USER GUIDANCE OF RESOURCE-ADAPTIVE SYSTEMS

João Pedro Sousa

Computer Science Department, George Mason University, 4400 University Drive, Fairfax VA, U.S.A.

Rajesh Krishna Balan

School of information Systems, Singapore Management University, 80 Stamford Road, Singapore

Vahe Poladian, David Garlan, Mahadev Satyanarayanan

Computer Science Department, Carnegie Mellon University, 5000 Forbes Avenue, Pittsburgh PA, U.S.A.

Keywords: Mobile Computing, Resource Adaptation, Self-adaptive Systems, Software Architecture.

Abstract: This paper presents a framework for engineering resource-adaptive software systems targeted at small

mobile devices. The proposed framework empowers users to control tradeoffs among a rich set of service-

specific aspects of quality of service. After motivating the problem, the paper proposes a model for

capturing user preferences with respect to quality of service, and illustrates prototype user interfaces to elicit

such models. The paper then describes the extensions and integration work made to accommodate the

proposed framework on top of an existing software infrastructure for ubiquitous computing.

The research question addressed here is the feasibility of coordinating resource allocation and adaptation

policies in a way that end-users can understand and control in real time. The evaluation covered both

systems and the usability perspectives, the latter by means of a user study. The contributions of this work

are: first, a set of design guidelines for resource-adaptive systems, including APIs for integrating new

applications; second, a concrete infrastructure that implements the guidelines. And third, a way to model

quality of service tradeoffs based on utility theory, which our research indicates end-users with diverse

backgrounds are able to leverage for guiding the adaptive behaviors towards activity-specific quality goals.

1 INTRODUCTION

Sophisticated software is increasingly being

deployed on small mobile devices, taking advantage

of their growing capabilities and popularity. Media

streaming is already found frequently in PDAs and

high-end cell phones. Soon, applications such as

speech recognition, natural language translation, and

virtual/augmented reality may leap from research

prototypes to widespread commercial use.

While software has enjoyed plentiful and stable

resources in the world of desktops (and to some

extent, of laptops,) resource variation needs to be

taken into account in smaller devices. Despite the

impressive capabilities of today’s mobile devices,

user expectations with respect to performance and

sophistication will continue to be set by the full-size

versions running on powerful desktops and servers.

Research in resource-adaptive applications takes

an important step towards addressing resource

limitation and variation (De Lara, Wallach, &

Zwaenepoel, 2001; Flinn & Satyanarayanan, 1999;

Yuan, Nahrstedt, Adve, Jones, & Kravets, 2006).

However, existing solutions either enforce

predetermined policies, or offer limited mechanisms

to control the application’s policies. In some cases,

the adaptation mechanisms focus strictly on network

conditions, enforcing policies that are established by

system designers before the system is deployed. In

other cases, users are offered limited control over the

policies, typically focusing on a single aspect of

quality of service, such as battery duration.

Unfortunately, those limitations prevent adaptive

systems from addressing two important issues. First,

user goals often entail tradeoffs among different

aspects of quality. For example, in the presence of

limited bandwidth, should a web browser skip

Pedro Sousa J., Krishna Balan R., Poladian V., Garlan D. and Satyanarayanan M. (2008).

USER GUIDANCE OF RESOURCE-ADAPTIVE SYSTEMS.

In Proceedings of the Third International Conference on Software and Data Technologies - SE/GSDCA/MUSE, pages 36-44

DOI: 10.5220/0001881500360044

 SciTePress

loading pictures in order to provide faster load

times? For browsing restaurant listings, a user may

prefer dropping images to improve load times; but

for browsing online driving directions, the user may

be willing to wait longer for the full page content.

Second, user activities may involve more than

one application, making it desirable to coordinate

resource usage and adaptation policies across

applications. For example, an activity that involves

simultaneous video streaming and downloading

email attachments may be best served when video

streaming consistently uses 80% of the bandwidth

and email does not attempt to go beyond 20%.

