Transcript Document
FAC 5.4 Maximizing Energy Efficiency in Data Centers with PAR4® Clemens Pfeiffer Chief Technology Officer 1 FAC 5.4 Maximizing Energy Efficiency in Data Centers with PAR4® This session will show how to overcome the high cost of energy in data centers by creating a test method and a metric for server performance. PAR4 is a comprehensive metric for IT energy efficiency tested by Underwriters Laboratories (UL) that complies with UL standard UL2640 and accurately measures IT equipment power usage. The PAR4 benchmark test suite provides a methodology for accurately measuring server power, consumption and energy efficiency. 2 What is PAR4? - Standardized energy efficiency rating for IT Equipment (Today: Servers only) Rating adjusted based on age of equipment Easy to measure transactions/watt Measures idle and peak power Adopted by UL as UL2640 Adopted by United Nations as part of AM0105 to earn carbon credits Free to download from www.par4.org 3 What do you measure and why? - Transactions per Watt Normalized Efficiency Rating - - Idle power consumption - - Allows to compare servers from multiple generations Degrades with age of the server Minimum power consumed when operating system is running Peak power consumption - Required minimum power allocation when racking equipment Power consumption even at low utilization Typically added to an inventory or change management system for optimal placement 4 Example setup with Starline Raceway How to measure PAR4? By Device By Rack Test Equipment setup with Ethernet cable connecting the raceway and laptop, and a power cable connecting the server. Test Equipment setup with Ethernet cable connecting the racetracks attached to data center racks. 5 An Example • Power consumption Off Idle Peak 7W 134W 255W • Efficiency Rating Green Gold Gold Silver 2010 2011 2012 2013 6 Why is it important? Typical provisioning today by - Size - De-rated power information Example for 10kW rack de-rated - 800W 740W (80%) - 13 Server per rack Example for 10kW rack by PAR4 - 235W peak power - 40 Server per rack Result: 3x capacity per rack 7 How to use PAR4 in your CMS/Asset Mgmt • • • • • • Min/Max values Power allocation Rack placement Inventory integration Utilization verification Aggregated totals 8 Using PAR4 for IT Upgrade Justification HP DL380 G6 Peak Power TPS HP DL360 G8 335W 59M Peak Power TPS 292W 380M Identical Capacity 20 DL380 G6 40U 7,700W 1,180M TPS 3 DL360 G8 3U 876W 1,140M TPS 90% Savings / Space & Power 9 A Real Case Example Peak Power: 335W TPS: 59M Example with about 1 MW of power 3000 DL380 G6 450 DL360 G8 6000U 450U 1 MW 131kW 177 B TPS 171 B TPS Peak Power: 292W TPS: 380M 10 A Warning: Don’t assume Power Consumption 3 Second Expanded Snapshots / TrueRMS 60 cycle/second IDLE 25% 100% PDU SPEC Power Tests: Stacked from 0% to 100% CPU Load 11 Rack Power Consumption 3 Scenarios • Uncapped, Always on • most often used today • Power Capped, Always on • to increase number of servers in the rack • Dynamically Managed • to optimize power cost per transaction 12 Uncapped / Always On 1 Network Switch Web Txn Max: 7,600M Txns Idle: 4,840W – Peak: 8,115W Web Servers 250W 20 Web Servers 165W idle / 292W peak 3 Application Servers 130W idle / 255W peak 2 Database Servers 450W idle / 620W peak Application Servers Database Servers 13 Capped at 200W / Always On 1 Network Switch Web Txn Max: 2,100M Txns Idle: 4,840W – Peak: 6,275W Web Servers 250W 20 Web Servers 165W idle / 292W peak / Capped 200W 3 Application Servers 130W idle / 255W peak 2 Database Servers 450W idle / 620W peak Application Servers Database Servers 14 Power Capping & Performance Power Capping kills performance & increases cost per transaction 400 300 200 100 0 Transactions in Mil 6.00 W/MTxn 4.00 2.00 0.00 No Cap Cap Cap Cap Cap Cap Cap Cap Cap 210 200 190 180 170 160 150 140 No Cap Cap Cap Cap Cap Cap Cap Cap Cap 210 200 190 180 170 160 150 140 Power for 25% Capacity Idle 10% 25% 50% 75% 100% 25% Power independent of use NOTE: Supermicro Dual L5520 CPU 15 Dynamically limited to 2,100M txns 1 Network Switch Web Txn Max: 2,100M Txns Idle: 2,530W – Peak: 4,027W Web Servers 250W 20 Web Servers 165W idle / 292W peak OFF 3 Application Servers 130W idle / 255W peak 2 Database Servers 450W idle / 620W peak Application Servers Database Servers 16 Results Uncapped/Always On Cap 200W/Always On Dynamic 7,600M Txns 2,100M Txns 7,600MTxns Idle 4,840W 4,840W 2,530W Peak 8,115W 6,275W – 23% 4,027W – 50% 2,100M Txns 2,100M Txns 2,100M Txns 5,744W 6,275W 4,027W 2.73W/MTxn 2.99W/MTxn 1.91W/MTxn 10% increase 30% decrease but NO spare capacity and Full capacity Capacity Max Load Max Load Watt W/MTxn Savings per MTxn 17 Other Considerations • • • • Prioritize server use by PAR4 rating Add less efficient capacity as needed Share spare capacity across applications Turn off backup equipment when not needed 18 The impact for a business continuity site Primary DC % Peak 8,115W Secondary DC % Peak 8,115W Max 5,744W OFF Max 5,744W Avg 5,400W Min 5,250W Idle 4,840W Min 5,250W Idle 4,840W Week 1 Week 2 Week 3 Week 4 Week 51 Week Week 2 Week 3 Week 4 Wee 100 Racks each, PUE 2.0: 2MW used $2M per year AFTER: Savings of ~$1Million 19 Next Steps Visit Session FAC 5.2 at 1:45pm to learn about Software Defined Power and its capabilities to manage your IT environment 20 FAC 5.4 Maximizing Energy Efficiency in Data Centers with PAR4® Thank You. To try out PAR4, please visit www.par4.org 21