PROFILING COMPUTER ENERGY CONSUMPTION ON

ORGANIZATIONS

Rui Pedro Lopes, Luis Pires, Tiago Pedrosa and Vasile Marian

Politechnic Institute of Braganca, Braganca, Portugal

Keywords:

Energy saving, Power management, Policy-based consumption optimization.

Abstract:

Modern organizations depend on computers to work. Text processing, CAD, CAM, simulation, statistical

analysis and so on are fundamental for maintaining high degree of productivity and competitiveness. Com-

puters in an organization, consume a considerable percentage of the overall energy and, although a typical

computer provides power saving technologies, such as suspending or hibernating components, this feature can

be disabled. Moreover, the user can opt for never turning off the workstation. Well deﬁned power saving

policies, with appropriate automatic mechanisms to apply them, can provide signiﬁcant power savings with

consequent reduction of the power expense.

With several computers in an organization, it is necessary to build the proﬁle of the energy consumption. We

propose installing a software probe in each computer to instrument the power comsumption, either diretly,

by using a power meter, or indirectly, by measuring the processor performance counters. This distributed

architecture, with software probes in every computer and a centralized server for persistence and decision

making tries to save energy, by deﬁning and applying organization level power saving policies.

1 INTRODUCTION

Many organizations depend heavily on computers

and other information technology equipment to work.

These resources consume a signiﬁcant amount of en-

ergy, contributing to the overall power expense.

Typically, computer equipment act in two roles:

as a server, providing remote access to network ap-

plications, such as network ﬁle systems, web servers,

e-mail and other, or as a desktop workstation, where

the user runs graphical applications and tools. The en-

ergy consumption is different in both situations, and

depend on the energy necessary to power the device,

and the use pattern.

The ﬁrst is conditioned by the technology used

in the computer components, such as the CPU, net-

work interface card, hard disk as well as the moni-

tor. Moreover, Power Management techniques, such

as ACPI (Hewlett-Packard et al., 2006), allow a con-

siderable reduction of power consumption. The use

pattern is related to how the user interacts with the

computer, it’s operating system and applications.

There is a high degree of uncertainty in predicting

the energy consumption for both approaches. Power

Management is not always used, either because the

equipment does not support it or because it is disabled

in the operating system, the user behaviour in relation

to turn-off rates is erratic, and the type and quantity

of equipment in use at the ofﬁce varies randomly.

Because of this uncertainty, the optimization of

computer energy consumption has to act on two axis:

a) enforcing power saving policies and b) appealing

to user behaviour.

2 COMPUTER ENERGY

CONSUMPTION

The energy used by a computer depends on two dif-

ferent aspects: the energy required to run the device

(power draw) and how the computer is used (use pat-

tern) (Webber et al., 2005).

2.1 Power Draw

The power used by a computer is in direct propor-

tion to the energy required by it’s components. The

memory system, composed of the main memory and

system caches consumes a signiﬁcant percentage of

energy, tending to increase with the cache size and

main memory. Several authors address this issue by,

171

Pedro Lopes R., Pires L., Pedrosa T. and Marian V. (2009).

PROFILING COMPUTER ENERGY CONSUMPTION ON ORGANIZATIONS.

In Proceedings of the 11th International Conference on Enterprise Information Systems - Software Agents and Internet Computing, pages 171-174

DOI: 10.5220/0002004801710174

 SciTePress

for example, optimizing the address translation and

compiler techniques for energy saving (Ashok et al.,

2004). Other proposals seek to optimize the power

state of memory as a function of the memory load (Di-

niz et al., 2007) or by transiting devices to low power

after unused for some time (Li et al., 2004).

Similar to the previous case, hard disk drives can

also be forced into a low power state if unused for

some time (Li et al., 2004). Some approaches adapt

the compiler to extend, as much as possible, the time

between disk accesses to allow switching the disk into

a hibernate state, thus saving energy (Hom and Kre-

mer, 2005). The insight into how disk drives use the

energy can also provide valuable information related

to power consumption (Zedlewski et al., 2003).

There is a large variety of computer processors on

the market nowadays and the fact is that every series

that is being produced has a speciﬁc hardware design

and implementation that affects the energy consump-

tion. Designing processors that can perform tasks ef-

ﬁciently without overheating is a major design factor

for many CPU manufacturers. In a desktop computer,

the CPU is one of the prominent energy consumer

component and successfully reducing it can represent

a considerable slice of the overall machine consump-

tion.

The measurement of the consumed energy can be

performed in real-time, either directly (by using a

power meter) or indirectly, by using the CPU’s per-

formance counters (Isci and Martonosi, 2003). With

proper clock speed control associated to an adequate

scheduling of processes, it is possible to reduce the

CPU power consumption (Weiser et al., 1996). Re-

ducing the clock speed allows lowering the supply

voltage, which can signiﬁcantly reduce power dissi-

pation (Hsu and Kremer, 2003).

2.2 Use Pattern

The randomness of the user behaviour in relation to

turn-off rates leads us to consider an automatic proce-

dure to turn off the computer based on speciﬁc poli-

cies. (Webber et al., 2005) shows that users turn-off

rates range from 0% to 91%, for example. On a dif-

ferent study, using stickers to remind the user to turn

off the computer when not in use had some effect.