This paper presents a framework for engineering

resource-adaptive systems that: (a) empower users to

control tradeoffs among a rich set of aspects of

quality, and (b) coordinate resource usage among

several applications. To develop such a framework,

important questions need to be answered: how to

represent user preferences in a way that can be used

to guide adaptation policies? How to elicit such

preferences? How to allocate resources among

applications, and how to coordinate their policies?

What APIs must applications expose to be amenable

to such coordination?

In the remainder of this paper, Section 2

proposes a model for capturing user preferences with

respect to quality of service, and illustrates prototype

user interfaces to elicit such models. The

implementation of the proposed framework builds

on an existing infrastructure for ubiquitous

computing (Garlan, Siewiorek, Smailagic, &

Steenkiste, 2002), which is summarized in Section 3.

Section 4 describes the extensions and integra-

tion work we made to accommodate the proposed

framework on top of the existing infrastructure.

Specifically, these extensions include: (a) interaction

protocols for coordinating resource allocation and

adaptation policies; and (b) guidelines for the

integration of adaptive applications into the

proposed framework.

Section 5 summarizes the evaluation of the

proposed solution from a systems perspective, and

describes a preliminary evaluation of usability, by

means of a user study. The results of the study

indicate that end-users with diverse backgrounds can

understand and use the proposed models of quality

of service to control the adaptive behavior of

applications towards specific goals.

Section 6 discusses related work, and Section 7

summarizes the main points of this paper.

2 USER PREFERENCES

Any adaptation or optimization process is guided by

a goal. In the case of adapting to resources in small

mobile devices, the goal is to optimize the quality of

service (QoS) perceived by the user. Work in this

area frequently addresses conserving resources, such

as battery charge, but that is just one way to

optimize for service duration, an aspect of QoS.

The conceptual framework that we adopt takes

into account that:

(1) Users may care about tradeoffs between

different aspects of QoS; e.g., latency vs. accuracy.

(2) Different services may be characterized by

different aspects of QoS. For example, for web

browsing, users may care about load times and

whether the full content is loaded (e.g., pictures); for

automatic translation, users may care about the

response time and accuracy of translation; for

watching a movie, users may care about the frame

rate and image quality.

(3) User preferences for the same service may

depend on the user’s activity. For example, a user

may prefer high frame rate over image quality for

watching a sports event over a network connection

with limited bandwidth, but might prefer the

opposite for watching a show on sculpture.

However, important questions cannot be

answered with this approach: for instance, how short

of a response time will satiate the user? And even if

accuracy is less important, what if it degrades so

much that the translations become unusable?

At the other end of the spectrum, preferences

may be expressed as an arbitrary function between

the multivariate quality space and a utility space

representing user happiness. For instance, the user

might indicate that he would be happy with medium

translation accuracy, as long as latency remains

under 1 second, and that he will be happy to wait 5

seconds for highly accurate translations. Although

fully expressive, designing mechanisms to elicit this

form of preferences from end-users is a hard

problem, and even more so if more than two aspects

of QoS are involved.

The model we propose sits between these two

extremes. User preferences are expressed as

independent utility functions for each aspect, or

dimension, of QoS. Such functions map the possible

quality levels in the dimension to a normalized

utility space U Þ [0,1], where the user is happy with

utility values close to 1, and unhappy with utility

values close to zero.

USER GUIDANCE OF RESOURCE-ADAPTIVE SYSTEMS

Figure 1: QoS preferences for a language translation service.

For each continuous QoS dimension the user

indicates two values: the thresholds of satiation and

of starvation. For example, the user might be happy

with response times anywhere under 3 second, but

may not accept response times over 20 seconds.

This is illustrated in

Figure 1b, where the thresholds

of satiation and starvation are represented by the

green (lighter) and red (darker) lines, respectively.

Currently, we use sigmoid functions to smoothly

interpolate between these two zones, the thresholds

marking the knees of the sigmoid. The utility

corresponding to each value of latency is indicated

by the scale at the top, ranging from a happy face

(

) for values beyond the satiation threshold, all the

way down to a cross (

), representing rejection, for

values beyond the starvation threshold.

