1. Quantifying the Environmental Advantages of Large-Scale Computing Vlasia Anagnostopoulou (vlasia@cs.ucsb.edu), Heba Saadeldeen, and Frederic T. Chong Department of Computer Science, University of California Santa Barbara

2. E-Business datacenter dilemma In addition, manufacturing and use phase of datacenters burden environment

3. Datacenter environmental implications Manufacturing: –A desktop computer requires 260 kg of fossil fuel + 6400 MJ of energy (Williams)‏ –An average cooling unit: 527 kg of primary materials –An average power unit: 8 ton Use phase (24/7) : –As of 2006, U.S. datacenters consumed 10% of total U.S. energy consumption, projected to double in 5 years –For lower growth projection: aggressive efficiency Therefore, environmental cost is significant!

4. Besides deployment costs, what are the environmental and operational costs with datacenter scale? Trade-offs among them?

5. Rest of presentation Overview of datacenter operation and characterization Datacenter and business models Model cost analysis –Environmental –Capital (CAPEX)‏ –Operational (OPEX)‏ Lessons Conclusion Future work

6. Overview of datacenter operation + characterization

7. Datacenter power distribution

8. Tier classification Tier-I: no redundancy Tier-II: redundancy N+1 Tier-III, Tier-IV, …

9. Cooling Operation

10. Datacenter-in-a-container Standard-sized container Very efficient air-flow –Better PUE (Power Usage Efficiency = Total Power/ IT-Power)‏ External cooling and power loops are same

11. Datacenter and business models

12. Datacenter and Business Model Datacenter case: Various datacenter sizes –Comp. room (1-2 racks), Small (20), Medium (50), Large (100)‏ –Based on vendor’s classification Building /container installation Cooling and power provisioned w/o redundancy (tier-I) and w. N+1 redundancy (tier-II) In case of comp. room, assume existing chilled-water loop Business case: Two representative types of business apps –E-commerce –Financial Simulated by TPC council’s TPC-C and TPC-H benchmarks (in respect)‏

13. Putting it all together Size# of businesses # of racks# of containers LocalComp.Room (TPC-H)‏ 1¾N/A Comp.Room (TPC-C)‏ 11.5N/A RemoteS18201 M47502 L951005

14. Model provisioning Strategy: capacity matching Not as precise as detailed model, but it is uniform and it does happen in practice(!)‏ Server provisioning –the state-of-the-art system from TPC council Cooling provisioning –from vendor’s specs, to match server heat load Power provisioning –to match server heat load + cooling load

15. Model cost analysis

16. Environmental cost -> Methodology For each size configuration: From vendor’s specs, add weights of power & cooling components Calculate amount of materials Use material breakdown tables to come up with amount of metals, plastic, and glass/ceramic Normalize over large configuration for comparison

17. Environmental cost -> Results

18. Environmental cost -> Explanation Material scaling dis-proportionality (Same trend for UPS)‏

19. CAPEX -> Methodology Here: Cooling and Power CAPEX (part of TCO)‏ Assumptions: –Loan with interest rate: 8% –Cooling & Power provisioning: 2x –Application-requirements double every 2 years –Small facility upgrade period is 4 years –Large facility upgrade is 6 months, except for Chiller –Life-cycle of 10 years

20. CAPEX -> Methodology

21. CAPEX-> Results Total capital costs: SizeCAPEX [Million $] Comp. Room6.1 S6.7 M5.4 L5.1

22. CAPEX->Results

23. OPEX -> Methodology Calculated PUE based on: –Active-power*work-hours + Idle-power*idle-hours –Power based on inefficiency related to size For container, used PUE from specs (same for all sizes)‏ SizePUE Comp. Room2.00 S1.76 M1.71 L1.69 Container (All sizes)‏1.25

24. OPEX->Results Total energy consumption: SizeEnergy in MWh Norm. Comp. R.687,400 Norm. S243,500 Norm. M238,000 Norm. L223,900

25. Lessons, Conclusions and Future Work

26. Lesson #1: Material efficiency Large (tier-I) installations are up to 53% more efficient Tier-II (w. N+1 redundancy) up to 75% Preferring a large installation can save up to: –95 tons of materials, from which: i.Primary metals: 62 tons ii.Plastics: 27 tons iii.Glass/Ceramic: 7 tons Because of disproportional use of materials in power and cooling manufacturing + effect of redundancy

27. Lesson #2: Operational efficiency Large (building) installations can have up to 16% better PUE Their operational energy consumption can be up to 67% less Containers can have up to 38% better PUE –(however, data from different sources)‏ Because the larger the datacenter, the less inefficiencies in power and cooling Although we don’t evaluate here, large datacenter have better practices and more staff resources

28. Lesson #3: Cost advantage A large installation can be up to 24% cheaper –Because of faster outpayment of loans However, a small datacenter installation is 10% more expensive compared to an equivalent # of comp. rooms –Because we assume that in the case of a comp. room deployment, the building’s chilled-water loop is used

29. Conclusion Quantified material, price, and operation efficiency with datacenter scale –Up to 75% material efficiency, 67% operational efficiency, and 24% in capital cost –Up to 95 tons less materials Container datacenters are even more efficient in their operation Exception is the deployment of a computer room, if it is to be hooked to the building’s chilled-water loop

30. Future work We plan to: Include more factors: –degradation of land –price of land –Operational savings because of VM migration Add staff and software expenses to OPEX Complete Life-Cycle Assessment: –Use LCA tools over manufacturing and use phases (e.g. GHG emissions, water pollution)‏ –Evaluate retirement options

31. The End Thanks for listening! Questions? Contact: vlasia@cs.ucsb.eduvlasia@cs.ucsb.edu URL: http://www.cs.ucsb.edu/~vlasia ArchLab: http://www.cs.ucsb.edu/~archlab Computer Science Dept.: http://www.cs.ucsb.edu