The Green/ Efficiency Metrics Conundrum Anand Ekbote Vice President, Liebert Monitoring Emerson Network Power
Agenda Green Grid, and Green/ Efficiency Metrics Data Center Efficiency Revisited Energy Logic I: Reducing Data Center Energy Consumption Increasing Data Center Compute Output Analysis Reveals Surprising Insights Energy Logic II: The Four Prioritized Efficiency-Improving Actions 2
The Green Grid A global consortium dedicated to advancing energy efficiency for data centers and business computing ecosystems by: Defining meaningful, user-centric models and metrics Developing standards, measurement methods, best practices and technologies to improve performance against the defined metrics Promoting the adoption of energy efficient standards, processes, measurements and technologies 3 3
Who s s In The Green Grid? Board of Director Member Companies 4
Contributor Members 1E ADP Avocent Brocade Communications Systems BT plc Chatsworth Products, Inc. Cisco COPAN Systems Delta Products Corporation Digital Realty Trust Eaton EMC Emerson Network Power Enterprise-Rent-A-Car Fujitsu Siemens Computers GmbH Greene Engineers Novell PG&E Pillar Data Systems Qimonda QLogic Rackspace Saft Power Systems Inc. SatCon Stationary Power Systems SunGard Data Systems Symantec Corporation Teradata Texas Instruments The Uptime Institute Trane Verari Systems, Inc. Verdiem Vette Corp Western Digital ZT Group Int'l Inc 5
The Green Grid Membership Membership Growth 160 140 120 100 80 60 40 20 0 Jan Feb Mar Apr May Jun Jul Aug Sept Oct Nov Dec Jan Over 175 members since February 2006 6
The Green Grid Technical Work Groups Data Collection and Analysis Focuses on investigations and analysis into the data center efficiency problem map EMEA Data Collection and Analysis Focuses on investigations and analysis into the data center efficiency problem map specifically for the EMEA region Data Center Technology and Strategy Focuses on existing and emergent technologies for data center efficiency Data Center Operations Focuses on use models, operational strategies, best practices and equipment standards Data Center Metrics and Measurements Focuses on data center characteristics, performance metrics, and measurement protocols for data collection 7
Agenda Green Grid, and Green/ Efficiency Metrics Data Center Efficiency Revisited Energy Logic I: Reducing Data Center Energy Consumption Increasing Data Center Compute Output Analysis Reveals Surprising Insights Energy Logic II: The Four Prioritized Efficiency-Improving Actions 8
IT Perspective on Energy Efficiency Top priority is delivering on service level agreements Performance - provide adequate compute capacity Reliability - redundancy at all steps Ability to support Security Does IT care about energy efficiency? Yes, but not if it impacts performance and reliability What if it frees up power and cooling capacity? Yes! If it does not impact performance and reliability 9
What Data Center Managers Are Looking For Objective vendor-neutral analysis Holistic view of the data center Quantification of savings from different strategies Prioritization of actions Actionable advice Tailored to different types of data centers 24x7 vs. 8x5; compute-intensive vs. transaction-intensive Payback / ROI analysis to help sell to management 10
Data Center Efficiency Revisited Data Center Efficiency = Data Center Output Energy Consumed Two Ways to Improve Efficiency: 1. Increase Data Center Output 2. Decrease Amount of Energy Consumed 11
Data Center Output: No Universal Measure Exists First we will address the issue of reducing energy consumption using Energy Logic, then we will turn our attention to addressing data center output. 12
Simple Data Center Layout (Energy Demand, Distribution and Supply) 2007 Emerson Network Power 13
Energy Logic: Roadmap to Reducing Energy Consumption Emerson Network Power approach to reducing data center energy consumption Sequential roadmap that starts with IT equipment and moves through to support infrastructure Emphasis is on cascade of savings Based on research and modeling Provides quantified savings and an estimated ROI Frees up power, cooling and space capacity without compromising availability or flexibility 2007 Emerson Network Power 14
Energy Logic Model 5,000 square foot Data Center 15
Energy Logic: The Cascade Effect 1 Watt saved at the server component level results in cumulative savings of about 2.84 Watts in total consumption 16
Energy Logic: Prioritized Energy Saving Strategies Higher AC Voltage improves efficiency 2007 Emerson Network Power 17
Cascade Savings Example - Lower Power Processor 2007 Emerson Network Power 18
Total Energy Logic Savings With All 10 Strategies Applied 2007 Emerson Network Power 19
Cooling, Space and Power Constraints Are Biggest Issues What is the biggest single issue you currently face? Source: Data Center Users Group Surveys 2007 Emerson Network Power 20