Preferences for discrete QoS dimensions are

represented using a discrete mapping to the utility

space.

Figure 1c shows an example where a table

indicates the utility of each level of accuracy.

The functions for each aspect of QoS are then

combined by multiplication, which corresponds to

an and semantics: a user is happy with the overall

result only if he is happy with the quality along each

and every dimension. Whenever a user task in-

volves more than one service, the overall utility

combines the QoS dimensions for all the services.

The relative importance of each aspect, modeled

as a weight w∈[0,1], is factored into the combined

utility. For example, for two aspects a and b, the

combined utility function is

uu .

. These weights

take the value 1 by default, but may be altered using

the slider bars on the right side in

Figure 1a-c.

To make it easier to use this model, we include

the notion of preference templates. This decision is

based on the principle of offering incremental

benefit for incremental effort, also known as gentle

slope systems (Myers, Smith, & Horn, 1992).

Figure

1a shows an example with two templates, fast and

accurate. If a template is selected, the associated

preferences are shown. In case a user wishes to fine-

tune these preferences, he may do so after selecting

the custom checkbox (

Figure 1b-c).

3 ARCHITECTURE BASELINE

The models of preferred QoS tradeoffs described in

Section 2 can be leveraged using two alternative

architectural strategies. Either individual applica-

tions are responsible for capturing and managing the

models, or the features required to do so are factored

out into a common infrastructure.

The latter approach has significant advantages in

terms of reuse. In addition to promoting the reuse

(avoiding replication) of code to support those

features across applications, there is also the reuse of

the knowledge about user preferences. For example,

once the preferred QoS tradeoffs for watching a

specific video stream are elicited from the user, that

knowledge resides with the infrastructure and can be

passed to the streaming application running on the

device that happens to be convenient to the user at

each moment: a cell phone, a laptop, etc.

Therefore, the architectural strategy we adopted

is to define a software infrastructure that: (a)

captures models of QoS tradeoffs, (b) coordinates

the resource usage across the applications supporting

the user’s activity, if more than one is involved, and

their adaptation policies based on the QoS models.

Rather than building such an infrastructure from

scratch, we extended an existing infrastructure

developed at Carnegie Mellon’s Project Aura, which

targets user mobility in ubiquitous computing

environments (Sousa, 2005). The remainder of this

(a) (b) (c)

ICSOFT 2008 - International Conference on Software and Data Technologies

section summarizes the Aura infrastructure, as well

as an existing library for resource adaptation,

Chroma, also related to Project Aura (Balan, Gergle,

Satyanarayanan, & Herbsleb, 2005).

Figure 2: The Aura infrastructure.

Aura supports a high-level notion of user

activities, such as preparing presentations or writing

film reviews. Such tasks may involve several

services. For instance, for preparing a presentation,

a user may edit slides, refer to a couple of papers on

the topic, check previous related presentations, and

browse the web for new developments.

Figure 2 shows a component and connector view

of the Aura infrastructure. The Prism component

captures and maintains models of user activities.

Specifically, each model enumerates the services

required to support the activity, how those services

are interconnected, if at all, preferences with the

respect to the kinds of applications to provide each

service (e.g., Emacs as opposed to vi for editing

text,) and service-specific settings.

The Environment Manager (EM) component

keeps track of the availability of services within an

environment. An environment in Aura refers to the

set of devices, software components and other

resources accessible to a user at a particular location.

Whenever a user indicates that he or she whishes

to start or resume an activity, Prism communicates

the corresponding activity model to the EM using

the service request protocol (SRP), and the two

components negotiate the configuration that best

supports the user’s needs and preferences. Once an

agreement is reached, the EM communicates with

the applications using the service announcement and

activation protocol (SAAP) to activate the services

and make the required interconnections, if any. After

that it passes a model of the concrete configuration

up to Prism (SRP). Prism uses this model to

communicate with the applications using the service

use protocol (SUP) and recover the preferred

settings for the activity; for example, the point at

which the user was previously watching a video.

