Energy Dumpster Diving - Example presentation- Paper written by: - PowerPoint PPT Presentation

Energy Dumpster Diving - Example presentation- Paper written by: M. A. Kazandjieva, B. Heller, P. Levis, C. Kozyrakis Stanford University In Proceedings of HotPower workshop 2009 Presented by: Simin Nadjm-Tehrani / Klervie Toczé / Rodrigo Moraes Department of Computer and Information Science (IDA) Linköping University, Sweden January 27, 2020

Overview Large computing systems  Individual elements contribute to consumption  Sources of waste difficult to identify  PowerNet sensing infrastructure  Power consumption of individual devices  Correlation with usage information  Analysis of consumption  Reveal device inefficiencies  Usage scenarios that waste energy  1

PowerNet Large-scale distributed sensing infrastructure  Provides  Per-device energy measures  Usage statistics  Deployed in real office building environment  2

PowerNet deployment Stanford Computer Science building  Office environment  Desktops  Monitors  Data center server rack  Small networking closet  Network switches  Some numbers  85 power meters  Utilisation data collected  15 desktops  10 servers  5 switches  3

PowerNet components 4

Wired power meters Watt's Up .NET with  Ethernet interface Disadvantages:  Requires Ethernet port  Difficult to configure  Relatively high power  consumption – 3 W Low sampling rate – 1 Hz  High monetary cost  Similar model shown here 5

Wireless power meter Implemented as a customised mote  Low power processor (1mA active, 1uA sleep)  Digital power meter chip  Characteristics  High rate of sampling – 14KHz  Configurable – TinyOS  No wired network required – mesh network  Lower monetary cost  Mote : resource-constrained device that can sense, process, and talk wirelessly to other motes 6

Mesh network Computer or server Gateway Sink Motes Motes can talk to each other without cables  Data is forwarded to the sink  Mote or computer that gathers data forwarded  Sink is wired to a gateway  Gateway provides out-of-network connectivity  Image from: The Basics of Wireless Sensor Networking and its Applications http://www.ida.liu.se/~rtslab/courses/wsn/Basics.pdf 7

Utilisation metering PowerNet monitors device usage  Desktops and servers  CPU utilisation  Python script tracks utilisation  Network switches  Traffic statistics for each port  Monitor hardware counters via SNMP   Server rack  Balanced workload 8

Data management and visualisation Data stored in a central server  Kept in a MySQL database  Power and utilisation data correlated  Data synchronised in time using timestamps  Analysis of consumption related to activity  Data visualisation through website  Line-chart visualisation of all data  Correlated power and utilisation graphs  Searches by meter name, type, or device category  9

Case study: Desktops High idle energy consumption (100 W)  Reduction  Put desktops in sleep mode when not used  User and CPU must be inactive  10

Case study: Desktops Predict when it is convenient to turn machines off  Machine usage models needed  Correlation between power consumption and CPU usage  11

Case study: Monitors Consumption comparable to desktops (40 – 130 W)  Usage pattern: almost always on, even if not in use  Consumption reduction by configuration parameters  Less brightness, less power consumption  Change desktop backgrounds (10% savings)  Total savings: 10 – 28%  Change of several parameters Change desktop backgrounds 12

Case study: Network Switches Network equipment is not energy proportional  Same energy consumption independently of the usage  Maximum usage  Maximum efficiency  HP switch consumes more due to:  Fan load  Backplane structure  13

Case study: Server rack 10 identical 1U servers in a server rack of 40 servers  Each server consumed 245 W  But server at top of the rack consumed 20% more power  Methodology used for reasoning about odd result  Swap top and bottom servers  Top part is warmer than bottom  Same workload in all of them  Replaced server on top increased from 250 W to 270 W  Previous top server consumption back to normal 245 W  14

Case study: Server rack  We should analyse more aspects than CPU usage alone Load  Temperature  Configuration  15 http://openclipart.org/detail/139525/server-rack-by-moini

Conclusions By analysing power consumption we can:  Reduce energy consumption  Rethink system designs  Power consumption and usage pattern  Needs to be measured/recorded before further savings  Insights revealed by PowerNet:  Monitor configuration can reduce consumption (25%)  Identical server machines can have different power  consumption depending on rack placement Network equipment is not energy proportional  16

Classification ICT Services Production Use End-of-life phase phase Resources phase Residues Recycling Design phase 17

Energy Dumpster Diving Discussion 18

Discussion Which is the cost of maintenance of this system? Cost of  the system itself? (not enough discussed in the paper) 19

Discussion Which is the cost of maintenance of this system? Cost of  the system itself? (not enough discussed in the paper) Do you think this is a permanent infrastructure? Or is just  to do the study? (discussion of key ideas of the paper) 20

Discussion Which is the cost of maintenance of this system? Cost of  the system itself? (not enough discussed in the paper) Do you think this is a permanent infrastructure? Or is just  to do the study? (discussion of key ideas of the paper) Should these meters be installed in all the systems by  default? (possible ideas to improve the work) 21

Discussion Which is the cost of maintenance of this system? Cost of  the system itself? (not enough discussed in the paper) Do you think this is a permanent infrastructure? Or is just  to do the study? (discussion of key ideas of the paper) Should these meters be installed in all the systems by  default? (possible ideas to improve the work) Do you think we should create a standard to declare power  values and utilisation values? (go further from the paper) 22

Discussion Which is the cost of maintenance of this system? Cost of  the system itself? (not enough discussed in the paper) Do you think this is a permanent infrastructure? Or is just  to do the study? (discussion of key ideas of the paper) Should these meters be installed in all the systems by  default? (possible ideas to improve the work) Do you think we should create a standard to declare power  values and utilisation values? (go further from the paper) Suitable for home? (go further from the paper)  23

Web Interface Additional material Powertron  http://powernet.stanford.edu/ 24

Another similar project Additional material The TrendMETER  http://trend.polito.it 25

Discussion The problem of the batteries of the sensors. Which is the  cost of maintenance of this system?(go further from the paper) Do you think this is a permanent infrastructure? Or is just  to do the study? (discuss the basic idea/approach of the paper) Should these meters be installed in all the systems by  default? (how to improve) Do you think we should create a standard to distribute  power values and utilisation values? Suitable for home? (where can apply this methodology...)  Cost of the system itself?  http://www.ida.liu.se/~TDDD50 26

May be a whole slide for TinyOS is too much because distracts the attention from the main topic TinyOS TinyOS is an operating system for resource-constrained  devices It offers you the tools to use the available features of  your hardware Written in nesC, a C dialect  27