Project start date: December 1, 2018
Project duration: 4 years
Funding institution: Croatian Science Foundation
One of the main research concerns in service-oriented computing is the question of maintaining a reliable execution environment. Services are self-contained applications deployed on remote servers which can be consumed using well-established interfaces. Thus, the Quality of Service (QoS) cannot be maintained by the service consumer as it does not have the direct control over the execution environment (exact server setup) nor does it have control over the parameters influencing the non-functional properties of the communication channel (throughput, latency, jitter and others).
This project focuses on two major problems in the area of dependable service computing. The first topic covers the issues of predicting the QoS parameters (such as reliably and availability) of atomic web services. Prediction is one of the popular methods to estimate the QoS parameters as active measurements can lead to degradation of service quality. More specifically, the project activates are focused on context-aware QoS prediction that gives estimation of QoS properties as seen by the service consumer. The other portion of research efforts of this project is focused at service selection. It is often the case that atomic services are used as building blocks to create more complex composite applications. In such cases, each atomic service has the purpose of handling a specific functional requirement of the application. However, multiple different web services that meet specific functional properties can be available. In such an environment, it is possible to select a service with more favourable QoS properties during runtime. This project focuses on developing efficient and accurate algorithms that can perform atomic service selection at the time of service composition invocation.
Some of the most notable approaches in the field of QoS prediction for web services utilized approaches adopted from the recommender systems. More specifically, the researchers successfully applied the method of collaborative filtering to solve this problem with reasonable accuracy. In general, collaborative filtering methods can be divided into memory-based and model-based approaches. The memory-based approach leverages statistical methods to perform predictions, while the model-based approach employs additional machine learning and data mining techniques. In our previous research efforts, two context-aware QoS prediction algorithms were developed: CLUS and LUCS. LUCS is a hybrid algorithm that leverages both memory and model based approaches, while CLUS is a model-based approach that leverages the k-means clustering to improve prediction accuracy. Recent research results indicate that there is room for further improvement in this area. The goal of this project is to put forward a novel prediction algorithm that will leverage modern machine learning approaches to further improve the prediction accuracy.
Selecting which services are to execute at runtime, taking into account their QoS properties, can be reduced to a classical NP-hard multi-criteria optimization problem. The researchers adopted many heuristic approaches that can solve this problem in polynomial time with some degree of accuracy. The goal of this project is to put forward an approach that periodically considers time frames in which each of the observed composite applications has several (one or more) pending, mutually independent requests. The proposed heuristic will be evaluated taking into account the problem size and time-accuracy trade-off based on the amount of heuristic assumptions and simplifications made in the algorithm.