The Aura connectors (SAAP, SRP, and SUP),

support the asynchronous exchange of XML

messages over TCP/IP. These are peer-to-peer

protocols, where each component may initiate

communication, as needed.

Figure 3: Integration of Chroma.

Chroma enables conventional applications to be

enhanced for adaptation, provided the applications

can carry out their operations using different tactics.

For example, a speech recognizer may have more

costly algorithms that deliver better results, or

simpler algorithms that demand fewer resources.

Additionally, Chroma supports the partitioning of

applications, shipping and running heavy

computations in remote servers when the available

resources, such as bandwidth, favor that option.

Chroma includes generic, application-independ-

ent parts: a Solver, a resource demand predictor, and

a number of resource monitors. The latter include

history-based monitors of available bandwidth,

battery charge, CPU and memory, both on the local

device and on remote servers (Narayanan, Flinn, &

Satyanarayanan, 2000). The resource demand

predictor forecasts the resource demand of each

tactic based on historical averages of actual demand.

The Solver determines the tactic with the highest

utility, given the available resources, by exhaustive

evaluation of all the tactics defined for the

application. The Solver is invoked by the applica-

tion before carrying out each unit of work; for

example, before recognizing each utterance, in the

case of speech recognition, or before rendering each

frame, in the case of virtual reality applications.

Figure 3 shows these parts and their interactions:

the thin arrows within Chroma represent information

flow as a result of method calls.

4 IMPLEMENTATION

The research in this paper involved extending the

Prism and EM components in Aura, as well as

integrating Chroma with the Aura protocols and with

the QoS models described in Section 2.

Prism

SRP

app 1 app N

SAAP

…

SUP

Chroma

Prism

adaptive app

Solver

rsrc availability rsrc

resource demand

predictor

resource

monitors

rsrc

rofiles

allocation

QoS models

rsrc usage

settin

USER GUIDANCE OF RESOURCE-ADAPTIVE SYSTEMS

Figure 6: Example resource allocation.

The Aura protocols were also extended to

include the flow of QoS models to the EM, over the

SRP, and to Chroma, over the SUP; as well as the

flow of resource information between the EM and

Chroma, over the SAAP. These flows are

represented as thicker arrows in Figure 3,

corresponding to the protocols in

Figure 2.

Prism is now in charge of capturing and

disseminating QoS models.

Figure 1 illustrates the

interfaces for eliciting user preferences for a service

with two aspects of QoS: latency (response time)

and accuracy. These models are disseminated to

other components in the format illustrated in Figure .

The use of XML as opposed to language-specific

data structures makes the models easier to exchange

between components written in different languages.

Prism creates user interfaces like the one in

Figure 1

dynamically, based on the QoS dimensions and

values in the XML representation of a model.

The EM is now in charge of determining and

disseminating the optimal resource allocations

among the applications supporting the user’s

activity. Figure 6 shows an example of resource

constraints that the EM may send to one application

via the SAAP. To determine those, the EM receives

(a) resource profiles via the SAAP, Figure , relating

the quality levels that each application can operate at

with the corresponding resource demands. (b) QoS

models via the SRP, Figure ; and (c) forecasts of

resource availability, which, unlike fine-grained

forecasts for adaptive applications, contain averages

over a large number of historical samples. Optimal

allocation uses an efficient global optimization

algorithm, which was published as a separate

research result (Anonymous, 2006).

Chroma is in charge of supporting the adaptation

policies within each application. Architecturally,

adaptive applications are built on top of the Chroma

library, and there is one instance of Chroma,

customized with application-specific tactics,

deployed with each application. Integrating such

applications involved wrapping them to mediate

between the Aura protocols and the Chroma APIs.

Since Chroma expects a generic utility function for

the Solver, plugging in a function that interprets the

QoS models passed via the SUP (Figure ) was fairly

straightforward.

5 EVALUATION

The proposed framework was evaluated both from

the systems and from the usability perspectives.