Energy Logic Addresses Space, Power & Cooling Constraints B E F O R E 65% space freed up from optimization from 5,000 sq. ft. / 465 sq. m. to 1,768 sq. ft. / 164 sq. m 43% cooling capacity and 33% power capacity saved A F T E R 2007 Emerson Network Power 21
Energy Logic: Payback Period 2007 Emerson Network Power 22
Energy Logic: 4 Key Takeaways 1. Start by reducing consumption at the IT equipment level and then work your way back through the supporting equipment Every watt saved at the equipment level has a cascading effect upstream. 2. Availability & Flexibility do not have to be compromised in order to increase data center efficiency. - Efficiency Without Compromise TM 3. High Density Architecture contributes toward increased efficiency - IT consolidation, Cooling Efficiencies 4. In addition to improving energy efficiency by reducing consumption, implementing these strategies frees up capacity of key constraints: Power, Cooling & Space Energy Logic White Paper Available Today 2007 Emerson Network Power 23
Data Center Efficiency: Importance of Measuring Data Center Output A Measure of Data Center Output is needed to help drive the right behavior for improving efficiency Lack of output metric limits focus and attention to the infrastructure (supply) side rather than on both the IT (demand) and infrastructure sides to consumption rather than on both output and consumption Data Center Efficiency = Data Center Output Energy Consumed 24
Measuring Data Center Output Challenges Data centers perform different types of work Processing-intensive for scientific and financial applications Data transfer-intensive for Web-based applications Data center requirements change as mix of workload shifts While a quantified measure of data center output does not exist, industry experts can agree that performance has improved dramatically over the last 5 to 10 years 25
IT Performance Improvement: 2002 2007 1998 2007 : 7400% Improvement (75X) 2002 2007 : 650% Improvement (7.5X) Raw Performance Gain 8000% 7000% 6000% 5000% 4000% 3000% 2000% 1000% 0% 75X 10X 1X 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 Source: Electronics Cooling magazine (Feb 2007) Belady, C., P.E., Hewlett-Packard, In the Datacenter, Power & Cooling Costs More than IT Equipment it supports 26
IT Performance Improvement: 2002 2007 2002 7.2 GFlops/ Server 2007 69.8 GFlops/ Server 2002 2007 870% Improvement (9.7X) Intel x86 2002 2007 TFLOPS 3.7 3.7 Servers 512 53 blades GFLOPS/server 7.2 69.8 Source: Intel 27
Introducing CUPS We introduce CUPS, or Compute Units per Second, as a temporary or placeholder measure for what will be the eventual universal metric for data center output Data Center Output Data Center Efficiency = = Energy Consumed CUPS Watts Consumed Based on information on performance gains, we assume CUPS has improved by 600% between 2002 and 2007 (compared to 650 %- Belady; 870% Intel) 2007 Emerson Network Power 28
How Does CUPS fit with Moore s s Law? CUPS 29
30 Server and Data Center Output and Efficiency Improvement 2002-2007 8.0 6.0 4.0 2.0 0.0 3000 2500 2000 1500 1000 500 0 1 5000 4000 3000 2000 1000 0 Total Server Power Draw (MW) 2293 Server Performance (MCUPS / Server) 7 2002 2007 321 1.8X 4027 2002 2007 7.0X 14.0X Server Efficiency (CUPS / Server Watt) 7.6X 2432 2002 2007 12.0 10.0 8.0 6.0 4.0 2.0 0.0 Data Center Efficiency (CUPS / Datacenter Watt) 1200 1000 800 600 400 200 0 Total Compute Output (TCUPS) 0.7 125 8.4X 9.8 2002 2007 1048 2002 2007
Data Center Efficiency Improved Dramatically from 2002 to 2007 Server efficiency improved over 650% (7.6x) Data Center efficiency improved over 735% (8.4x) If computing demand in 2007 was the same as in 2002, 2007 power consumption would have been <1/8th of 2002 consumption. Gets the Most Attention % Increase 1400% 1200% 1000% 800% 600% 400% 200% 0% 1300% 738% 59% 2002-2007 Consumption Output Efficiency 31
Data Centers: 0.3% of U.S. Energy Consumption Note: This ties in with EPA s assessment that data centers consume 1.5% of total US electric power. Source: Energy Information Administration / Annual Energy Outlook 2008 US EPA, Report to Congress Bureau of Transportation Statistics, National Transportation statistics 32
Efficiency Improvement: Cars vs. Computers CAGR 53.0% CAGR 0.8% If fuel efficiency had kept pace with data center efficiency improvement, cars would get 163 miles to the gallon! 33
Impact of Increase in IT Output Increase in IT output has had dramatic impact on business, economy and society: Significant improvement in productivity through automation of tasks and processes, and increased collaboration Better and faster decision making driven by availability of richer real-time information and communication Wider utilization of best cost resources around the world, driving global economic development Increased level of conveniences and benefits at the individual and societal level Increase in energy consumption has been small relative to increase in output -- and benefits to economy and society. 34
