Distributed Load Management e-Energy 2010 Kiril Schrder C. v. O. - PowerPoint PPT Presentation

Power and Cost Aware Distributed Load Management e-Energy 2010 Kiril Schröder C. v. O. University of Oldenburg schroeder@informatik.uni-oldenburg.de Daniel Schlitt Marko Hoyer Wolfgang Nebel OFFIS – Institute for Information Technology OFFIS – Institute for Information Technology C. v. O. University of Oldenburg daniel.schlitt@offis.de marko.hoyer@offis.de nebel@informatik.uni-oldenburg.de

Motivation 1/2  High energy consumption in data centers  Rising performance, slower rising energy efficiency of servers  Forecast 2020, Greenpeace: 623 TWh (DCs+Network)  Saving energy through better capacity utilization  Server consolidation using virtualization  Higher efficiency at runtime with dynamic load management  Through live migration  Now: reactive, Better: proactive using load prediction 14.04.2010 Power and Cost Aware Distributed Load Management 2

Motivation 2/2  Enterprises maintaining several data centers  Different data center efficiency and location characteristics  Operation  Temperature, humidity -7°C  Energy price  Energy supply  Sun and wind power +5°C  Application  Data center network  Marketplace of computing power 14.04.2010 Power and Cost Aware Distributed Load Management 3

Vision of Distributed Load Management 1/3  Data Center (Server, Cooling, UPS etc.)  Database Replication  Network  Live Migration (migration costs)  Problems  Data Security and Privacy,  Control,  Responsibility etc.  like in Cloud Computing 14.04.2010 Power and Cost Aware Distributed Load Management 4

Vision of Distributed Load Management 2/3 Simple example for migrations:  Constant work load  Three data centers with different energy prices  Lowest prices rotating Reality, much more difficult:  Varying loads  Unsure forecasts  Different migration costs  Need dynamic (partial) replication  Good locations can change 14.04.2010 Power and Cost Aware Distributed Load Management 5

Vision of Distributed Load Management 3/3 Related Work  IBM: Utilize Sun Power  Asfandyar Qureshi et al., MIT: “Cutting the Electric Bill for Internet -Scale Systems”  Query Distribution  Full Replication  Traces for  Electricity Prices  Load & Traffic 14.04.2010 Power and Cost Aware Distributed Load Management 6

dependent P max Potential Analyses power consumption load P min independent standby active server load  Simulation Setup P standby  Simple data center model 0 % 50 % 100%  (Semi-)Homogeneous environment work load  Typical parameter values in Germany  Savings hot air free air cooling always active  Energy (~Temperature): less than 5 %  Present energy prices: up to 10 % cold accumulator/ cooling unit free air cooling control  Dynamic energy prices: up to 40 % control and hybrid mode  Electricity Regulation humidifier active in full climate chiller  Reducing peak demand  Enable even demand (50 %) free air cooling cold air  Increasing power station efficiency 14.04.2010 Power and Cost Aware Distributed Load Management 7

Ongoing and required future work  Detailing simulation setup energy supply renewable energy with with time varying private  weather-dependent Modeling: peak loads households feed-in dynamic  Data Center Abstraction (in progress) electricity pricing regional data center  Network (in planning) energy internet access provider video on demand cloud computing  Load Prediction (proceeded) data center as an adaptive regulating outsourcing bulk consumer of services  Load Management (in progress) cooperating  data centers Adaptive Replication (in progress) S ME s  Embedding in Smart Grids (in planning) energy demand and price aware load management 14.04.2010 Power and Cost Aware Distributed Load Management 8

Summary  Data center comprehensive load management of virtual machines  High saving potentials: Costs > Energy  Several application areas  Energy and cost aware cloud service  Large enterprises, network of cooperating enterprises  Marketplace of performance (open cloud)  Embedding in smart power grids  Demand regulation  Increasing efficiency  Many challenges  Technical: assurance of service level agreements, network coordination …  Political: (legal) obligation, different national laws … 14.04.2010 Power and Cost Aware Distributed Load Management 9