From the systems perspective, we verified that

the proposed solution makes optimal adaptation

decisions, and that it does so efficiently, that is,

quickly and without consuming significant resources

itself. (Balan, Satyanarayanan, Park, & Okoshi,

2003) tested Chroma running language translation

and speech and face recognition applications on a

PDA under a wide range both of available resources

and of user preferences. These tests verified that the

Solver consistently picks the tactic that delivers the

highest utility under the available resources.

In (Anonymous, 2006) we evaluated the

efficiency of EM’s resource allocation running on a

1.6 GHz CPU with 512 MB of RAM. The latency

of finding the optimal allocation is 200 ms on

average (standard deviation 50 ms) for configura-

tions requiring from 1 to 4 services, when 4 to 24

alternative suites of application are available to

provide those services, and when the search space of

combined QoS levels reaches up to 15,000 points.

Reevaluating the resource allocation every 5s, the

<thresholds good="3" bad="20"

unit="second"/>

</function>

</QoSdimension>

</function>

</QoSdimension>

</utility>

Figure 4: Representation of the preferences in Figure 1.

<head>latency accuracy cpu bdwdth</head>

<units>second none % Kbps</units>

</QoSprofile>

</service>

Figure 5: Example QoS profile.

<rsrc id="bdwdth" avg="800" var="100"

u="Kbps"/>

</constraints>

ICSOFT 2008 - International Conference on Software and Data Technologies

EM uses on average 3% of CPU cycles. The

optimality of decisions was verified analytically.

The remainder of this section focuses on the

evaluation of usability. For that, three criteria were

considered: the expressiveness of the QoS models,

the ease of eliciting them, and the ease of using them

to control adaptation. With respect to expressive-

ness, our experience with multiple examples, some

illustrated in the user study described below,

indicate that the proposed models are expressive

enough for a wide range of practical situations.

A user study investigated whether end-users can

express their preferences and control adaptation

using the proposed QoS models.

This study consists of using a natural language

translator running on a mobile device. The quality

of translation observed by users varies, since the

translator runs either simple algorithms locally, or

more sophisticated ones on a remote server,

depending on the availability of bandwidth and of

capacity in the server. To prevent limitations in the

capabilities of the actual translation application

(limited dictionaries, etc.) from affecting the results

of the study, we replaced a human for the remote

translation server. This technique is well accepted

and known as a Wizard of Oz experiment.

The study focused on answering the following

questions: first, can users understand and use

templates to achieve a goal? Second, can users think

of and manipulate preferences in terms of

thresholds? Third, do they find it easy? And fourth,

can users interpret the effects of specifying different

preferences in the application’s adaptive behavior?

The participants were drawn from a population

with homogeneous education level and age group,

but diverse technical background. Ten students in

the age group 18-29 were drawn among the

respondents to a posting, 5 of which from

computing-related fields (computer science,

electrical and computer engineering, logic) and the

remaining 5 from other fields (business, physics,

literature). Incidentally, 6 were male and 4 female.

Participants individually performed an

experiment that lasted 30 minutes, after being given

a 30 minute introduction to the experiment,

methodology and tools. Participants were asked to

follow the think aloud protocol (Steinberg, 1991),

and their voice and actions on the screen were

recorded using video capturing software

(TechSmith). After the experiment, the participants

completed a short questionnaire.

The scenario for the experiment revolved around

a conversation with a foreign language speaker

(Spanish in this case) aided by translation software.

To prevent serious misunderstandings in a real

situation, users of the translation software would be

able to check the accuracy of translation by having

the Spanish translation translated back to English

and spoken (using speech synthesis) on the user’s

earphones. Users would press a go-ahead button to

synthesize the Spanish translation only if they were

happy with the accuracy of translation.

During the experiment, participants were asked

to input sentences of their own making, listen to the

output of the double translation, and rate the

accuracy. The training included calibrating the

participants' rating of accuracy using the following

scale: high, if the meaning is fully preserved;

medium, if the meaning is roughly preserved; and

low, if the meaning is seriously distorted.

Participants were asked to pursue different QoS

goals during each part of a three-part experiment.