However, after some weeks, the percentage dimin-

ished (Newsham and Tiller, 1994). Only automatic

approaches were able to maintain the percentage re-

duction.

3 POLICIES

The main idea of Policy-based Consumption Opti-

mization is the deﬁnition of high level procedures –

policies – that will rule the behaviour of the power

management mechanisms. The meaning of policy in

this context is very simple: it is one or more rules that

describe the action(s) to be taken when speciﬁc con-

dition(s) exist. It can be expressed semantically as:

if(condition)then(action)[on(target)] (1)

The condition can include, but not limitted

to, time constraints, calendar constraints, key-

board/mouse activity, processor activity, running pro-

cesses footprint, performance registers content.

actions include actions related to changing the

state of the computer in terms of energy consumption:

turn-off, turn-on (Wake on LAN), suspend, hibernate.

Policies can be structured in global-level and user-

level. The former are of the responsibility of the man-

ager and the latter of the responsibility of each user.

The user must be able to specify that a computer will

never be switched off, regardless of keyboard/mouse

activity, and that the computer will never be switched

off if an application program is running (Bray, 2006).

In other words, users must have the freedom to choose

switch off criteria that is best suited to their own work

patterns (Newsham and Tiller, 1994).

3.1 Power Probes

The accuracy, or the overall quality, of a policy, de-

pends on the use patterns of the organization’s com-

puters. A deep knowledge on how the computer is

used, when it is being used, by which user, which

processes are running, and others, provide valuable

insight into how energy can be saved. To retrieve this

information is necessary a probe (from now on, we

will use the term agent to reference the probe) in each

computer. The agent’s responsibility is to read in-

strumentation parameters related to processes, perfor-

mance registers, keyboard and mouse activity, and to

send it to a central database. Aditionally, the agent is

also responsible for applying policies, thus suspend-

ing, hibernating or shutting-down the computer.

4 ARCHITECTURE

Proﬁling computer energy use in an organization as-

sumes a distributed power management mechanism,

able of gathering the use pattern of several comput-

ers spread over the organization and of applying poli-

cies for energy saving. Each computer has a soft-

ware agent, to perform instrumentation on the CPU’s

ICEIS 2009 - International Conference on Enterprise Information Systems

172

performance counters, keyboard and mouse activity,

building the running processes and applications map

and enforcing actions. Moreover, the agent can access

mean and instant values for voltage, current, power

and energy, provided by a hardware device, accessi-

ble through USB. The instrumentation data is sent to

a centralized server, for persistence and data mining.

The hardware device is custom made and is in-

stalled in serial with the power supply. It has several

internal registers, which can be read through USB.

These values are used to correlate with the instrumen-

tation values of the performance counters. This will

allow to get the consumption scenario in several com-

puters using a single power meter.

The communication between the agents and the

central server is performed through Web Services,

providing multi-platform and multi-language services

for persistence, queries and management.The server

also provides a Web based interface, to allow the def-

inition, modiﬁcation and removal of power saving

policies.

4.1 Power Probe – Agent

From a power consumption point of view, computers

are intrinsically heterogeneous. Differences in CPU

technology, manufacturer, graphics, number of hard

disk drives and so on make the process of proﬁling

power consumption a difﬁcult task. Moreover, with-

out solid information about all the computers in the

organization it is not possible to act, in terms of en-

forcing power saving policies.

Agents provide a common interface to obtain in-

strumentation values of the computer as well as pro-

viding a means to act on it. In other words, each

agent periodically sends information related to the

computer power consumption and enforces policies

received from the central server.

Each agent has a set of sensors, modules respon-

sible for getting and processing samples. For the pur-

pose of the work presented in this paper, sensors mon-

itor running processes, perform periodic samples on

CPU performance counters, receive indication of key-

board and mouse activity, perform samples on the ex-

ternal power meter. Sensors also havea policy engine,

responsible for running and enforcing policies.

The communication between agents and the pol-

icy server is performed through Web Services. This

provides a interoperable communication mechanism,

independent of the operating system, network tech-

nology and languages used.

4.2 Policy Server

The policy server is a central repository for power

saving policies and power consumption history. It

also provides a Web based user interface, to allow

deﬁning, updating and removing policies. It is de-

veloped in a multitier approach, isolating the persis-

tence layer from the presentation layer with a business

layer.

The business layer exports Web Services based

interfaces, to allow remote operation on the server.

Each module is encapsulated in an Enterprise Java

Bean (EJB), preventing a direct access to the data in

the database. In the work presented in this paper, the

following modules are available: user and resources

managament, policy deﬁnition, samples repository.

The user and resources management is related to

the authentication and authorization of users in the or-

ganization. Each user is associated with an access list,

which describe the resources he is allowed to control

through the deﬁnition of policies. If a user tries to de-

ﬁne a policy for a computer he has no authorization,

the operation fails and the attempt is logged, for later

inspection.

The policy deﬁnition module allows the user to

write policies to the resources which he has access to.

