Zero-carbon Compute Project

Andrew A ChienUniversity of Chicago and Argonne National Laboratory

achien@cs.uchicago.edu

October 1, 2019

Objective: Launch a consortium of industry and academic partners thatwill align, focus, and publicize efforts on the fundamental challenges of com-puting’s carbon and environmental impact. Create a Zero-carbon and zeronet-waste path for the computing industry. Recruit a critical mass of part-ners first one, then several of the clusters. Grow funding and participationto critical mass.

Project Approach and Potential Impact

Figure 1: Computing’s massive e-waste legacy

Cloud computing’s rapid growthhas reached a scale where its power-use carbon footprint and e-wasteenvironmental impact is a signif-icant element of the global prob-lems of climate change and envi-ronmental pollution [1, 2, 3]. De-spite efforts to promote renewableenergy generation and offset poweruse, cloud computing faces poten-tial societal limits as cloud poweruse approaches 10% of grid powerin some regions, and is already the#1 driver for new power generationand transmission in many parts of the world [4].

The exclamation mark for these troubling challenges is the recent emer-gence of machine learning as a major new growth factor for computing.With explosive growth Driven by exciting new capabilities and the abilityto “learn function” from data, machine learnig is accelerating the already

1

fast capacity growth of datacenters [5, 6, 7]. The confluence of these factorsis creates a powerful urgency for the creation of new zero-carbon and zeronet waste practices. As computing professionals, we all believe in comput-ing’s potential to be a positive force for the environment, and for society toview it as such.

Figure 2: Computing mixed with Re-newables

Research Background and PlanOver the past 5 years, the Zero-Carbon Cloud research project(ZCCloud) has pursued new ap-proaches that use power pricinginformation and renewable poweravailability to reduce the carbonfootprint of cloud computing, andcreate new cloud software that en-ables productive use of the vari-able and intermittent computingresources that arise [8, 9, 10, 11,12, 13]. The promise of theseapproaches and research successeshave inspired new research collab-orations with major cloud playerssuch as Google and Intel as well as underpinned the creation of ambitiousnew startups creating “stranded power” driven intermittent computing re-sources [14]. We described a number of the open research challenges in arecent white paper [15]

While our growing research project continues apace, any single projecthas limitations. The urgency and scope of the problem frame an urgent needfor a consortium analogous to the Open Compute Project[16] (which fostersopen hardware designs in cooperation between data center operators (cus-tomers) and hardware system product designers, with the goal of enablingmore rapid innovation and adoption (deployment). We describe potentialmajor vectors this proposed Zero-carbon Compute Project could initiate:

Zero-Carbon Computing Operations - Grid Coupling Projects thatdesign, evaluate, and document both new techniques and best practices, forcoupling large-scale computing resources to the power grid with goals of 1)reducing carbon impact of data center power use, 2) increasing renewableabsorption of the power grid, 3) reduced power cost, 4) increased power gridresilience and 5) avoiding negative impacts on other power grid customers.

2

Figure 3: Explosive growth of Machine-learning Computing Requirements(drives environmental impact).

Examples include:

• Best practices and open source software for single-site, multi-site powerpurchasing, renewables coupling – onsite, remote long-term power-purchase agreements.

• Best practices for coupling local energy storage with the grid (localmanage, collective manage, load cost savings, overall grid “social wel-fare”)

• Power market designs for fair coupling of collections of data centerswith broader power purchase and dispatch. Provable, auditable posi-tive grid impact (carbon, renewable absorption, prices for others) forthe cloud power users

In general, pursue projects that drive increased capability of the cloudand grid ecosystem to reduce carbon impact of growing cloud data centerspower use. Partner Types: hyperscale cloud operators, enterprise/edgecloud operators, university data centers, government agencies. Entities thatoperate computing facilities at scale. Power grid companies (IndependentSystem Operators). Researchers.

Zero-Carbon Computing Operations - Cloud / HPC Software Projectsthat design, evaluate, and document approaches for software to support dy-namic operation of cloud infrastructure and applications. This dynamismincludes both changes in capacity (e.g. a datacenter operating at X, then

3

0.7*X, then 1.2*X of a series of time intervals) and changes in availability(e.g. a service replica in an availability zone becoming unavailable due topower availability change. These changes may come with advance warnings(24-hour, 1-hour, 5-minutes), or be predictable. Examples include:

• Standard interfaces for service availability and capacity managementfor key cloud and distributed cloud services. Systematic exposure offlexibility information to support flexibility, and to export changes inavailability or capacity.

• Distributed infrastructure services and applications - designs and opensource – that tolerate availability and capacity change whilst main-taining high reliability and quality. Resource management systems forcritical user-responsive, deferrable, and even optional work.

• Distributed protocols and open-source implementations of heteroge-nous capacity and availability tolerant protocols. Adaptive versionsto optimize based on measured or guaranteed properties. Adaptiveversions based on prediction of properties.