Within each part, we simulated resource variation

and asked the participants to evaluate the changes

both in latency and accuracy of translation. During

the first two parts, the QoS goals could be satisfied

by preference templates. During the third part, the

specific goal could only be achieved by customized

preferences. The participants were not directed as to

whether or not to use templates in any case.

12345

12 3 4 5

(a) Templates (b) Thresholds

Figure 7: Likert scale evaluation of preferences’

specification (5-fully favorable, to 1-unfavorable).

Whenever the QoS goals could be met by a

template, the participants did use templates in 17 out

of 20 cases. In the remaining 3 instances, the

participants were still able to achieve the goals using

customized preferences. When asked about the

clarity and usefulness of templates, 8 participants

were fully favorable, while 2 didn’t recognize a

benefit in having templates – see Figure 7a.

All 10 participants were able to manipulate the

thresholds in customized preferences for achieving

the required QoS goals. Specifically, the experiment

was set in such a way that the thresholds in one

dimension needed to be made stricter, while relaxing

the other dimension, under penalty of the goal not

being achievable.

When asked about the clarity of using thresholds

to specify preferences, 8 participants were fully

favorable, while 2 thought some alternative strategy

USER GUIDANCE OF RESOURCE-ADAPTIVE SYSTEMS

could be preferable – see Figure 7b. One of these

participants suggested that an X-Y representation the

tradeoff might be clearer. However, there are two

reasons why that may not be such a good idea. First,

it would be hard to show and manipulate tradeoffs

with more than two aspects of QoS. Second, the

actual tradeoff changes with the availability of

resources: with plentiful resources, high levels may

be attainable along all aspects; but with low

resources, to privilege one aspect may have a severe

impact on others.

The participants were able to interpret the effects

of different preferences in the application’s adaptive

behavior. To verify this, we tested the hypothesis

that when resources change participants perceive a

change in the QoS, with a greater impact along the

QoS dimension for which the preferences are laxer.

For that, after each translation the participants

evaluated which QoS dimension changed the most

relative to the previous translation: a noticeable

change in accuracy with similar latencies, a

noticeable change in latency with similar accuracies,

no noticeable changes, etc. Participants then related

those changes with the strictness or laxness of the

preferences along each QoS dimension. The

participants were not informed of when or in which

direction resources would change.

Figure 8: Regression performed on experiment data.

Figure 8 shows the results of correlating which

dimension had stricter preferences with which

dimension was perceived to have changed the most.

The correlation coefficient is negative, meaning that

whenever user preferences were stricter along one

dimension, the participants perceived a greater

fluctuation on the other dimension (caused by

underlying resource fluctuations). When asked

about how easy it was to use the interfaces in

Figure

1 to customize preferences, 5 participants were fully

favorable while the other 5 thought the interfaces

could be improved.

This user study demonstrates that end-users can

both define their preferences, and interpret the

results of such definitions in the system’s adaptive

behavior. A control loop is therefore formed,

enabling users to pursue concrete QoS goals. The

practicality of the control loop is confirmed by the

fact that all participants were easily able to achieve

concrete QoS goals.

6 RELATED WORK

Similarly to the proposed framework, others have

leveraged techniques from microeconomics to elicit

utility with respect to multiple attributes. In the

Security Attribute Evaluation Method (SAEM), the

aggregate threat index and the losses from successful

attacks are computed using utility functions (Butler,

2002). The Cost Benefit Analysis Method (CBAM)

uses a multidimensional utility function with respect

to QoS for evaluating software architecture alterna-

tives (Moore, Kazman, Klein, & Asundi, 2003). Our

work is different from SAEM and CBAM in that it

is geared towards mobile computing.

A body of work addressed battery duration in

mobile devices. For example (Yuan et al., 2006),

presented OS extensions that coordinate CPU

operation, OS scheduling, and media rendering, to

optimize device performance, given user preferences

concerning battery duration. The QoS models in our

framework are significantly more expressive, since

they support a rich vocabulary of service-specific

aspects of QoS.

