Leveraging Renewable Energy in Data Centers: Present and Future Ricardo Bianchini Department of Computer Science Collaborators: Josep L. Berral, Inigo Goiri, Jordi Guitart, Md. Haque, William Katsak, Kien Le, Thu D. Nguyen, Jordi Torres
Billion KWh/year Billion KWh/year Motivation Data centers = machine rooms to giant warehouses Consume massive amounts of energy (electricity) 90 60 30 0 2000 2005 2010 Electricity consumption of US DCs [JK 11] 2% 1.5% 270 240 210 180 150 120 90 60 30 0 2000 2005 2010 Electricity consumption of WW DCs [JK 11]
MMT/year Motivation Electricity comes mostly from burning fossil fuels 100% 80% 60% 40% Others Renewables Nuclear Natural Gas 120 116 112 108 35th 34th 20% Coal 104 0% US World 100 Nigeria Data Centers Czech Rep. Electricity sources in US & WW [DOE 10] CO 2 of world-wide DCs [Mankoff 08] Can we use renewables to reduce this footprint?
Outline DC energy usage and carbon footprint Reducing carbon with renewables: 2 approaches Our target and research challenges Software for leveraging solar energy Parasol: our solar micro-data center Current and future works Conclusions
Greening DCs: Grid-centric approach Pump renewables into the grid Pros: If the grid is available, power is available DC operator need not worry about renewable plants Plants can be placed at the best possible locations Cons: Energy losses of ~15% [IEC 07] Dependence on the power grid or diesel generators Example: Google buys wind power from NextEra
Greening DCs: Co-location approaches (1) Build DC near a renewable plant or (2) self-generate Pros: Reduced energy losses: ~5% No dependence on the grid Lower peak-power/energy costs, after amortization period (2) Cons: Location may not be good for DC (1) or renewable plant (2) Energy may have already been committed (1) Need to install and maintain renewable plant (2) Examples: Microsoft built DC near hydro plant in OR (1) Apple is building 20MW solar array in NC (2)
Outline DC energy usage and carbon footprint Reducing carbon with renewables Our target and research challenge Software and hardware for leveraging solar energy Current and future works Conclusions
No approach is perfect Our target Different DC operators may take different approaches Co-location or self-generation with solar and/or wind Pros: Clean and available Cons: Space and cost
g CO2e per KWh over lifetime Solar and wind are clean 1000 900 800 700 600 500 400 300 200 100 0 [Sovacool 08]
g CO2e per KWh over lifetime Solar and wind are clean 1000 900 800 700 600 500 400 300 200 100 0 [Sovacool 08]
Solar is more available in the US Wind Solar [NREL 12]
Efficiency rates of PV modules Space: Solar PV efficiencies are increasing [IEA 10]
Generation as % of max 25 Space: Solar PV capacity factors today 25 20 15 10 5 1Q 2Q 3Q 4Q Year 20 15 10 5 0 Rutgers, USA 0 [PVOutput 12]
1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 2007 2009 2011 2013 2015 2017 2019 2021 2023 2025 2027 2029 2011 Dollars per Watt Cost of solar PV energy is decreasing 20 16 12 Installed 8 4 0 Panels Inverters [DOE 11,Solarbuzz 12] Grid electricity prices have been increasing: 30%+ since 1998 [EIA 12]
1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 2007 2009 2011 2013 2015 2017 2019 2021 2023 2025 2027 2029 2011 Dollars per Watt Cost of solar PV energy is decreasing 20 16 12 Installed spike in demand world-wide recession back to historical levels 8 4 0 Panels Inverters [DOE 11,Solarbuzz 12]
1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 2007 2009 2011 2013 2015 2017 2019 2021 2023 2025 2027 2029 2011 Dollars per Watt Cost of solar PV energy is decreasing 20 16 12 Installed 8 4 0 Panels Inverters < 1/2 of current cost [DOE 11,Solarbuzz 12] With incentives, the installed price can go down by another 50-60%
Solar space and cost: Present and future Space as a factor of rack area Present Future (2020-2030) Density per rack 8kW (200W 1U servers) ~47x ~24x 2kW (25W 0.5U servers) ~12x ~6x Assuming 30% server utilization, 50% solar energy, NJ capacity factor, and 1 row of panels Cost per Watt Present Future (2020-2030) ~$2.30 < $1.20 Assuming self-generation and federal + NJ incentives Time to amortize cost Present Future (2020-2030) ~12 years < 6 years Assuming above costs, NJ capacity factor, and NJ grid energy prices
