[COMPUTER ARCHITECTURE] Final Presentation_Spring 2014
Download as PPTX, PDF0 likes220 views
This document discusses power provisioning for large computer systems like data centers. It first shows the hierarchy for distributing power in a data center and issues that can lead to inefficient power usage like staged deployments and variable loads. It then analyzes the power usage profiles of different server components and workloads like web search, finding that actual peak power often differs from nameplate ratings. Finally, it explores approaches for improving power efficiency, such as CPU voltage/frequency scaling and reducing idle power consumption.
[COMPUTER ARCHITECTURE] Final Presentation_Spring 2014
- 1. Power Provisioning for a Warehouse-Sized Computer Christin Panjaitan M10202815
- 2. 1. Data Center Power Provisioning Fig 1. Simplified datacenter power distribution hierarchy 2
- 3. Inefficient use of the power budget Staged Deployment Fragmentation Conservative Equipment Ratings Variable Load Statistical Effects 3
- 4. 2. Power Estimation • Observe power usage profile of a typical server Table 1. Component peak power breakdown for a typical server 4
- 5. Nameplate vs Actual Peak Power Nameplate indicates maximum power draw of that machine. Nameplate is important to supply power to the machine. According to their benchmark, the maximum is 145 W System Total 213 W Power Supply Efficient 85 % Nameplate : 251 W 5
- 6. 3. Power Usage Characterization Baseline characterization of the power usage of three scale workloads : • Websearch • Webmail • Mapreduce Evaluation : Set of servers selected are running well-tuned workloads above and typically at high activity levels. Rack = 40 machines PDU = 20 racks (20 x 40 = 800 machines) Cluster = 5000 machines 6
- 7. Fig 2. Websearch-CDF of power usage normalized to actual peak 7
- 8. Fig 3. Webmail -CDF of power usage normalized to actual peak 8
- 9. Fig 4. Mapreduce -CDF of power usage normalized to actual peak 9
- 10. Fig 5. CDF of a Real Datacenter 10
- 11. Power Savings Approaches 1. CPU Voltage / Frequency Scalling CPU Voltage and frequency scaling (DVS) is a technique to manage energy consumption. Fig 6. Impact of CPU DVS at Datacenter Level 11
- 12. 2. Improving Non-Peak Power Efficiency Fig 7. Idle power as fraction of peak power in 5 server configurations 12
- 13. Fig 8. Power and Energy Savings Achievable by reducing 13