User studies done in mid-to-late 1990s have

demonstrated that stability (e.g., absence of jitter) is

more important than improvement for certain

aspects of QoS (Wijesekera, Varadarajan, Parikh,

Srivastava, & Nerode, 1998). Our framework recog-

nizes the importance of these results and ensures, by

explicit resource allocation, that adequate resources

are available for applications to provide service

while maximizing the overall utility.

Dynamic resolution of resource allocation policy

conflicts involving multiple mobile users is

Correlation Coefficient t-value Significant at 95%

-0.6 -4.27 Yes

How to interpret a correlation: the correlation

coefficient denotes the slope of the line that best fits the

data. A positive/negative coefficient means that an

increase in the x-axis corresponds to an increase/decrease

in the y-axis. If the coefficient is zero, the data cannot be

approximated by a straight line (there is no correlation

between the x values and the y values).

Student's t-test of significance: indicates the

likelihood that the correlation in the data sample

corresponds to a real correlation in the general

population. A commonly accepted threshold is 95%

confidence. Statistics manuals contain tables of t-

statistics for each size of the data sample. The t-test

consists of comparing the t-value calculated for the

correlation with the lookup t-statistic. If the absolute t-

value is larger than the t-statistic, then the correlation is

significant with 95% certainty.

Sample: 40 data points relating two variables (38

degrees of freedom), for which the t-statistic is 2.024 for a

95% confidence.

ICSOFT 2008 - International Conference on Software and Data Technologies

addressed in (Capra, Emmerich, & Mascolo, 2003)

using sealed bid auctions. While this work shares

utility-theoretic concepts with our configuration

mechanisms, the problem we solve is different. Our

work has no game-theoretic aspects and addresses

resource contention by multiple applications

working for the same user on a small mobile device.

From an analytical point of view, closest to our

resource allocation algorithm are Q-RAM (Lee,

Lehoczky, Siewiorek, Rajkumar, & Hansen, 1999),

Knapsack algorithms, and winner determination in

combinatorial auctions. By integrating with generic

service discovery mechanisms in the EM, our work

provides an integrated framework for service

discovery, resource allocation and adaptation.

7 CONCLUSIONS

Resource adaptation can play an important role in

improving user satisfaction with respect to running

sophisticated software on small mobile devices.

However, today, many applications implement

limited solutions for resource adaptation, or none at

all. The primary reasons for that are: (a) the cost of

creating ad-hoc adaptation solutions from scratch for

each application; and (b) the difficulty of

coordinating resource usage among the applications.

Because it is hard for an individual application to

even know which other applications are actively

involved in supporting a user activity, individual

applications frequently trample each other in their

quest for resources.

This paper proposes a framework for resource

adaptation where a number of features are factored

out of applications into a common infrastructure.

First, user preferences with respect to overall

QoS tradeoffs are elicited by an infrastructural

component such as Prism. These models are

expressed using a rich vocabulary of service-specific

QoS aspects. Furthermore, a preliminary user study

indicates that end-users can understand and leverage

these models to pursue concrete QoS goals.

Second, resource allocation among applications

is coordinated by another infrastructural component

such as the EM. This component receives QoS

profiles from applications, and efficiently computes

the resource allocations that optimally support the

QoS goals, given forecasts of available resources for

the next few seconds.

Third, adaptation to resource variations at a time

granularity of milliseconds is facilitated by a

common library, such as Chroma. This library saves

application development costs by providing common

mechanisms for (a) monitoring available resources,

(b) profiling the resource demands of alternative

computation tactics, and (c) deciding dynamically

which tactic best supports the QoS goals, given

resource forecasts for the next few milliseconds.

Additionally, this paper clarifies concrete APIs

that adaptive applications need to support for being

integrated into the framework. These APIs are

realized as XML messages, which may be

exchanged within the mobile device, or across the

network, if the infrastructural components are

deployed remotely.

In summary, the proposed framework makes it

easier to develop and integrate applications into

coordinated, resource-adaptive systems.

Furthermore, our research indicates that end-

users with diverse backgrounds are able to control

the behavior of such systems to achieve activity-

specific QoS goals.

