Several major contributions were made by this thesis towards research into QoS-based web service composition in the cloud. It addressed three important challenges of the problem which is described as an NP-Hard problem. They include multiple-conflicting QoS objectives, multiple constraints and network performance. The thesis also made contributions to the research into network
performance prediction with emphasis on how to accurately estimating end-to- end network latency. In addition, the thesis also made several contributions towards evolutionary algorithm research by introducing new evolutionary algorithms to address the problem. A detailed discussion of the contributions is presented below in the next subsections.
6.2.1 Prediction problem (Chapter 3)
The prediction problem has been extensively studied by previous works leading to the development of several network performance prediction algorithms. However, the algorithms had difficulty in making accurate estimates due to several reasons such as centralized architecture (in the case of EDM) or adoption of a general coordinate update strategy (in the case of NMF) which caused erroneous estimates. Hence, they are not suitable in a cloud environment which usually requires accurate representations of RTTs between web service nodes. This work contributed to the study of prediction algorithms by developing a learning-based non-negative matrix factorization algorithm (LANMF) to improve accuracy of estimating RTTs. LANMF encodes each web service node coordinate as an automaton where each automaton consists of its coordinate update strategy, set of actions and action probabilities. At each iteration, LANMF selects the update strategy with the best probability of leading to minimum prediction error. This is a clearly different strategy from pervious works which generally use the same update strategy for all node coordinates. An extensive comparison of LANMF against other prediction algorithms demonstrated that it had better prediction accuracy than them.
6.2.2 New Methods for Network-aware Web Service Composition in the Cloud (Chapter 4)
This problem has been studied by recent research targeted at development of web service composition algorithms that search for QoS optimal solutions in the cloud. However these algorithms lacked the ability to address real-world
issues in the cloud. The first issue is that recent works ignore the impact of network performance on composite service selection. In practice, network performance metric such as network latency plays a crucial role in determining overall performance of a composition in the cloud. The second issue is that the poor optimality of current techniques makes them unsuitable in dealing with the problem. The third issue is that, due to the large scale nature of our cloud environment, current techniques have poor scalability which makes them a bad choice for tackling the problem. Lastly, current constraint handling strategies are incapable of dealing with a situation where all QoS attributes considered during optimization process are “lower is better”. This work enriched the study of QoS-based web service composition in the cloud by developing evolutionary algorithms that successfully address the issues. This work also contributes to evolutionary algorithm research by presenting for new algorithms; INSGA, KNSGA, NMPSO and NFOA. INSGA provided novel ND-Crossover and ND-Mutation operators which search for low latency and QoS-optimal solutions. KNSGA introduced a new K-mean based K-Mutation operator for searching for web service nodes in the same cluster (in term of network proximity) to certain reference web service nodes. NMPSO uses best particles form different populations to guide search for near-optimal Pareto set. Lastly, NFOA uses non-dominated sort fruit fly search to find network positions of composite services with low latency and optimal QoS. All four algorithms adopted a unique constraint penalty function that rewarded solutions which satisfy QoS constraints and penalized those ones that didn’t satisfy constraints. The penalty function was developed to suite “lower is -better” QoS attributes which were considered in this thesis. An extensive evaluation of the algorithms shows that they have good optimality and scalability when compared to previous works. Among the four algorithms, INSGA shows the best optimality albeit at the cost of scalability. Still, INSGA’s scalability was better than linear integer programming (LIP) and only slightly worse than the other algorithms. NMPSO, KNSGA and NFOA
demonstrated better balance between optimality and scalability than INSGA, Particle swarm optimization, NSGA-II, and LIP.
6.2.3 A New Method for Network-aware Service Composition in Dynamic Environment (Chapter 5)
In contrast with the previous contribution, this work focused on addressing QoS optimization in a dynamic environment where there are constant changes in QoS of web services in the cloud. It is motivated by the discovery that previously proposed techniques were only effective in QoS optimization if QoS of web services remain unchanged. This work proposed a technique known as cellular automata-based NSGA-II algorithm (CellGA) to tackle the problem in a dynamic environment. CellGA adopts new Cell-Crossover and Cell-Mutation operators. The novelty in the operators lies in their ability to use different cellular automata (CA) rules to decide which genes need be altered to arrive at superior children. The rules depend on the global state of a gene which in turn rely on local states of gene’s CA neighbourhood. Experiments conducted confirmed that CellGA has better optimality when compared to the algorithms presented in Chapter 4. It also showed good scalability due to the use of qualitative RTT values which were obtained from LANMF.