 x o = arg maxx Ψ(x): bargaining solution maximizes social welfare; e n = Ψ(x ∗ , x ∗ ) − Ψ(x ∗ , 0): the actual marginal social welfare ∆ −n n −n generated by AP n; Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 97 / 147 Property of Concurrent NBS Concurrently Moving Tragedy The payoff of each AP under the concurrent bargaining equals to the worst-case payoff that it can achieve under the sequential bargaining (as the last AP). Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 98 / 147 Impact of Grouping in Concurrent Bargaining Intra-Grouping Benefit Under the concurrent bargaining, grouping of APs always benefits APs in the group. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 99 / 147 Impact of Grouping in Concurrent Bargaining Intra-Grouping Benefit Under the concurrent bargaining, grouping of APs always benefits APs in the group. No Inter-Grouping Benefit Under the concurrent bargaining, grouping of APs does not affect the APs not in the group. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 99 / 147 Simulations S equential Bargaining 16 h5i denotes the merged group: {5,6,7,8,9,10}. 14 12 Payoff of APO Payoffs of APOs Concurrent Bargaining 10 h5i 8 6 4 4 2 2 3 1 0 1 2 3 4 5 6 7 Groupin g Structure 8 9 10 C Bargaining AP 1 16 16 14 14 12 12 AP 5 AP 6 10 10 AP 7 8 8 6 AP 9 Group 6 4 4 2 2 0 1 2 3 4 5 6 7 Groupi ng Structure 8 AP 2 AP 3 AP 4 AP 8 0 9 10 10 123456789 Group Structure Left figure: Payoffs of APs under sequential bargaining I I Early-mover advantage Positive intra-grouping effect, positive inter-grouping effect. Right figure: Payoffs of APs under concurrent bargaining I I Concurrently moving tragedy Positive intra-grouping effect, no inter-grouping effect. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 100 / 147 Summary Study an offloading market with one MNO and multiple APs. Propose two one-to-many bargaining protocols. Analyze the impact of bargaining protocols and grouping structure. Next Step: What happen if there are multiple MNOs? Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 101 / 147 An Iterative Double Auction Mechanism for Mobile Data Offloading Joint work with Lin Gao (CUHK) George Iosifidis & Leandros Tassiulas (University of Thessaly) IEEE WiOpt 2013 (Best Paper Award), IEEE ToN 2015 Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 102 / 147 Mobile Data Offloading Market BS2 MU13 BS1 AP1 MU11 MU21 AP2 MU24 AP4 MU14 MU32 AP3 BS3 MU33 MU31 Multiple MNOs and multiple APs Each MNO can lease multiple APs Each AP can offload for multiple MNOs Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 103 / 147 Key Problems From the MNO’s Perspective: How much traffic should each MNO offload to each AP, and how much to pay? From the AP owner’s Perspective: How much traffic should each AP offload for each MNO, and how much to charge? Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 104 / 147 System Model Each MNO is represented by one Base Station (BS) I , {1, ..., I }: the set of BSs N , {1, ..., N}: the set of APs BS2 MU13 BS1 AP1 MU11 MU21 AP2 MU24 AP4 MU14 MU32 AP3 BS3 MU33 MU31 Example: I = {1, 2, 3} and N = {1, 2, 3, 4}. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 105 / 147 For Each BS i ∈ I xin : offloading request to AP n x i , (xin , ∀n ∈ N ): offload request vector to all APs Ji (x i ): the utility (cost reduction) function of BS i I Positive, increasing, and jointly strictly concave I AP-specific: depending on x i , not just the total traffic I Intuition: offloading cell-edge traffic will lead to more cost reduction Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) P n∈I xin June 2015 106 / 147 For Each AP n ∈ N yin : offload admission for BS i y n , (yin , ∀i ∈ I): offload admission vector for all BSs; Qn (y n ): the cost function of AP n I Positive, increasing, and jointly strictly convex. Cn : capacity constraint Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 107 / 147 Market Outcome A feasible market outcome is where BSs and APs reach an agreement: xin = yin , Jianwei Huang (CUHK) ∀n ∈ N , ∀i ∈ I. Mobile Data Offloading (Tutorial) June 2015 108 / 147 A Benchmark Problem Social Welfare Maximization (Efficiency) maximize X i∈I subject to variables Ji (x i ) − X Qn (y n ) ......Social Welfare n∈N (i) P (ii) xin = yin , ∀n ∈ N , i ∈ I, ......Feasibility i∈I yin ≤ Cn , ∀n ∈ N , ......Capacity constraint x i , y n , ∀n, ∀i. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 109 / 147 Socially Optimal Solution Socially Optimal KKT ∂Ji (x i ) ∂Qn (y in ) − µin = 0, (A2) : − µin + λn = 0, ∂xin ∂yin X (A3) : λn · yin − Cn = 0, (A4) : µin · (yin − xin ) = 0, (A1) : i∈I (A5) : xin = yin . Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 110 / 147 Challenge: Information Asymmetry The utility function Ji (x i ) is the private information of BS i: I Not known by other BSs, APs, and any market coordinator The cost function Qn (y n ) is the private information of AP n: I Not known by other APs, BSs, and any market coordinator It is difficult to achieve efficiency (social welfare maximization). I Conflict of interests: BSs want to offload more traffic with less payment, while APs want to admit less traffic with more payment. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 111 / 147 A First Throught A traditional approach: Two-sided Market → Double Auction I A market controller or broker acts as the auctioneer; I BSs and APs act as bidders; I The auctioneer decides the allocation and payment rules such that all bidders truthfully disclose their private information. Direct application of double auction does not work here I Every bidder may have infinite amount of private information due to the continuity of the utility/cost function. I According to [Myerson’1983], there does not exist a double auction that possesses all the following properties: F F F F Efficiency Individually rationality Incentive compatibility Budget balanced Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 112 / 147 Our Approach: Iterative Double Auction (IDA) Our proposed approach: Iterative Double Auction I Conducts one double auction in each round. Next round Auctioneer Auctioneer Payment rule Allocation rule Updating Payment rule Allocation rule Bidder Bidder Disclore all private information Signaling his private information Fig. Double Auction vs Iterative Double Auction Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 113 / 147 How IDA Works? Key Steps Step 1 (Rules): The auctioneer broadcasts the payment rule hi (·) to every BS i and the reimbursement rule ln (·) to every AP n; Step 2 (Bidding): Every BS i determines his bids pin to every AP n. Every AP n determines his bid αin to every BS i. Both aim at maximizing their respective objectives. Step 3 (Allocation): The auctioneer determines the allocation xin and yin between every BS i and AP n, aiming at maximizing a public auxiliary objective function: XX αin 2 W (x, y ) , pin log xin − y . 2 in i∈I n∈N Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 114 / 147 How IDA Works? Key Steps Step 1 (Rules): The auctioneer broadcasts the payment rule hi (·) to every BS i and the reimbursement rule ln (·) to every AP n; Step 2 (Bidding): Every BS i determines his bids pin to every AP n. Every AP n determines his bid αin to every BS i. Both aim at maximizing their respective objectives. Step 3 (Allocation): The auctioneer determines the allocation xin and yin between every BS i and AP n, aiming at maximizing a public auxiliary objective function: XX αin 2 W (x, y ) , pin log xin − y . 2 in i∈I n∈N Question: What are the conditions that guarantee us to achieve the social optimal allocation? Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 114 / 147 Analysis of IDA Step 3: Allocation Auctioneer Optimal KKT Jianwei Huang (CUHK) () Socially Optimal KKT Mobile Data Offloading (Tutorial) June 2015 115 / 147 Analysis of IDA Step 2: Bidding Individual Optimal Bids () Socially Optimal Bids * Step 3: Allocation Auctioneer Optimal KKT Jianwei Huang (CUHK) () Socially Optimal KKT Mobile Data Offloading (Tutorial) June 2015 115 / 147 Analysis of IDA Step 1: Rules Socially Optimal Rules * Step 2: Bidding Individual Optimal Bids () Socially Optimal Bids * Step 3: Allocation Auctioneer Optimal KKT Jianwei Huang (CUHK) () Socially Optimal KKT Mobile Data Offloading (Tutorial) June 2015 115 / 147 Social to APs and BS execute IDA al 100 IDA - Convergence 50 Convergence of IDA 00 20 40 60 80 100 120 140 Step − t The IDA algorithm converges to the socially optimal solution. Fig. 3. Evolution of social welfare produced by the IDA. 0.4 0.2 Gap y − x 0 −0.2 BS 1, AP 1: y11 −x11 BS 1, AP 2: y21−x12 −0.4 BS 2, AP 1: y −x 12 −0.6 