REFERENCES

Balan, R. K., Gergle, D., Satyanarayanan, M., & Herbsleb,

J. (2005). Simplifying Cyber Foraging for Mobile

Devices (Tech. Report No. CMU-CS-05-157R).

Pittsburgh, PA: Carnegie Mellon University.

Balan, R. K., Satyanarayanan, M., Park, S., & Okoshi, T.

(2003). Tactics-Based Remote Execution for Mobile

Computing. Paper presented at the USENIX Intl

Conference on Mobile Systems, Applications, and

Services (MobiSys), San Francisco, CA.

Butler, S. (2002). Security Attribute Evaluation Method. A

Cost-Benefit Approach. Paper presented at the Intl

Conf in Software Engineering (ICSE), Orlando, FL.

Capra, L., Emmerich, W., & Mascolo, C. (2003).

CARISMA: Context-Aware Reflective mIddleware

System for Mobile Applications. IEEE Transactions

on Software Engineering, 29(10), 929-945.

De Lara, E., Wallach, D., & Zwaenepoel, W. (2001).

Puppeteer: Component-based Adaptation for Mobile

Computing. Paper presented at the USENIX

Symposium on Internet Technologies and Systems

(USITS), San Francisco, CA.

Flinn, J., & Satyanarayanan, M. (1999). Energy-aware

Adaptation for Mobile Applications. ACM SIGOPS

Operating Systems Review, 33(5), 48-63.

Garlan, D., Siewiorek, D., Smailagic, A., & Steenkiste, P.

(2002). Project Aura: Toward Distraction-Free

Pervasive Computing. IEEE Pervasive Computing,

1(2), 22-31.

Lee, C., Lehoczky, J., Siewiorek, D., Rajkumar, R., &

Hansen, J. (1999). A Scalable Solution to the Multi-

Resource QoS Problem. Paper presented at the IEEE

Real-Time Systems Symposium (RTSS).

Moore, M., Kazman, R., Klein, M., & Asundi, J. (2003).

Quantifying the Value of Architecture Design

USER GUIDANCE OF RESOURCE-ADAPTIVE SYSTEMS

Decisions: Lessons from the Field. Paper presented at

the Intl Conf on Software Engineering (ICSE),

Portland, OR.

Myers, B., Smith, D., & Horn, B. (1992). Report of the

‘End-User Programming’ Working Group. In B.

Myers (Ed.), Languages for Developing User

Interfaces (pp. 343-366). Boston, MA: Jones and

Barlett.

Narayanan, D., Flinn, J., & Satyanarayanan, M. (2000).

Using History to Improve Mobile Application

Adaptation. Paper presented at the 3rd IEEE

Workshop on Mobile Computing Systems and

Applications (WMCSA).

Poladian, V., Sousa, J. P., Garlan, D., & Shaw, M. (2004,

May). Dynamic Configuration of Resource-Aware

Services. Paper presented at the 26th International

Conference on Software Engineering, Edinburgh, UK.

Sousa, J. P. (2005). Scaling Task Management in Space

and Time: Reducing User Overhead in Ubiquitous-

Computing Environments (Tech. Report No. CMU-

CS-05-123). Pittsburgh, PA: Carnegie Mellon

University.

Steinberg, E. (Ed.). (1991). Plain language: Principles and

Practice. Detroit, MI: Wayne State University Press.

TechSmith. Camtasia Studio. Retrieved March, 2008,

from www.techsmith.com

Wijesekera, D., Varadarajan, S., Parikh, S., Srivastava, J.,

& Nerode, A. (1998, April). Performance evaluation

of media losses in the Continuous MediaToolkit.

Paper presented at the Intl Workshop on Multimedia

Software Engineering (MSE), Kyoto, Japan.

Yuan, W., Nahrstedt, K., Adve, S., Jones, D., & Kravets,

R. (2006). GRACE-1: Cross-Layer Adaptation for

Multimedia Quality and Battery Energy. IEEE

Transactions on Mobile Computing, 5(7), 799-815.

ICSOFT 2008 - International Conference on Software and Data Technologies