Usually, the process of development of services available as web applications considers only functional requirements. Since, an evergrowing number of users take advantage of di®erent kinds of communication channels and devices, this process must be revised by considering new aspects: quality of service (QoS), user pro¯les and technical characteristics of channels. In previous works, we proposed a methodology that provides a rational to formalize the redesign process of existing services to support multi-channel access. This paper extends our approach and highlights how the QoS dimensions can be considered quantitatively in the di®erent phases of the methodology. Moreover, an hybrid approach that allows the QoS evaluation, during the development of a service, is proposed.
Usually, a service available as a web application is characterized by the
functionalities that it provides to the ¯nal users. Therefore, during the development
of these services, only functional requirements are considered. Actually, the
possibility to deliver multi-channel applications underlines the necessity to
characterize each service with a well-de¯ned set of qualities of service (QoS). In fact,
even if a service ful¯lls all of its functional requirements by providing the
required features, it can still be unacceptable if, for example, availability is too
little, performance is too poor, or usability does not meet end-user expectations.
Therefore, traditional development processes need to be rethought to take into
account non-functional requirements of di®erent nature (technological, social,
organizational, etc.) at the right stage of the development. In the MAIS
(Multichannel Adaptive Information Systems) project [
The de¯nition of a quality model will form a proper foundation for identifying, analyzing, and specifying the large number of quality requirements. This quality model provides a tool that can be used to turn these general high-level quality requirements into detailed measurable descriptions. Such model is based on an ontology of qualities that helps in classifying and evaluating a quality with a precise description/de¯nition and with its relations with other qualities. Relations could be of di®erent nature: a quality could be a composition of other simpler qualities, a quality could be a re¯nement of another, two qualities could be independent of each other and ¯nally relations could re°ect di®erent perspectives (provider, users, mediator, . . . ). Such relationships, extracted from the ontology, can be modeled by a QoS tree in which the root represents the analyzed quality, the children nodes are the composing qualities and the leaves are the composing technical characteristics.
The ontology that is under construction in the MAIS project speci¯es for each quality the following attributes: independency, observability, controllability, negotiability and measurability. The independency attribute states whether the quality is primitive or composed. In the latter case the composing qualities must be explicit by the composition rules needed to calculate the analyzed quality. Such rules include linear composition, which requires the de¯nition of weights associated with composing qualities, and non-linear composition, which requires the de¯nition of functions or explicit tables. The others attributes state whether the quality is observable, controllable or negotiable. A quality is observable when the user may only measure its value. Instead, if the user may also express a preference the quality is de¯ned as controllable. Finally, a quality is negotiable when it's possible to establish a process of negotiation between the user and the platform. Moreover, the attribute measurability states how the quality is measured (metric, method of measure, max and minimum value).
It's necessary to underline that, except independency and measurability, the others attributes may have di®erent values according to di®erent factors. A ¯rst factor deals with the quality of the service delivered by a provider with respect to the quality perceived by the client. A second factor deals with the context (business or technological) in which a quality is considered. A third factor deals with the service domain: a quality which is relevant in a certain domain could be irrelevant, or even not measurable, in others. For this reason, the ontology has to be instantiated according to application domain, the context and the prospective of analysis.
The remainder of the paper discusses an hybrid approach to evaluate the value of a QoS. Section 2 will present an overview of the reference methodology. Sections 3 and 4 will detail the QoS evaluation using a running example. Conclusions are drawn in Section 5. 2
In the MAIS project, we are developing a methodology for design/redesign
services that addresses quality aspects in multi-channel contexts. In [
The main objective of the second phase, "High-level redesign", is to redesign the service architecture in the light of new requirements promoted by the new channels and by domain characteristics. Special consideration is given to behavior modeling to address the interaction between the user and the service according to functional requirements. Moreover, in this phase quality requirements related to user and domain needs must be considered. Therefore, it is necessary to locate the quality dimension related to user quality requirements that are described in an high-level language. The ontology and the QoS model previously described are used to locate, de¯ne and classify the QoS in the right mode. The located quality is quanti¯ed and modeled using an extension of standard UML [uml:omg], proposed by OMG, that permits to model QoS, constraints and the relationships between them. The modi¯ed UML diagrams, that are the output of this phase, de¯ne the architecture of the service considering only abstract requirements. In other terms, no speci¯c technologies or user characteristics are addressed.