The policy, after validated, is stored in the database,

through the persistence layer. Then the user has the

possibility of activating it, which translates to sending

the policy to the target agents. Deactivating the policy

implies the removalof the agent in which it is running.

Finally, the samples module is responsible for re-

ceiving and storing the samples from the agents in-

stalled in the organization’s computers. It also pro-

vides some statistics to the authorized users, to allow

optimizing policies according to the power saving po-

tential.

4.3 User Behaviour

The most important factor for saving energy is user

behaviour, such as turning off devices at night or

enabling power management (Newsham and Tiller,

1994; Webber et al., 2005). By providing adequate

information as well as adequate tools, which can au-

tomate many of the user actions, it is possible to inﬂu-

ence energy consumption reduction to a greate extent.

Each agent has the possibility to popup a message,

providing non-intrusive information about the power

saving history in his computer. The most simple mes-

sage is “You do not havepower managementactivated

in this computer. Do you want to turn it on? It would

not have any impact on your work.”. A more elab-

orate example: “Your computer will disconnect one

PROFILING COMPUTER ENERGY CONSUMPTION ON ORGANIZATIONS

173

hour after 8 p.m. However, the computer is usually

idle after 5 p.m. Changing the policy would save 3

hours of energy consumption”.

We believe that a discrete number of messages for

month would appeal effectivelly to the user behaviour,

thus contributing to reducing the overall expense.

5 CONCLUSIONS

Ofﬁce equipment, in particular computers, demand a

signiﬁcant amount of power, contributing to the over-

all energy consumption. Proﬁling the power con-

sumption of an organization require a distributed ar-

chitecture, based on remote agents and a centralized

server. The agents are responsible for instrumenta-

tion of computers, and the central server is responsi-

ble for storing and processing the information sent by

the agents. Based on the energy proﬁle, it is possible

to deﬁne power saving policies, rules of behaviour for

power management mechanisms, to be enforced by

each agent.

The information from all the agents provide a

valuable insight into how computers all over the or-

ganization use energy, as well as an opportunity to

optimize power consumption based on the typical de-

mand.

In this paper, we present and architecture for pro-

ﬁling and enforcing policies. It is based on a multitier

application server with a Web frontend, and a web

services middleware, to cope with the diversity and

heterogeneity of enterprises’ networks and systems.

REFERENCES

Ashok, R., Chheda, S., and Moritz, C. A. (2004). Cou-

pling compiler-enabled and conventional memory ac-

cessing for energy efﬁciency. ACM Trans. Comput.

Syst., 22:180–213.

Bray, M. (2006). Review of computer energy consumption

and potential savings.

Diniz, B., Guedes, D., Meira, J. W., and Bianchini, R.

(2007). Limiting the power consumption of main

memory. In Proceedings of the 34th annual inter-

national symposium on Computer architecture, pages

290–301, San Diego, California, USA. ACM.

Hewlett-Packard, Microsoft, Intel, Technologies, P., and

Toshiba (2006). Advanced conﬁguration and power

interface speciﬁcation.

Hom, J. and Kremer, U. (2005). Inter-program optimiza-

tions for conserving disk energy. In Proceedings of

the 2005 international symposium on Low power elec-

tronics and design, pages 335–338, San Diego, CA,

USA. ACM.

Hsu, C. H. and Kremer, U. (2003). The design, implemen-

tation, and evaluation of a compiler algorithm for cpu

energy reduction. ACM SIGPLAN Notices, 38:38–48.

Isci, C. and Martonosi, M. (2003). Runtime power moni-

toring in high-end processors: Methodology and em-

pirical data. In Proceedings of the 36th annual

IEEE/ACM International Symposium on Microarchi-

tecture. IEEE Computer Society Washington, DC,

USA.

Li, X., Li, Z., David, F., Zhou, P., Zhou, Y., Adve, S., and

Kumar, S. (2004). Performance directed energy man-

agement for main memory and disks. In Proceed-

ings of the 11th international conference on Archi-

tectural support for programming languages and op-

erating systems, pages 271–283, Boston, MA, USA.

ACM.

Newsham, G. and Tiller, D. (1994). The energy consump-

tion of desktop computers: measurement and savings

potential. Industry Applications, IEEE Transactions

on, 30:1065–1072.

Webber, C. A., Roberson, J. A., McWhinney, M. C., Brown,

R. E., Pinckard, M. J., and Busch, J. F. (2005). After-

hours power status of ofﬁce equipment in the usa.

Journal: Energy (the International Journal); Journal

Volume: 31; Journal Issue: 14; Related Information:

Journal Publication Date: 11/2006.

Weiser, M., Welch, B., Demers, A., and Shenker, S. (1996).

Scheduling for reduced cpu energy. Kluwer Interna-

tional Series in Engineering and Computer Science,

pages 449–472.

Zedlewski, J., Sobti, S., Garg, N., Zheng, F., Krishna-

murthy, A., and Wang, R. (2003). Modeling hard-disk

power consumption. FAST2003 - 2nd USENIX Con-

ference on File and Storage Technologies.

ICEIS 2009 - International Conference on Enterprise Information Systems

174