In general, pursue projects that create software - infrastructure and ap-plications – that enable tolerance and themselves can tolerate underlyingdata center dynamic capacity management to reduce carbon impact. Part-ner Types: cloud application software companies, cloud operators, enter-prise/edge cloud operators, university data centers, government agencies.Researchers.

Zero Net waste - Datacenter Computing Lifecycle Projects thatdesign, evaluate, and document optimization of datacenter hardware life-cycle for reduced, and ultimately zero net waste operations. A focus onextending operational lifetime for hardware through exploitation of low-costpower and physical infrastructure (remote, low-cost facilities), and a lifecycleacross facilities – or even organizations.

• Best practices for extending the operational lifetime of hardware, bycreating new lifepaths across sites, operational practice, progression oflow-cost spaces.

• Creation and tracking of lifetime and “net waste rate” metrics thatcapture and track progress of organizations and the entire industryacross these lifepaths.

• Framing of new partnership models (traditional data centers and novellifetime-extender sites), service share, compute kickback, and zero-carbon operation.

4

In general, any projects that create new opportunities to extend thelifetime of datacenter computing hardware, promote adoption of these tech-niques, and document their adoption. Thereby, reducing the environmentalimpact per compute used. Partner Types: computing hardware vendors,hyperscale cloud operators, enterprise/edge cloud operators, university datacenters, government agencies. Researchers.

Zero Net waste - Mobile and Internet-of-Things Devices Scopeand examples analogous to above, but appropriate for mobile and internet-of-things devices.

Expected Impact

Create new technical paths and industry movement vectors that enable thecontinued growth of computing in the face of increasing societal concern[6, 2] and downright resistance [17].

Plans

Y1H1 Define Project clusters, Recruit Partners, Define Governance

Y1H2 Initiate Project, Execute Projects in Teams, Publish Whitepapers

Y2 Define Additional Clusters, Recruit Partners

Acknowledgements

This research was supported by in part by the National Science Foundationunder Awards CNS-1405959, CMMI-1832230, and CNS-1901466. Supportfrom the CERES Center for Unstoppable Computing, Google, Samsung,and Intel is also gratefully acknowledged.

References

[1] A. A. Chien, “Owning computing’s environmental impact,” Commun.ACM, vol. 62, no. 3, Feb. 2019.

[2] N. Jones, “How to stop data centres from gobbling up the worlds elec-tricity,” Nature, September 12 2018.

[3] U. N. University”, “The global e-waste monitor 2017,” United Nations,Tech. Rep., December 2017.

5

[4] Greenpeace, “Clicking clean: Who’s winning the race to build a GreenInternet, 2018 edition,” 2018, http://www.greenpeace.org/.

[5] Amodei and Hernandex, “Ai and compute,” openai.com/blog/ai-and-compute/, May 2018, 10x growth per year.

[6] K. Hao, “Training a single ai model can emit as much carbon as fivecars in their lifetimes,” Technology Review, June 2019.

[7] E. Strubell, A. Ganesh, and A. McCallum, “Energy and policyconsiderations for deep learning in NLP,” CoRR, vol. abs/1906.02243,2019. [Online]. Available: http://arxiv.org/abs/1906.02243

[8] F. Yang and A. A. Chien, “ZCCloud: Exploring Wasted Green Powerfor High-Performance Computing,” in IPDPS’16, 2016.

[9] K. Kim, F. Yang, V. Zavala, and A. A. Chien, “Data centers as dis-patchable loads to harness stranded power,” IEEE Transactions onSustainable Energy, 2016, dOI 10.1109/TSTE.2016.2593607.

[10] A. A. Chien, F. Yang, and C. Zhang, “Characterizing curtailed anduneconomic renewable power in the mid-continent independent systemoperator,” AIMS Energy, vol. 6, no. 2, pp. 376–401, December 2018.

[11] A. A. Chien, R. Wolski, and F. Yang, “Zero-carbon cloud: A volatileresource for high-performance computing,” in SHPC’15. IEEE, 2015.

[12] C. Zhang, V. Gupta, and A. A. Chien, “Information models: Creatingand preserving value in volatile cloud resources,” in 2019 IEEE In-ternational Conference on Cloud Engineering (IC2E), June 2019, pp.45–55.

[13] H. Nguyen, C. Zhang, Z. Chen, and A. A. Chien, “Real-time serverless:Enabling application performance guarantees,” submitted for publica-tion, August 2019.

[14] “Lancium,” www.lancium.com.

[15] A. A. Chien, C. Zhang, and H. Nguyen, “Zero-carbon cloud: Re-search challenges for datacenters as supply-following loads,” Universityof Chicago, Tech. Rep. CS-TR-2019-08, February 2019.

[16] “Open compute project,” www.opencompute.org.

6

[17] B. Tarnoff, “To decarbonize we must decomputerize: why we need aluddite revolution,” The Guardian, September 2019.

7