Instead, the "Context adaptation" phase takes into account the actual
deployment environment to evaluate and adapt the abstract assumptions with
respect to actual technical characteristics of channels and user pro¯les. In
particular, the quality assumptions perform in the high-level redesign phase must
be evaluated. Therefore QoS trees, that show the analyzed qualities and the
relationships among other qualities, are extracted. These trees are selected using
the ontology and the QoS model. Moreover, it's necessary to consider for each
class of domain users the aspects of the user pro¯le (for example, experience
and preference) that are related to the analyzed qualities. These characteristics
and the previously extracted QoS trees are used to de¯ne the quality level
requests by the di®erent classes of ¯nal users. The hybrid approach, proposed in
the following sections, is used to perform this task. This approach permits the
quanti¯cation of a quality (root of the tree) using a bottom-up approach and
linear/non-linear compositions. After this task, a comparison between the level
of quality de¯ned in the high-level redesign phase (service quality request) and
the levels request by the di®erent classes of users (user quality request) must be
performed. The comparison permits, using information related to the analyzed
context, the de¯nition of the level of quality that the service must provide. The
evaluation of the compatibility between this value and the available technology
is made using the hybrid approach and the previously extracted QoS trees. If the
evaluation has a negative result, it's necessary to resolve the discovered
incompatibility. Examples of this task are proposed in the following sections. Instead,
if the evaluation has a positive result, the context adaptation phase is completed
and the output of the methodology is a set of UML diagrams that models the
multi-channel service along with its quality characteristics. Such a model will be
exploited to actually implement and deploy the service to make it available to
clients.
The QoS evaluation technique will be presented by a running example. Let us
assume that the qualities of interest for the end user are: the usability, the
service execution time and the service availability. Then, in the second phase of the
methodology the QoS trees reported in Figure 2 are extracted from the ontology.
A tree node n represents a quality dimension, while edges between a node n and
its children indicate a dependency between quality dimension n and the quality
dimension corresponding to its children. The dependency could be expressed by
an explicit expression, for example the service execution time (i.e. the expected
delay between the time instant when a request is sent and the time when the
result is obtained) is given by the sum of the service response time (i.e., the
time required to process the request by the Service Provider (SP)
infrastructure) and the network transmission time (the time required to transmit request
and response). In the same way, the service availability can be expressed by the
product of the SP availability and the channel availability (the channel in the
MAIS project includes the network, the network interface, the application
protocol and the end-user device. In a mobile environment the channel availability
could be lower than the network availability). We do not introduce any
assumption on the properties of functional dependencies. Dependency could be linear,
non linear or could even be expressed in tabular form if the explicit formulation
is unknown. This latter situation happens very frequently in practice. Among
the 321 quality dimensions classi¯ed in the MAIS project [
The dependency among quality dimensions could also be qualitative. In the example reported in Figure 2 usability depends on learnability, comprehensibility operability and pleasantness. Furthermore, comprehensibility depends on ScreenQoS and NetworkInterfaceQoS. ScreenQoS depends on the quality attributes of the device screen (resolution, size, etc).
In order to evaluate quantitatively the value of usability, the Simple Additive
Weighting (SAW) technique is adopted as proposed in [
First note that the leaves of the tree correspond to technical characteristic tc of the device. Each device can be associated with a tuple < t1; t2; : : : ; tc; : : : ; tC > where each value tc assumes the value of the corresponding technical characteristic. Considering the range of values proposed for Resolution, Size and Color depth of available screens, examples of possible tuples are the following: T1 =< 800x600; 5:0"; : : : ; 32bit > ; . . . ; Tn =< 1024x640; 19:0"; : : : ; 64bit >
The domain expert associates each tuple with a value in [0; 1], i.e. de¯nes a function v(T ). This value represents the level of quality of the parent node when the corresponding tuple is selected. Note that, as shown in Figure 3a the assignment could be non uniform in the interval. In the proposed example, T1 and Tn determine di®erent ScreenQoS values. Note that, the mapping is domain dependent and could depend also on the user pro¯le. In this latter case, the mapping can be modeled as a function v(T; U ) which associates the quality value v to a tuple T and the user pro¯le U P .
