GreenBS: Design Energy Efficient Cellular Network Infrastructure Chunyi Peng University of California, Los Angeles Joint work with Suk-Bok Lee, Songwu Lu, Haiyun Luo, Hewu Li IBM Student Workshop for Frontiers of Cloud Computing 2011 Hawthorne, NY UCLA WiNG Surging Energy Consumption in ICT ICT: information and communication technology systems 3% of world-wide electricity consumed by ICT systems CO2 emission, comparable to global air travel Operation cost from a huge power bill Rising energy consumption at 16-20%/year E.g., data centers, internet infrastructure, cellular networks, etc Moore’s law: 2x power every 4~5 years by 2030 Our focus is energy efficiency of cellular networks One large scale of ICT systems Applicable to other ICT systems IBM F2C2 2011 C Peng (UCLA) 2 UCLA WiNG Energy Consumption in Cellular Networks 0.1w X 5B = 0.5GW <10% (~1%) Mobile Terminals 1~3kw X 4M = 8GW 10kw X 10K = 0.1GW >90% (~99%) Cellular Infrastructure ~80% at BS The key to green cellular network is on BS network Source: Nokia Siemens Networks (NSN) IBM F2C2 2011 C Peng (UCLA) 3 UCLA WiNG Roadmap Overview Problem and root cause Existing solutions Our solution Characterizing 3G dynamics Exploiting dynamics in design Working with 3G standards Evaluation Summary IBM F2C2 2011 and Insights C Peng (UCLA) 4 UCLA WiNG Case Study in a Regional 3G Network Power (Kw) Current Ideal Load: (#link in 15min) Power-load curve in a big city with 177 BSes (3G UMTS) Non-energy-proportionality (Non-EP) to traffic load IBM F2C2 2011 C Peng (UCLA) 5 UCLA WiNG Root Cause for Energy Inefficiency Each BS is non-EP 2000 Power (w) l500 l000 500 load Large portion of consumed energy even @ zero traffic load as long as the BS is on. IBM F2C2 2011 C Peng (UCLA) 6 UCLA WiNG Root Cause for Energy Inefficiency Traffic is highly dynamic Fluctuate over time Be uneven at BSes BS traffic load (norm) 100 BS1 BS2 BS3 BS4 80 60 40 20 0 08/30 08/31 09/01 09/02 09/03 09/04 09/05 09/06 Low usage at night Date Large energy overhead at light traffic => non-EP. Turn off BS completely to save more energy! IBM F2C2 2011 C Peng (UCLA) 7 UCLA WiNG Goals and Challenges 1. System-wide energy proportionality (EP) How to design EP network with non-EP BS components? 2. Negligible performance degradation How to meet location-dep. coverage & capacity requirements ? 3. 3G standard compliance How to support energy efficiency w/o changing 3G standard? IBM F2C2 2011 C Peng (UCLA) 8 UCLA WiNG Existing Solutions Optimization-based approach subject to C1,C2… constraints Component-based approach Practical issues unaddressed Theoretical analysis only e.g., on cooling, power amplifier No system-wide solution Complement our approach Clean slate design e.g., C-RAN Re-architect the 3G infrastructure Communication and computation intensive IBM F2C2 2011 C Peng (UCLA) 9 UCLA WiNG Our Solution Characterizing multi-dimensional dynamics Exploiting dynamics in design Working with 3G standards Evaluation IBM F2C2 2011 C Peng (UCLA) 10 UCLA WiNG Temporal Dynamics is Pervasive Low Peak-to-idle traffic is > 5 at 40~80% BSes BS traffic load (norm) 100 BS1 BS2 BS3 BS4 80 60 40 20 0 08/30 08/31 09/01 09/02 09/03 09/04 09/05 09/06 100 Percentage (> ratio) average utilization under dynamic load Region 1 Region 2 Region 3 Region 4 80 60 40 20 0 Date 0 4 8 12 16 20 Ratio of peak-to-idle traffic load Large saving potential for quiet hours IBM F2C2 2011 C Peng (UCLA) 11 UCLA WiNG Temporal Dynamics is Stable Temporal pattern is near-term stable Traffic at each BS is quite stable on a daily basis Autocorrelation with 24-hour lag is >0.92 at 70% BSes Day-to-day variation (|Curr – Prev|/Prev) is <0.2 at 70% BSes Region 1 Region 2 Region 3 Region 4 >70% BS 0.92 0.93 0.94 0.94 >90% BS 0.83 0.83 0.90 Autocorrelation with 24-hour-lag 0.90 Traffic is predictable. Make a case for traffic profiling IBM F2C2 2011 C Peng (UCLA) 12 UCLA WiNG Spatial Dynamics Deployment varies at locations Dense in big cities 20+ neighbor (<1KM) 100 90 80 CDF (%) 70 60 50 40 30 Region Region Region Region 20 10 0 0 5 10 15 20 25 30 1 2 3 4 35 40 Num. of neighbor BSs within 1 Km Rich BS redundancy ensures coverage. IBM F2C2 2011 C Peng (UCLA) 13 UCLA WiNG Spatio-temporal Dynamics Traffic Peak hours are different Multiplexing gain ~ 2 at peak hours 0.3 is also diverse at various locations Lower bound for the ratio of capacity to traffic Subregion A (Residence) 5 Subregion B (Business) Region Region Region Region 0.25 4 TM Gain Probablity 0.2 0.15 0.1 0.05 