Solar space and cost: Present and future Space as a factor of rack area Present Future (2020-2030) Density per rack 8kW (200W 1U servers) ~47x ~24x 2kW (25W 0.5U servers) ~12x ~6x Assuming 30% server utilization, 50% solar energy, NJ capacity factor, and 1 row of panels Cost per Watt Present Future (2020-2030) ~$2.30 < $1.20 Time to amortize cost Present Future (2020-2030) ~12 years < 6 years Assuming above costs, NJ capacity factor, and NJ grid energy prices Wind takes ~12x less space and is ~3x cheaper Assuming self-generation and federal + NJ incentives
Power Main challenge: Supply of power is variable! Solar power Workload Now Batteries and net metering are not ideal We need to match the energy demand to the supply
Main challenge: Supply of power is variable! Many research questions: What kinds of DC workloads are amenable? What kinds of techniques can we apply? Should we allow programmers to specify what can be done? How well can we predict solar availability? If batteries are available, how should we manage them? Can we leverage geographical distribution? Building hardware & software to answer questions
Outline DC energy usage and carbon footprint Reducing carbon with renewables Our target and research challenges Hardware and software for leveraging solar energy Current and future works Conclusions
Green DC software Follow the renewables [HotPower 09, SIGMETRICS 11] Duty cycle modulation with sleep states [ASPLOS 11] Quality degradation for interactive loads [UCB-TR 12] Adapt the amount of batch processing [HotPower 11] Delay jobs while respecting deadlines GreenSlot [SC 11], GreenHadoop [Eurosys 12]
Overall delay-until-green approach Predict green energy availability Weather forecasts Schedule jobs Maximize green energy use If green not available, consume cheap brown electricity May delay jobs but must meet deadlines Send idle servers to sleep to save energy Manage data availability if necessary
GreenHadoop scheduling Estimate the energy required by jobs (EWMA) Job1 Job4 Job5 Job3 Job6 Job2
GreenHadoop scheduling Assign green energy first Job1 Job4 Job3 Job5 Job6 Job2 Off-peak On-peak Off-peak Predict energy availability (weather forecast) Power Now Time
GreenHadoop scheduling Assign cheap brown energy Job1 Job4 Job3 Job5 Job6 Job2 Off-peak On-peak Off-peak Previous peak Power Now Time
GreenHadoop scheduling Assign expensive energy Job1 Job5 Current power Active servers Job3 Job2 Job4 Job6 Off-peak On-peak Off-peak Active servers Power Now Time
GreenHadoop scheduling As time goes by the number of active servers changes Active servers Power Now Time
Energy (kwh) Energy prediction vs actual 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 Prediction Actual Actual data from the Rutgers solar farm (scaled down to our 16-node cluster)
Energy (kwh) Energy prediction vs actual 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 cloud cover rain thunderstorm Prediction Actual Actual data from the Rutgers solar farm (scaled down to our 16-node cluster)
GreenHadoop for Facebook workload Hadoop Brown consumed Green consumed Green produced Brown price GHadoop 31% more green 39% cost savings Brown consumed Green consumed Green predicted Brown price
Outline DC energy usage and carbon footprint Reducing carbon with renewables Our target and research challenges Software and hardware for leveraging solar energy Current and future works Conclusions
The Rutgers Parasol Project
Parasol: Our hardware prototype Unique research platform Solar-powered computing Remote DC deployments Software to exploit renewables within and across DCs Tradeoff between renewables, batteries, and grid energy Free cooling, wimpy servers, solid-state drives Full monitoring: resources, power, temperature, air
Parasol details Installed on the roof Steel structure Container to host the IT 16 solar panels: 3.2 kw peak Backup power Batteries: 32 kwh Power grid IT equipment 2 racks 64 Atom servers (so far): 1.7 kw 2 switches and 3 PDUs Cooling Free cooling: 110 W or 400 W Air conditioning: 2 kw Heating: 3 kw
Outside and inside Parasol
Outline DC energy usage and carbon footprint Reducing carbon with renewables Our target and research challenges Software and hardware for leveraging solar energy Current and future works Conclusions
Current and future works DC placement with probabilistic guarantees GreenNebula Smart management of energy sources Green SLAs Tradeoff between performance and green energy use Collect and make sense of the monitoring data
Conclusions Reduce the carbon footprint of ICT, data centers Topic is interesting and has societal impact Lots left to do More info -- http://parasol.cs.rutgers.edu