Furthermore, each node of the tree is associated with a weight and the value or score of a parent node is calculated by multiplying the child's weight by its value and then summing across all children. Weights are normalized, i.e. their sums equals to 1, hence the quality value of every node ranges in [0; 1]. Figure 3b shows how to evaluate the value of comprehensibility when the device selected for the end user corresponds to screen Tk and networkInterface Tj . The process can be iterated for other dimensions (learnability, operability, etc.) and can be applied bottom up in order to obtain the value for the tree root, i.e., the service usability in the example. 4
The quality evaluation technique allows verifying design hypotheses. If the technical characteristics are ¯xed, then the relevant quality values can be determined. A design hypothesis is veri¯ed if the technical characteristics provide quality values greater or equal to given threshold Bk ¯xed at design time (let us consider for simplicity positive quality dimension, i.e. attributes such that the higher the value the higher the quality for the end user). For example the service design is accepted if the technology characteristics guarantees 0.99% of availability and a usability greater than 0.7. Quality thresholds can be ¯xed a priori as a desired characteristic of the service but could also be determined by end user pro¯les.
In the MAIS framework a user pro¯le is a set of characteristics of the users
that can be exploited for further customization of services. The study and the
determination of the user pro¯les require a preliminarily in-depth analysis of
habits, preferences, behaviors, which is out of the scope of this paper. Pro¯les
de¯ne service requirements for individual and group of users. In order to
improve the system adaptability and usability of the provided services, speci¯c
peculiarities of each user should be highlighted. An example of personalization
is given by the analysis of Activity Participations and Body Functions of each
user as presented in [
The quality evaluation of an end user pro¯le starts with the creation of the UP/QoS Matrix that de¯nes the dependencies between the QoS previously described and the UP characteristics. An example of QoS/UP Matrix is reported in Table 1.
QoS dimensions (Comprehensibility, Learnability, Operability, Pleasantness) are the columns of the matrix, user pro¯le dimensions are (Is ltf pref:, Is ltf skills, ICF relational capabilities, ICF body f unction, Expertise, Deliverypre ¡ f erences) the rows of the matrix. The "X" highlights the dependency between a QoS and a UP dimension. In the proposed example, the Operability dimension is related to Is ltf skills (ability to perform a particular operation), ICF rela ¡ QoS
UP Is ltf pref:
Is ltf skills ICF rel capab: ICF body f unct:
Expertise Delivery pref:
QoS
UP Is ltf pref:
Is ltf skills ICF rel capab: ICF body f unct:
Expertise Delivery pref:
Comprehensibility Learnability Operability Pleasantness
X
X X X X tional capabilities (capacity to interact with the system), ICF body f unction (physical and psychological condition of the user), Expertise and Delivery preferences.
In a second phase (see Table 2), a weight is associated with each identi¯ed dependency with a procedure similar to the SAW technique discussed for QoS trees. Weights are numeric values that represent the level of in°uence between QoS and UP dimensions. Numeric values are domain dependent and therefore assigned by domain experts. The sum of weights of each columns has to be equal to 1.
Comprehensibility Learnability Operability Pleasantness 0.1
If for example we assume that the value of Is ltf pref , ICF body f unct, Expertise and Delivery pref are equal to 0.2, 1, 0.8 and 0.8 then the value of comprehensibility required by the user pro¯le is 0.82. If we consider the example above shown in Figure 3, then the set of technical characteristics selected to deploy the service does not satisfy the user pro¯le requirements and the design hypotheses have to be revised.
The assumptions revision proceeds by identifying the most violated constraints. A di®erent set of technical characteristic should be identi¯ed in order to improve the quality measure which corresponds to the most violated constraint and which does not introduce further constraints violation. In our framework this phase is supported in a semi-automatic manner as it will be discussed in Section 4.1.
