^{1}

^{2}

^{2}

^{1}

^{2}

OWL-S, one of the most important Semantic Web service ontologies proposed to date, provides a core ontological framework and guidelines for describing the properties and capabilities of their web services in an unambiguous, computer interpretable form. Predicting the reliability of composite service processes specified in OWL-S allows service users to decide whether the process meets the quantitative quality requirement. In this study, we consider the runtime quality of services to be fluctuating and introduce a dynamic framework to predict the runtime reliability of services specified in OWL-S, employing the Non-Markovian stochastic Petri net (NMSPN) and the time series model. The framework includes the following steps: obtaining the historical response times series of individual service components; fitting these series with a autoregressive-moving-average-model (ARMA for short) and predicting the future firing rates of service components; mapping the OWL-S process into a NMSPN model; employing the predicted firing rates as the model input of NMSPN and calculating the normal completion probability as the reliability estimate. In the case study, a comparison between the static model and our approach based on experimental data is presented and it is shown that our approach achieves higher prediction accuracy.

Web Services are interfaces that describe a collection of operations that are network-accessible through standardized protocols. Web services in the Semantic Web are described through ontologies, which represent formally the service features by using a semantic mark-up language that follows a logical paradigm. One of the most used ontologies to specify Semantic Web service compositions is OWL-S (Ontology Web Language for Services) [

Recently, various studies have discussed how to effectively model and predict the reliability of service compositions based on the aggregations of the reliabilities of its constituent activities. These studies share a common idea that they all try to fit historical reliability data of activities into assumed distributions (deterministic, exponential, geometrical, or general distributions) and employ these obtained distributions as the static model input into the static stochastic models (continuous Markovian model, discrete Markovian model, or PERT (Program Evaluation and Review Technique, [

Our aim in this paper is therefore the development of a dynamic reliability prediction approach for service compositions, with special attention to control flow modelling of OWL-S processes and the dynamics of service quality in runtime environment. This research employs the OWL-S service processes as the example. The main innovation includes a feature-completed translation from OWL-S to Non-Markovian stochastic Petri net (NMSPN for short) and an analytical reliability method using the predicted quality parameters (using the ARMA model) as model input. In the case study, we obtain the experimental reliability data of actual service ontologies and conduct a comparative study between our approach and the static models. The comparison suggests that our dynamic prediction model produces less errors and achieves higher prediction accuracy.

During the last decade, the QoS of web services and service compositions has received a lot of attention in the research community. Reference [

In this case, a composite service is equivalent to a static PERT network with deterministic edges. The service completion time can be easily calculated as the longest execution path of the PERT network. Similar works can be found in [

It is easy to see that the aforementioned assumption of deterministic task execution times and normal completion rates of service tasks is unrealistic for real-world scenarios. Recent studies therefore assume randomly distributed variables as the model input instead. To simplify the solution, exponential or geometric distributions of tasks are assumed in [

Recently, a step forward is taken by assuming generally distributed task execution times and reliabilities. For example, the work in [

Autoregressive/moving average (ARMA) [

The innovations

To predict the future value of an established ARMA

The details of these iterative steps can be found in [

In this section, we introduce the translation rules for OWL-S. Because the syntax of OWL-S is too vast, we restrict the translation into a subset (OWL-S elements such as input-binding, output-binding, data manipulation, boolean condition evaluation, preconditions, and results are abstracted away and omitted), which describes the control flows of the activity executions and message exchanges. This subset is mainly specified by the

Firstly, translation rule for atomic process is presented. The atomic process in OWL-S is a description of a service that expects one (possibly complex) message and returns one (possibly complex) message in response. It is invoked through the

According to the discussion above, an atomic process in OWL-S can be translated into the NMSPN model given in Figure

NMSPN model of the atomic process.

In the following, we present translation rules for composite processes. All composite processes are composed of atomic processes. The

We start with the

NMSPN model of the sequence process.

