Data centers & energy: Did w id we ge get it t it backwards ds? Adam Wierman, Caltech
The typical story about energy & data centers:
The typical story about energy & data centers: Su Sust stainab able d dat ata a centers Remember: The cloud is (often) more efficient than the alternative.
The typical story about energy & data centers: But maybe we got it backwards?
Renewable energy is coming! …but incorporation into the grid isn’t easy
Today’s grid Generation Load Key Constraint: Generation = Load (at all times) low uncertainty
Today’s grid Generation Load Key Constraint: Generation = Load (at all times) controllable low uncertainty (via markets)
Tomorrow’s grid Key Constraint: Generation = Load (at all times) less controllable low uncertainty high uncertainty
1) Huge 1) ge pri price v vari riability, leading to generators opting out of markets! 2) 2) More c e conven enti tional res eser erves es n needed eeded, countering sustainability gains! Key Constraint: Generation = Load (at all times) less controllable low uncertainty high uncertainty
1) Huge 1) ge pri price v vari riability, leading to generators opting out of markets! 2) 2) More c e conven enti tional res eser erves es n needed eeded, countering sustainability gains! Key Constraint: Generation = Load (at all times) less controllable low uncertainty high uncertainty
Grid needs huge growth in demand response (or storage) Data centers are a promising option 500 k 500 kW-10 100 M 0 MW each …they are large loads …usage is growing quickly …highly automated 10 10-15% grow growth/year …they have significant flexibility
Build lding m manage geme ment 5% in 2 min / 10% in 20min [LLNL] e.g. cooling, lightin cost g, … flexibility Data centers are a promising option …they are large loads …usage is growing quickly …highly automated 10+ years of research into …they have significant flexibility energy-efficient data centers
Build lding m manage geme ment 5% in 2 min / 10% in 20min [LBNL] e.g. cooling, lighting, … cost flexibility Data centers are a promising option …they are large loads …usage is growing quickly …highly automated 10+ years of research into …they have significant flexibility energy-efficient data centers
Workloa load m manage geme ment 10-30+% in 10-60min [LBNL,HP] e.g. demand shaping, geographical load balancing, quality degradation, … cost flexibility Data centers are a promising option …they are large loads …usage is growing quickly …highly automated 10+ years of research into …they have significant flexibility energy-efficient data centers
Microgri rogrid manage geme ment 10-100% in 5-30min e.g., Battery management, local PV, Backup generation cost flexibility Data centers are a promising option …they are large loads …usage is growing quickly …highly automated 10+ years of research into …they have significant flexibility energy-efficient data centers
A new story about energy & data centers: Data centers are valuable resources for making the grid sustainable =
What t is is th the p pote otential of of data ta ce center de demand r respon onse?
A case s A study tudy: PV data center interactive workload Optimally placed, fast charging rate storage batch workload PUE
A A case s study tudy:
1 1 MWh Wh if geogra raph phica cal l load b bala lanci cing g is used! A case s A study tudy: $1-5 m $1 milli llion on c cost!
Where a are we we toda today? Time of use pricing Coincident peak pricing Wholesale markets Ancillary service markets Emergency DR Data centers rarely participate … and if they do it is highly inefficient
coincident peak warnings peak Where a are we we toda today? Time of use pricing Coincident peak pricing Wholesale markets Ancillary service markets - Risky to participate Emergency DR - Few opportunities for utility to extract response For more see [Liu et al 2013]
Where a are we we toda today? Time of use pricing Coincident peak pricing Wholesale markets Ancillary service markets Emergen ency cy D DR
How ca ow can we we do do bett tter? Algorit ithm desig sign fo for da data c cen enter ter p parti ticipation Eng ngine neering ng: + [Camacho et al 2014], [Chen et al 2013, 2014], [Ghamkhari et al 2012, 2014], [Aikema et al 2012, 2013], [Irwin et al 2011], [Urgaonkar et al 2013, 2014], [Li et al 2012, 2013], [Liu et al 2013, 2014] Economic ics: s: New ew marke ket des t designs [Liu et al 2014], [Chen et al 2014],[Ghamkhari et al 2013], [Li et al 2013], [Wang et al 2014], [Ren et al 2015], [Wierman et al 2014], [Zhang et al 2015]
How ca ow can we we do do bett tter? Performance is priority #1 Algorit ithm desig sign fo for da data c cen enter ter p parti ticipation Eng ngine neering ng: + Highly regulated, change is difficult Economic ics: s: New ew marke ket des t designs …bu but, a adopt ption re repre prese sents s a huge c challe llenge
A A sta tarti ting p poin oint: Col oloca ocated (mu (multi-tenant nt) d ) data cent enter ers
A A sta tarti ting p poin oint: Col oloca ocated (mu (multi-tenant nt) d ) data cent enter ers Hyp yper er-scale ( (e.g. .g. goog google): 7 7.8 .8% Ent nterprise: 53% 53% Coloc location on: 3 : 37% …of total data center industry electricity usage 29
Why col Why coloca ocated data cent enter ers? Building ope g opera ration i is sepa parated f from rom com computing pri priorities …but th t the e da data ta c cen enter ter w would l like to ke to participate i te in dem demand r res esponse! e! + On-site generation provides backup! + Market power isn’t an issue! Set u et up incen enti tives es for ten tenants ts + No regulation – can do whatever they want! + Tenants are heterogeneous in workloads! Lot’ Lot’s of of wor work to b k to be don one…
The typical story about energy & data centers: But maybe we got it backwards? Key p y point nts: 1) ) We ne need t to move b beyo yond a a “myo yopic” f focu cus o on n a da data ce centers t to co cons nsider a a “sy syst stem-wi wide de” v view o w of sustaina nability. ty. 2) It is t is im important to c to con onsider m mor ore th than g goo oogle-like “ e “hyper per-sc scal ale” d dat ata a centers. s.
Data centers & energy: Did w id we ge get it t it backwards ds? Adam Wierman, Caltech
Recommend
More recommend