Applying Energy Logic: Improvements in Compute Efficiency CUPS / Datacenter Watt 5 IT Actions 5 Infrastructure Actions Base 604 1,335 2,198 0 500 1000 1500 2000 Low Power Processor High Efficiency Power Supply Power Management Features Blade Servers Server Replacement Virtualization 2.2x Efficiency Improvement! Power Distribution Architecture Cooling Best Practices All Ten Energy Logic Steps Variable Capacity Cooling 3.6x Efficiency Improvement!! High Density Cooling Monitoring & Optimization 35
Energy Logic II: 4 Key Steps Most impactful ways to improve data center efficiency: 1. Speed up refresh cycle for IT technology Blades provide a modular platform for continued improvement 2. Implement server power management policies 3. Virtualize 4. Adopt a high-density architecture Energy Logic II White Paper Available Today 36
Low-Power Processors AS3AP Transactions / Second (Higher is Better) Source: Anandtech 37
High-Efficiency Power Supplies LBNL reported power supply efficiency 72% - 75% at 30% load New power supplies have substantially higher efficiencies 89% - 91% @ 30% load Right-size your power supply Typically server power supplies are oversized to accommodate maximum server configuration Even though most servers are shipped at much lower configurations Higher losses associated with oversized power supplies 2007 Emerson Network Power 38
High-Efficiency Power Supplies (cont.) Which power supply will you choose? Power supply A: 91% efficient at nameplate rating, or power supply B: 93% efficient at nameplate rating? Power supplies are never at nameplate rating Dual power supplies are never loaded at >50% under normal conditions Spec the power supply which is more efficient at 10% - 35% load Typical Load in Dual Power Supply Configuration Nameplate Rating 90% 85% Efficiency 80% 75% 70% 2007 Emerson Network Power 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 2007 Emerson Network Power Percent of Full Load 39
Blade Servers Comparison of hardware for rackmount servers & blade servers Blade servers consume about 10% less power compared to equivalent rackmount servers Common components in chassis fans, communication cards, etc. Blades enable high-density architecture! 40
Leveraging the Power of Blade Servers Blades provide a modular platform from which to upgrade to higher efficiency components as technology advances Components (processors, power supplies, fans, etc.) can be replaced in a modular fashion Less administrative overhead compared to replacing whole servers; and no virtually operational disruption Virtualization can be implemented on blades Enable high-density architecture! 41
High-Density Architecture Traditional Architecture Fan Power- 8.5kW per 100 kw of Cooling Average entering air temperature of 80-84 F High-Density & Base Cooling Cooling Unit Fan Power- 3.5 kw per 100 kw of Cooling Average entering air temperature of 96-98 F Higher efficiency gains from cooling closer to the source Fan power reduces by up to 65% Higher entering temperature Lower capex due to lower floor space, chiller and switchgear capacity 42
Server Virtualization Before Virtualization Typical Virtualization Architecture Logical Server Virtualization increases server utilization by decoupling hardware and software Multiple logical servers on a single physical server Energy savings with fewer number of servers Consolidation ratios of 8:1 are typical Source: VMware 43
Server Power Management Server processors have power management features built in Can reduce power draw when processor is idle Typically power management features are turned off Turning on power management feature reduces processor idle power to ~45% of peak or less Test your OS / applications for latency 44
Defining Criteria for a Data Center Efficiency Metric A Measure of Data Center Output, even if less-than-ideal, can help drive the right energy-saving behaviors Effective measure vs. Ideal or Fair Measure Three criteria an effective measure must fulfill: 1. Most importantly, does it drive the right behavior? 2. Must be published at device level so that users can evaluate competing technologies 3. Must be scalable to the data center, allowing the output of the devices to be added together to produce an overall measure of data center efficiency Data Center Output Data Center Efficiency = Energy Consumed 45
Energy Logic Shows Using PUE Does Not Drive Right Behavior Measure Un-optimized Data Center Five IT Actions Only Five Infrastructure Actions Only Fully Optimized Data Center Behavior Dictated by Measure PUE 1.9 1.9 1.5 1.5 Five IT actions have no impact on PUE, only infrastructure actions do. CUPS / Data Center Watt 604 1673 (+177%) 955 (+58%) 2198 (+264%) IT actions have a greater impact on efficiency than infrastructure actions; doing both provides max benefit. Total Facility Power PUE = IT Equipment Power *PUE: Power Usage Effectiveness Data Center Efficiency = Data Center Output Energy Consumed 46
Questions? Anand Ekbote Anand.Ekbote@Emerson.com Vice President, Liebert Monitoring Emerson Network Power