21 BS 2, AP 2: y22−x22 −0.8 −1 −1.2 0 20 40 60 80 100 120 Step − t of of thethegap xin and y(x) i.e.,admitted yin − xdata in , and in . (y). Fig. Evolution 4. Evolution gapbetween between requested [1] Cisco, “Cisco Forecast Upda [2] Bloomberg: “A [3] FemtoForum, “ [4] AT&T Press R Additional Cit [5] BT Wifi Press Customers”, Ju [6] Cisco, “Makin [7] Republic Wire [8] Spectrum bridg [9] R. B. Myerso Bilateral Tradi [10] R. P. McAfe Economic The [11] P. Maille, B. Pricing of Inte [12] D. P. Paloma Network Utilit 1451, 2006. [13] F. P. Kelly, A Networks: Sha of Oper. Res. S [14] L. Johansen, “ [15] K. Lee, I. Rh How Much Ca [16] N. Ristanovic Efficient Offlo ρ11 = 0.74. Finally, the payments of the BSs 1, 2 and 5 are p21 =Mobile 6.29,Data andOffloading p51 = (Tutorial) 6.63, respectively. Notice June 2015 11 = 7.3, Jianwei Huang p (CUHK) 116 / 147 Properties of IDA Properties of IDA Efficient I The IDA mechanism achieves the social welfare maximization; Weakly Budget Balanced I I The auctioneer does not lose money by organizing an IDA; If there is no capacity constraint, the auctioneer neither lose money nor gain money by organizing an IDA (strongly budget balanced); Incentive Compatible I All bidders (price-taking) act in a truthful manner; Individually Rational I All bidders achieve non-negative utilities. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 117 / 147 Summary Consider multiple MNOs offloading to multiple APs. Iterative double auction mechanism that satisfies all desirable properties. Next Step: Do we always offload traffic from cellular to Wi-Fi? Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 118 / 147 Enabling Crowd-Sourced Mobile Internet Access Joint work with Lin Gao (CUHK) George Iosifidis & Leandros Tassiulas (Yale University) IEEE INFOCOM 2014 Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 119 / 147 Imbalance of Mobile Internet Access Different users have different access technologies and access speeds: 3G/4G, femtocell, Wi-Fi. Different networks have different congestion levels even at the same time and location. How to effectively take advantage of and integrate heterogeneous network access capabilities? Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 120 / 147 Crowd-Sourced Mobile Internet Access 3G/4G Wi-Fi Femtocell Share the best mobile internet connection(s) among users. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 121 / 147 Commercial Cases Open Garden (http://opengarden.com) M-87 (http://www.m-87.com/) Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 122 / 147 Key Problems How to achieve an efficient and fair network resource allocation? I I Who will download data for whom, and how much? Who will route data from each host to each client, and how much? How to encourage the user participation and cooperation? I how to compensate the hosts and the relays for their efforts? Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 123 / 147 Crowd-Sourced Mobile Internet Access Internet 3G Base Station WiFi Router 4G 3G 4G WiFi 4G Base Station WiFi Bluetooth Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 124 / 147 Crowd-Sourced Mobile Internet Access Internet 3G Base Station WiFi Router 4G 3G ta ta Relay Gateway Data (Host) Da Da 4G WiFi Bluetooth Data Gateway (Host) Data WiFi 4G Base Station Client Client Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 125 / 147 Key Features Three types of roles I Host (Gateway): Downloading data from Internet I Relay: Forwarding data for others I Client: Consuming data I A mobile user may have multiple roles Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 126 / 147 Key Features Three types of roles I Host (Gateway): Downloading data from Internet I Relay: Forwarding data for others I Client: Consuming data I A mobile user may have multiple roles User-provided networking I Mobile users can access internet through the hosting of other users. Multi-hop accessing I Mobile users can access internet through the relay of multiple devices. Access bonding I Mobile users can access internet through multiple access links. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 126 / 147 System Model A set of mobile users: I = {1, 2, ..., I } For each user i ∈ I: Internet Ci ei pi WiFi Bluetooth I Cij, e s ij User i Cji, e r ij User j ci , cij , cji , j ∈ I: link capacity; I ei , eijs , eijr , j ∈ I: unit energy consumption; I pi : usage-based pricing for accessing Internet. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 127 / 147 Client Model When user i ∈ I is a client. Internet y1(i) User 1 User 2 (i) y1 I I I yi(i) y2(i) y2 (i) User i (client) ... yI(i) yj(i) User j (i) yj : the data downloaded via host j for client i; P (i) y (i) = j∈I yj : the total data consumed by client i; Ui y (i) : the utility function of client i. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 128 / 147 Host Model When user i ∈ I is a host (gateway). Internet yi(1) yi(2) ... yi(I) yi(2) yi(1) User 2 User 1 yi(3) User i (host) yi(j) User j yi(4) User 4 User 3 I I I I (j) yi : the data downloaded via host i for a client j; P (j) yi = j∈I yi : the total data downloaded via host i; ei · yi : the total energy consumption for downloading data; pi · yi : the total payment for downloading data; Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 129 / 147 Host Model When user i ∈ I is a host (gateway). Internet yi(1) yi(2) ... yi(I) yi(2) yi(1) User 2 User 1 yi(3) User i (host) yi(j) User j yi(4) User 4 User 3 I I I I I (j) yi : the data downloaded via host i for a client j; P (j) yi = j∈I yi : the total data downloaded via host i; ei · yi : the total energy consumption for downloading data; pi · yi : the total payment for downloading data; Downloading capacity constraint: yi ≤ ci . Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 129 / 147 Relay Model When user i ∈ I is a relay. Internet xij (n) , n=1,...,I User i (relay) (n) n∈I : I xij, I ejir · P I eijs · P xji (n) , n=1,...,I User j the data relayed from user i to user j, for client n; (n) n xji : total energy consumption for receiving data from user j; (n) n xij : total energy consumption for sending data to user j. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 130 / 147 Relay Model When user i ∈ I is a relay. Internet xij (n) , n=1,...,I User i (relay) (n) n∈I : I xij, I ejir · P I eijs · P I I xji (n) , n=1,...,I User j the data relayed from user i to user j, for client n; (n) n xji : total energy consumption for receiving data from user j; (n) n xij : total energy consumption for sending data to user j. P (n) P (n) Relay capacity constraints: n xij ≤ cij , n xji ≤ cji P (n) P (n) (n) Flow balance constraint: j xji + yi = j xij , n ∈ I Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 130 / 147 User Payoff Payoff of each user i ∈ I: Ji (x i , y i ) = Ui − Pi − Ei (n) I y i = {yi }n∈I : Downloading matrix; I x i = {xij }j,n∈I : Relaying matrix; I Ui : Utility of user i (as a client); I Pi : Total payment of user i (as a host for internet access); I Ei : Total energy consumption of user i (as a host and/or relay); (n) To maximize the payoff, each user only wants to be a client, but not as a host or relay. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 131 / 147 Our Goal Mechanism design to address incentive, efficiency, and fairness issues I Encouraging the user participation and cooperation; I Achieving an efficient and fair network resource allocation. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 132 / 147 Incentive Design Challenges Users may not want to participate in the crowd-sourced system I For example, those without a demand for Internet access; Users may not want to download or relay data for others I For example, user i may not want to download data for user 4. Internet yi(1) yi(2) ... yi(I) yi(2) yi(1) User 2 User 1 yi(3) User i (host) X yi(j) User j User 4 User 3 Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 133 / 147 Solution: Virtual Currency Key idea: User pays certain virtual currency to those who send data to him (I give you money, you give me data). User j xji(1) ... xji(n) ... xji(I) zji(1) ... zji(n) ... zji(I) User i (n) zji : the virtual price that user i pays j for receiving data (of client n); P (n) n zji (n) · xji : the total virtual money that user i pays j Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 134 / 147 Modified Payoff with Virtual Currency Modified payoff of each user i ∈ I: Ji (x i , y i , z i ) = Ui − Pi − Ei + Vi I I (n) z i = {zij }j,n∈I : Virtual payment matrix; Vi : Total virtual currency evaluation of user i; Modified payoff maximization takes care of incentive issues. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 135 / 147 Efficiency and Fairness Issues How to achieve an efficient and fair network resource allocation? I Efficiency: The aggregate payoff of all users is maximised. I Fairness: Every user achieves a satisfactory payoff; Our Solution: Nash Bargaining Solution Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 136 / 147 Nash Bargaining Solution Nash Bargaining Problem (NBP) max x i ,y i ,z i ,∀i Πi∈I (Ji (x i , y i , z i ) − Ji0 ) s.t., (a) Ji ≥ Ji0 (Ji0 : disagreement point) (b) Capacity constraints; (c) Flow balance constraint; (d) Virtual current budget constraint. The NBP problem has a unique optimal solution. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 137 / 147 Nash Bargaining Implementation Centralized Implementation I A central control node collects all the required network information, and computes the Nash bargaining solution. Decentralized Implementation I Iterative updating: Users update their individual decisions sequentially and repeatedly, and signals to neighbors until convergence. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 138 / 147 Simulation An example with 6 nodes I I Blue Bar: Downloading/relaying data; Red Bar: Consuming data; 2 1.242 2 1.242 0.0 1 5 1.242 1.242 2.295 2.295 1.143 2.295 2.295 0.059 1.880 0.059 1 2.295 6 2.295 0.059 6 4 51 2.245 0.059 0.059 3 0.059 3 0.059 Standalone (Independent) Standalone (Independent) Operation Operation Left: Independent Operation. Jianwei Huang (CUHK) 2 1.240 1.240 0.0 2.295 2.295 0.097 46 0.337 2 0.239 0.097 5 1 2.247 0.337 0.239 1.1431.704 0.304 1.880 0.366 2.245 6 0.350 3 0.410 2.247 0.304 0.350 4 0.366 0.046 0.059 5 1.704 4 0.046 3 0.410 0.044 0.044 UPN Bargained Operation UPN Bargained Operation Right: Crowd-sourced Operation. Mobile Data Offloading (Tutorial) June 2015 139 / 147 Summary We study the crowd-sourced mobile internet access system, in particular, we answer I I How to achieve an efficient and fair network resource allocation? How to encourage the user participation and cooperation? We propose a Nash bargaining solution with virtual currency, which addresses the incentive, efficiency, and fairness issues. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 140 / 147 Going Beyond Offloading Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 141 / 147 Blurring the Cellular/Wi-Fi Boundary Mi-Fi turn cellular signal into Wi-Fi signals Social bandwidth trading: Karma Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 142 / 147 Why Wi-Fi Complements Cellular? Wi-Fi can be the primary access technology with cellular as a coverage supplement Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 143 / 147 Where Will Wi-Fi Go in The Future? Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 144 / 147 Where Will Wi-Fi Go in The Future? It’s Up to You and Me. Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 144 / 147 Related Publications G. Iosifidis, L. Gao, J. Huang, and L. Tassiulas, “An Double Auction Mechanism for Mobile Data Offloading Markets,” IEEE/ACM Transaction on Networking, 2015 (conference version received Best Paper Award in IEEE WiOPT 2013) M. Cheung and J. Huang, “DAWN: Delay-Aware Wi-Fi Offloading and Network Selection,” IEEE Journal on Selected Areas in Communications, June 2015 L. Gao, G. Iosifidis, J. Huang, L. Tassiulas, and D. Li, “Bargaining-based Mobile Data Offloading,” IEEE Journal on Selected Areas in Communications, June 2014 G. Iosifidis, L. Gao, J. Huang, and L. Tassiulas, “Enabling Crowd-Sourced Mobile Internet Access,” IEEE INFOCOM, May 2014 L. Gao, G. Iosifidis, J. Huang, and L. Tassiulas, “Hybrid Data Pricing for Network-Assisted User-Provided Connectivity,” IEEE INFOCOM, May 2014 H. Yu, M. Cheung, L. Huang, and J. Huang, “Predictive Delay-Aware Network Selection in Data Offloading,” IEEE GLOBECOM, December, 2014 M. Cheung, R. Southwell, and J. Huang, “Congestion-Aware Network Selection and Data Offloading” (invited), CISS, March 2014 Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 145 / 147 More Information http://ncel.ie.cuhk.edu.hk/content/wifi-data-offloading http://ncel.ie.cuhk.edu.hk/content/user-provided-networks Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 146 / 147 Contact Google “Jianwei Huang” http://jianwei.ie.cuhk.edu.hk/ Jianwei Huang (CUHK) Mobile Data Offloading (Tutorial) June 2015 147 / 147