The

NMSPN model of the choice process.

The

NMSPN model of the split process.

The

NMSPN model of the split-join process.

The

NMSPN model of the if-then-else process.

The

NMSPN model of the any-order process.

Both the

NMSPN model of the repeat-while process.

NMSPN model of the repeat-until process.

In this section, we conduct a case study to illustrate the effectiveness of the translation introduced above. The case study is based on the frequently used

NMSPN model of the any-order process.

As discussed earlier, timed transitions correspond to two kinds of activities in the atomic processes, namely, the timeout activities and activities of executing invoked services. The former always have invariable delays (prescribed by system settings) and the latter always have nondeterministic delays. Instead of the static evaluation of model inputs, we dynamically fit the historical data of these nondeterministic delays into a time series model and predict the future firing rates using the ARMA method. The prediction is implemented as follows.

For any nondeterministic timed transition, its empirical firing rate at time

Thus, the firing rates measured at different times can be described as a sequence of data with equal time intervals:

Employing

In this section, we introduce the analytical methods to predict reliability of OWL-S process. This method takes the predicted fire rates (based on ARMA methods discussed earlier) and the NMSPN representations as model inputs. We use process-normal-completion-probability (PNCP) as the metric of reliability. From the NMSPN view, PCNP denotes the probability that the

To predict PNCP, we first have to predict reliabilities of individual atomic processes. Let

According to earlier discussions, the

Prediction of composite processes is easier. For example, the reliability of

The reliability of

The prediction of

As discussed earlier, the

To prove the feasibility and accuracy of our model, we execute the OWL-S sample given in Section

Using the soapUI [

Firing rate series of LocateBook.

Firing rate series of ShipmentManagement.

Firing rate series of LoadUserProfile.

Firing rate series of SpeficyPaymentMethod.

Firing rate series of PutInCart.

Firing rate series of ValidateUserMail.

We also obtain the SOAP failure connection rates of the six atomic processes as 0.0042, 0.0019, 0.0068, 0.0075, 0.0054, and 0.0048. That is to say, firing probabilities of

Using the ARMA as the prediction model and the firing rates series shown in Figures

The comparison of the predicted PNCP and the actual reliabilities.

Figure

In this paper, we present a comprehensive dependability prediction model for OWL-S processes. We first introduce a set of translation rules to map the process-level elements of OWL-S into the Non-Markovian stochastic Petri net (NMSPN). Based on the NMSPN representations, we employ the ARMA-based time series method to predict the future firing rates of the execution delays of invoked external services. Using these firing rates as model input, we introduce an analytical method to calculate the process-normal-completion-probability as the predicted future reliability. In the case study of real-world OWL-S ontologies, we show that our approach achieves higher prediction accuracy than the static evaluation model.

Based on the current research, we are also considering further studies as follows:

developing tools for automatic translation of OWL-S processes and reliability/dependability computation. We are currently cooperating with the Center of National Software Engineering of Peking University to integrate our prediction models into their automatic formal translation tools. It is hoped that a new tool for automatic Petri-net-translation will be realized;

introducing methods to predict other metrics such as performance, mobility, reputation, maintainability, and reusability.

Developing selection and scheduling algorithms. Based on our research, the reliability of a given composite OWL-S service process can be calculated. However, sometimes we are more interested in knowing, for a given process definition, how to choose from many functionally identical but qualitatively different services and schedule services to achieve the best reliability. We are therefore developing selection and scheduling algorithms based on the NMSPN-based reliability model.

The authors declare that there is no conflict of interests regarding the publication of this paper.

This work is supported by NSFC foundations of China under Grant nos. 61103036 and 61202347, Young Scientist Foundation of Chongqing no. cstc2013kjrc-qnrc0079, Fundamental Research Funds for the Central Universities under Project no. CDJZR12180012, and National Key Technology RD Program of China under Project no. 2011BAH25B03.