Greening Geographic Load Balancing
Data Center Demand Response: Coordinating IT and the Smart
GridZhenhua [email protected] Institute of
TechnologyDecember 18, 2013Acknowledgements:Adam Wierman1, Steven
Low1, Yuan Chen2, Minghong Lin1, Lachlan Andrew3, , Cullen Bash2,
Niangjun Chen1, Ben Razon1, Iris Liu1 1California Institute of
Technology, 2HP Labs, 3Swinburne University of Technology
job market12Sustainable IT IT for sustainabilityEnergy
efficiency of IT systemIT as a demand response providerbroader
impacts2Renewables are coming3Cumulative capacity has grown by 72%
from 20002011
Wind and solar grow fastest (13x and 51x)
Source: Gelman, R. (2012). 2011 Renewable Energy Data Book
(Book). Energy Efficiency & Renewable Energy (EERE)Worldwide
Renewable Electricity Capacity increase in renewable3Challenges
with renewables4
Generation
TimePower12 AM12 AMGeneration = Demandat all timesat all
locations
DemandKey constraint: predictablecontrollablelow
uncertaintyGeneration follows Demandtraditional approach:
generation follows supply4Challenges with renewables5
GenerationGeneration = Demandat all timesat all locations
DemandKey constraint: responsiveless controllablehigh
uncertaintyDemand follows Generation(to some extent)
expensivechallenges with renewable integration5Need huge growth
in demand response6Data centers are a promising option
Wind and Solar capacities are growing 15~40% per year large
loads: 500kW~50MW each
increasing fast: 10~15% per year
significant flexibilitiesOthers vs DC6Data center
flexibilitiescooling, lighting, 5% of consumption can be shed in 2
min [LBNL2012] 10% of consumption can be shed in 20 min [LBNL2012]
workload managementTemporal demand shaping [Sigmetrics12][3
patents]HP Net-Zero data center, 2013 Computerworld Honors
LaureateGeographical load balancing
[Sigmetrics11][GreenMetrics11][IGCC12]Best student paper award at
ACM GreenMetrics 2011Best paper award at IEEE Green Computing
2012Pick of the Month in the IEEE STC on Sustainable Computing
onsite backup generators & storage7previous work: design
& implementation, distributed and online algorithms with
theoretical guarantee7
Geographical load balancingpower and coolingspatial
flexibility
future: robustness when adding new data center or failure8Data
center flexibilitiescooling, lighting, 5% of consumption can be
shed in 2 min [LBNL2012] 10% of consumption can be shed in 20 min
[LBNL2012] workload managementTemporal demand shaping
[Sigmetrics12][3 patents]HP Net-Zero data center, 2013
Computerworld Honors LaureateGeographical load balancing
[Sigmetrics11][GreenMetrics11][IGCC12]Best student paper award at
ACM GreenMetrics 2011Best paper award at IEEE Green Computing
2012Pick of the Month in the IEEE STC on Sustainable Computing
onsite backup generators & storage9great opportunitiesData
center demand response today10coincident peak pricing
(CPP)timecustomer power usagesystem peak hour(decided by
utility)coincident peak demandcustomers peak demandMany
programsTime of use (ToU) pricingWholesale marketAncillary service
market
Monthly bill = fixed charge + usage charge + peak charge +
coincident peak charge CPP is popular
how CPP works10CPP in practiceRates at Fort-Collins Utilities,
Colorado, USA11CP is very important!fixed charge:
$101.92/monthusage charge rate: $0.0245/kWhpeak charge rate:
$4.75/kWcoincident peak (CP) charge rate: $12.61/kWExample: average
demand 10MW, peak demand 15MW, CP demand 14MWMonthly bill = fixed
charge + usage charge + peak charge + coincident peak charge
$101.92$176,400$71,250$176,540
CPP is very important11DC management is
challenging12Uncertainties in CPonly known at the end of the
monthParticipating CPP program is risky!algorithm
designchallenge1213mind f(d; t)expected cost optimizationdata
mining for patternsless accurate with renewablesrobust
optimizationmind Et[f(d; t)]mind maxt [f(d; t)]online
algorithmoptimal competitive ratioExtensions warning signals backup
generator & local renewables workload & renewable
prediction errorswhat is f
two approaches
performance guarantee
extension1314mind f(d; t)expected cost optimizationrobust
optimizationTimePower12 AM12 AMperiods with high probability to be
CPTimePower12 AM12 AMmake the demand flatLimited demand
responsemarket designlimited demand response14Potential of data
center demand response15
Goal: minimize voltage violation with large PV generation20MW DC
3MWh storage=voltage violation ratewith 20% flexibility optimal
location & fast charge rate
great potential from the societys perspective15Pricing data
center demand response16
supply function si(p)Pricing data center demand
responseefficiency loss due to user strategic behavior
[XLL2013]17
market-clearing price p
supply function bidding
but when we have data centers works well when no user has large
market powerPricing data center demand response18
price p
prediction-based pricing supply function
Pricing data center demand response19
prediction-based pricing supply si(p)efficiency loss is
independent of market powerbut depends on prediction accuracy
parameter in supply functionfor quadratic cost function
20supply function biddingprediction-based pricing vsefficiency
loss depends on market powerefficiency loss depends on prediction
accuracy
supply function biddingprediction-based pricing
supply function biddingprediction-based pricing21supply function
biddingprediction-based pricing vsincorporating power networkvalue
of locationoptimal power flowlearning from user
responseexploitation vs explorationtheory of
quantization[BSXY2012]Pick of prices during learning stageDesign
demand response menufuture work
connection2122
demand response
flexibilitiescloud platform23Thank you!References[LBNL2012]
Ghatikar, Girish, et al. "Demand response opportunities and
enabling technologies for data centers: Findings from field
studies." LBNL-5763E. 2012.
[XLL2013] Yunjian Xu, Lina Li, Steven Low. On the Eciency of
Parameterized Supply Function Bidding with Capacity Constraints.
2013.
[BSXY2012] Bergemann, Dirk, et al. "Multi-dimensional mechanism
design with limited information." Proceedings of the 13th ACM
Conference on Electronic Commerce. ACM, 2012.24Model for prediction
based pricing25
user
for each realizationcost function
supply
utility
penalty
social objective
offline optimal
Model for prediction based pricing26
utility
penalty
social objective
offline optimalperformance evaluation
competitive ratio
Theorem