Our quantitative approach allows also evaluating the lower bound to be provided by each leaf in order to satisfy assumptions. In the example above by considering the weight assignment 0.2, 0.8 for ScreenQoS and NetworkInterfaceQoS, one can determine that in order to satisfy UP requirements by modifying only the ScreenQoS attribute, then ScreenQoS has to be set equal to 2.9 (this can be derived by the relation 0:2 ¤ ScreenQoS + 0:8 ¤ 0:3 >= 0:82), which is impossible since the value attributed to every node can be at most 1. Vice versa, if the NetworkInterfaceQoS is set equal to 0.85, then the UP constraints is satis¯ed and hence the design hypothesis is veri¯ed.
If, vice versa, design hypotheses are veri¯ed, then in the same way we can determine for each leaf the range such that constraints are satis¯ed.
In the MAIS framework we are implementing a semi-automatic tool which support the designer in the assumption revision which will be presented in the next Section. The service design/re-design can be modeled as an optimization problem which can be formulated as: identify the set of choices for the technical characteristic relevant for the end users and for the service requirements which minimizes design costs.
Let us consider the quality tree reported in Figure 4 and let us assume ¯rst that the quality tree extracted includes only qualitative dependencies, hence the quality value for the root attribute can be determined by a linear expression. Let us indicate with wk the weights associated with the quality attributes of the ¯rst level of the tree, while wlk are the weights of the second level associated with node k. Let us assume that the overall value of the quality tree depends on a single set of technical choices I. Let us indicate the tuples for technical choice i as T i; T2i; : : : ; Tji : : : ; Tni . Every alternative j for the technical choice i can be 1 associated with: { vi;j : the quality value, assigned by the domain expert, for alternative j; { ci;j : the cost associated with alternative j.
For example, if the technical choice i is the end user device the cost ci;j is the cost of provisioning of a given client device (which is proportional to the number of end users). If the technical choice i is the network bandwidth, ci;j is the cost of the network connection. Let us indicate with xi;j the binary decision variable of our model. xi;j is equal to 1 if the j alternative for the technical choice i is selected and 0 otherwise. The optimization problem can then formulated as: P1) min Pi2I
Pn
j=1 ci;j xi;j
Pi2I xi;j = 1; Pk wk Pl wlkvi;j xi;j ¸ B
8i xi;j 2 f0; 1g (1) (2) where the constraints family 1 guarantees that exactly one alternative for each technical choice is selected, while equation 2 guarantees that the quality value provided by the solution is greater than the requirement B, hence the design hypothesis is veri¯ed.
The problem above is a NP-hard linear integer programming problem [
m=1 cmym PM m=1 wmym · W
ym 2 f0; 1g setting xi;j = 1 ¡ yi¤n+j and considering that for every programming problem with objective function F (x), the solution of the problem min F (x) is also the solution of the problem ¡ max F (x) then by relaxing constraint family 1, P1) is a KP. P1) is NP-hard, hence the design/redesign problem is NP-hard even if we consider only one quality tree and we assume that the quality dependencies are qualitative and hence the relation which can be derived by applying the SAW technique are linear.
In real projects, the design/re-design methodology faces several quality trees and non linear dependencies among quality variables. We are developing a local search approach which is based on the following steps: { if the design hypothesis is violated, ¯nd a feasible solution by focusing iteratively on the most violated constraint; { the feasible solution obtained in the ¯rst step (or the solution which corresponds to the design hypothesis if it is veri¯ed) is improved by exploring the current solution neighborhood in order to to ¯nd a quasi-optimum solution; { the optimization technique implements a quality tree partitioning, in order to solve with integer linear programming tools, problems for qualitative dependencies.
We are developing an hybrid optimization approach which interleaves the solution of linear integer programming problems with non-linear problems. We only require to be able to evaluate the value of a quality variable from its children and this requirement is always satis¯ed since in the worst case scenario quality dependencies are expressed by enumeration, i.e., in tabular form. 5
In our previous works we have proposed a methodology for the process of
redesign of existing services. Current work is focused on two ongoing research
activities. We are extending our methodology to support the design of new
services in adaptive multi-channel applications [
Acknowledgment The work presented in this paper has been partially supported by FIRB MAIS project - Multi-channel Adaptive Information Systems: models, methodology, qualifying object-oriented platform and architectures for the °exible on-line information systems.