0 1 2 3 4 3 2 1 10AM12PM2PM 4PM 6PM 8PM10PM 10AM12PM2PM 4PM 6PM 8PM10PM Peak hour distributions in two sub regions 08:00 10:00 12:00 14:00 16:00 18:00 20:00 22:00 Multiplexing gain: sum(maxTraffic)/sum(traffic) Large saving potential even at peak hours IBM F2C2 2011 C Peng (UCLA) 14 UCLA WiNG Roadmap Overview Problem and root cause Existing solutions Our solution Characterizing 3G dynamics Exploiting dynamics in design Working with 3G standards Evaluation Summary IBM F2C2 2011 and Insights C Peng (UCLA) 15 UCLA WiNG Issue I: How to Satisfy location-dependent Coverage & Capacity Constraints? Once a BS turns off, clients in its original coverage should still be covered ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✔ ✔ ✔ Even if the total capacity is enough, it may fail to serve mobile clients due to coverage issue. ⇒ provide location-dependent capacity IBM F2C2 2011 C Peng (UCLA) 16 UCLA WiNG Solution I: Building Virtual Grids 1. Divide into BS virtual grids to decouple coverage constraints 2. Within a Vgrid, turn on/off BSes s.t. cap >= load ✗ ✗ ✔i j ri + d(i,j) < Ri rj + d(i,j) < Rj IBM F2C2 2011 ✗ ✗ ✔ ✗ ✔ ✗ C Peng (UCLA) ✗ ✗ ✗ ✔ ✗ 17 UCLA WiNG Issue II: How to Estimate Traffic Load? Q1: At what time scale is traffic load predictable? How often to estimate? Exploit near periodicity over consecutive time-of-the-day Q2: What to estimate? Instantaneous traffic load vs. traffic upper-envelope? Tradeoff between accuracy and efficiency. IBM F2C2 2011 C Peng (UCLA) 18 UCLA WiNG Solution II: Profiling Traffic Envelope 1. Traffic aggregation Sum Output 2. Discrete 24 intervals Estimate S, D, EV Stat 3. Estimate and profile update IBM F2C2 2011 C Peng (UCLA) 19 UCLA WiNG Issue III: How to Minimize On/Off Switches? ✗ ✔ ✔ ✗ 8:00 ✔ ✔ ✗ ✗ ✔ ✗ ✔ ✔ 10:00 12:00 ✗ ✔ ✔ ✗ 14:00 What if frequent On/Off switches happen? Frequent on/off switching is undesirable Large ramp-up time when on Reduced lifetime for cooling and other subsystems How often to switch on/off? Over 24-hour period, consistent with traffic characteristics IBM F2C2 2011 C Peng (UCLA) 20 UCLA WiNG Solution III: Smooth Switches Monotonically increasing ON from idle peak Monotonic OFF from peak idle 1) Find Smax for peak hours 3) Find St when traffic ✔ ✗ 2) Find Smin for idle hours (Smin ≤ Smax) At most ONE on/off switch per BS per 24 hours IBM F2C2 2011 C Peng (UCLA) 21 UCLA WiNG Roadmap Overview Problem and root cause Existing solutions Our solution Characterizing 3G dynamics Exploiting dynamics in design Working with 3G standards Evaluation Summary IBM F2C2 2011 and Insights C Peng (UCLA) 22 UCLA WiNG Working with 3G Standard Expand/shrink coverage at ON Bses Cell breathing technique When neighbor BSes turn OFF/ON 1 Migrate 2 OFF 3 1 2 3 clients from OFF BSes to ON BSes Before they turn off Leverage handover procedures 1 Coordinate 2 2 2 OFF 3 1 2 3 BSes at RNC via Iu-b interface Information collector and distributor IBM F2C2 2011 C Peng (UCLA) 23 UCLA WiNG Roadmap Overview Problem and root cause Existing solutions Our solution Characterizing 3G dynamics Exploiting dynamics in design Working with 3G standards Evaluation Summary IBM F2C2 2011 and Insights C Peng (UCLA) 24 UCLA WiNG Energy Saving in Four Regions Use two-month real traces in four regional 3G networks Region 1 Region 2 Region 3 Region 4 Eold (K. kwh) 9.81 2.63 8.58 9.18 Eour (K.kwh) 4.64 1.4 5.94 7.03 E Gain (%) 52.7% 46.6% 30.8% 23.4% missRatio 6.7e-4 #BS(weekday) 34~97 80.0% Spatial Dynamics 7.9e-4 8.2e-4 Temporal 8~32 79~122 Dynamics Energy saving gain 1.9e-5 104~142 60.0% Region 1 Region 2 40.0% Region 3 Region 4 20.0% 0.0% allsaving min-weekday max-day min-end IBM F2C2 2011 C Peng (UCLA) max-end 25 UCLA WiNG Summary It is feasible to build a practical solution to “green cellular infrastructure” Especially in the big cities with dense BS deployment Especially at late evenings to early dawn with light traffic Build EP system using non-EP components Exploiting inherent dynamics in time and space IBM F2C2 2011 C Peng (UCLA) 26 UCLA WiNG Insight Build an EP data center Dynamic traffic over time and space Exploit traffic dynamics to reduce the active # of hosts Easier without strict coverage constraints Flexibility to coordinate server clusters Management and monitoring functions IBM F2C2 2011 C Peng (UCLA) 27 UCLA WiNG THANK YOU! Any Questions? More details in “Traffic-Driven Power Saving in Operational 3G Networks”, ACM Mobicom’11. http://www.cs.ucla.edu/~chunyip/proj.html IBM F2C2 2011 C Peng (UCLA) 28