A Computer Scientist Looks at (Energy) Sustainability Randy H. Katz University of California, Berkeley NSF Sustainable Energy-Efficient Data Management Workshop 2 May

Working Definition Sustainability: strategies that meet society’s present needs without compromising the ability of future generations to meet their own needs Satisfaction of basic economic, social, and security needs now and in the future without undermining the natural resource base and environmental quality on which life depends – Energy, Water, Natural Resources, Environment, … – Green Manufacturing, Transportation, Agriculture,... – Increase efficiencies and minimize bad side-effects Use less, use what you need, use better 2

Role of Information Technology “Energy permits things to exist; information, to behave purposefully.” W. Ware, 1997 Observe-Analyze-Act – Observe: Sense, Monitor, Collect – Analyze: Organize, Model, Infer, Plan – Act: Control, Actuate, Execute, Manage 3

Energy “Spaghetti” Chart 4 Quads (10 15 BTUs) data 4

Sources and Loads Dispatchable Sources Oblivious Loads Non-Dispatchable Sources Aware Loads 5

Supply- versus Load- Following Most expensive, least efficient energy Latency involved in bringing capacity on line (or probability of exceeding) 6 Base Capacity Intermediate Capacity Peaker Capacity Load-following Supply Demand Response: Incentivize reduced loads during times of peak demand Demand Side Management: Shift demand to reduce peak loads, e.g., Supply-following Loads Load Duration Curve

21 st Century Energy Infrastructure Energy: the limited resource of the 21st Century Role of IT: Information Age approach – Overarching Principle: bits follow current – Pervasive actionable awareness of energy availability and consumption, on fine time scales Exploit information to match sources to loads, manage buffers, integrate renewables, signal demand response, and take advantage of locality 7

8 Instrumentation Models Controls Building OS Plug Loads Lighting Facilities Building Instrumentation Models Routing/Control Grid OS Demand Response Load Following Supply Following Grid Facility-to- Building Gen-to- Building Instrumentation Models Control Compressor Scheduling Temperature Maintenance Supply-Following Loads Storage- to-Building Instrumentation Models Power-Aware Cluster Manager Load Balancer/ Scheduler Web Server Web App Logic DB/Storage Machine Room MR-to- Building Energy Networks Gen- to-Grid uGrid- to-Grid Building- to-Grid

Datacenter as a Supply-Following Load 1.Degree of Freedom: On-demand + scheduled workloads 2.Principle: Power proportionality from non- power proportional components 3.Sustainability: Maximize use of renewable sources 9

Supply-side Challenge: Wind High variability of wind energy is an impediment to its large- scale penetration in traditional Grid/Load architectures Single LocationMultiple Locations vs. Time of Year Std dev / mean +/-11MW in 20 min 10

Load-side Challenge: Power Proportionality Scheduling agility: workload awareness and resource allocation Wikipedia interactive workload + HPC batch workload 11 Time Static Load Provisioning: 100% overprovisioning over worst expected case Dynamic Load Arrival Dynamic Capacity Management Over- provisioning or Spinning Reserve Work Capacity

Energy-Aware System Architecture 12 Energy Usage Price/ Renewable Energy Wake/Sleep Utilization, Response Times, Dropped Requests Add/Remove Nodes Delay/Degrade Jobs Database / SAN Load Balancer / Job Scheduler Clients TRANSITION AWAKE SLEEPING Web App Web Server Workload/ Request Statistics Web App App Server Web Server Batch Server Internet Power Subsystem: Metering, Distribution, Battery Control Grid Electrical Power Cluster Manager Load Monitor Workload Prediction Cluster Provisioning Energy Policy Charge/Discharge Battery Observe Analyze Act

Server Efficiencies Operating Range Server Class Machines (similar figure for netbook/embedded class nodes) 13 Better Measured

Effectively Scaling Work Capacity and Power 14 Measured experiment on LoClu

Price/Elastic Workload Response 15 Requests degraded Response rate maintained Energy and cost reduced Wikipedia workload

Batch Processing and Slack 16 Run Immediately, Grid-Oblivious 54% Grid, 46% Wind Greedy, Grid-Aware 30%, 70% Wind Grid energy down, wind energy up

Information Overlay to the Energy Grid 17 Conventional Electric Grid Generation Transmission Distribution Load Intelligent Energy Network Load IPS Source IPS energy subnet Intelligent Power Switch Conventional Internet

Aware Co-operative Grid 18 Observe-Analyze-Act: Deep instrumentation Waste elimination Efficient Operation Shifting, Scheduling, Adaptation Forecasting Tracking Market Availability Pricing Planning Power Proportional Cluster as a Model System applied to the Smart Grid—now distributed

19 Smart Buildings Cory Hall Soda Hall SDH

Smart Buildings BMS Cyber PhysicalBuilding Light Transport Process Loads Occupant Demand Legacy Instrumentation & Control Interfaces Pervasive Sensing Activity/Usage Streams BIM BITS PIB Activity Models Multi-Objective Model- Driven Control Building Integrated Operating System External HVAC Electrical Fault, Attack, Anomaly Detect &Management Control Plan and Schedule Physical Models Human-Building Interface 20 Observe – Analyze – Act

Building Observation 21

22 Environmental Operational Building Analysis

23 Lighting Servers PDUs, CRACs Building Analysis HVAC IT and Plug Load Soda IT-intensive CS Building

Building Action 24 Soda Hall Etchevery Hall 50 Ton Chiller 200 Ton Chiller 10 months 2 months Scott McNally Bldg Manager

Building OS 25 sMAP: Simple Measurement and Actuation Profile for Physical Information StreamFS: Storage System for energy sensor data Scheduling and Slack/Supply-following Loads

26 Summary Awareness of Load and Supply – Load-Following: match load with managed supply – Demand Response: reduce load to meet supply – Supply-Following: schedule work to exploit knowledge of available supply—essential for non-dispatchable sources like wind and solar Key idea: make information actionable – Observe-Analyze-Act – Information overlay on cluster, machine room, building-scale “grids” – Interface sensors, facilities, clusters, and buildings to information buses at a variety of scales

27 Conclusions Smart Clusters, Smart Buildings, Smart Grids – Use less energy Right provisioning for expected + reserve vs. peak – Use the energy you need: Power proportionality – Use better energy Integrate renewables

28 Thank You!