Architecture-based availability prediction and service recommendation for cloud computing

Abstract

Cloud computing has become an essential pillar in the development of modern IT systems. IoT solutions such as smart factories and connected cars, or finance and e-commerce systems, just to name a few, have formed a strong availability dependency to the cloud. Consequently, it becomes increasingly important to assess the availability of a cloud application during its development phase to choose suitable cloud services to assure high service quality and to define robust service-level agreements with clients. However, one single cloud data center is already a highly complex system of hard- and software components, making availability assessments of cloud applications a challenging task. While researchers acknowledge the significance of infrastructure and communication faults in the cloud as important aspects of the availability prediction, they usually model either the (compute) infrastructure or the communication part of a cloud service, disregarding that a cloud application consists of multiple interconnected services with potentially different replication degrees and common cause failures. Especially with the introduction of the Function-as-a-Service (FaaS) paradigm, which have a large number of redundant service instances, availability models become computationally infeasible when modeling k-out-of-n services with large n. Usually, availability assessments guide design choices by probing for different cloud services that fit availability requirements or other constraints such as cost simultaneously. However, this task becomes also increasingly challenging in its own right due to the vast amount of potential service offerings and individual configuration options. As a result, recent developments in cloud application modeling have introduced a wide range of technology-agnostic cloud modeling languages, introducing abstractions or patterns that do not require ample knowledge on concrete cloud services anymore. However, the initial challenge remains: the more abstract a pattern, the larger the solution space, and the longer the time to find a suitable solution. This thesis addresses both challenges. First, it proposes a hierarchical availability model implementing a novel availability model utilizing Bayesian networks for the prediction task. Second, it presents a service recommendation system based on a novel pattern-based cloud modeling language called Clams, which provides a framework for custom search criteria in cooperation with meta-heuristics. In detail, the availability model enables developers to model cloud applications at any preferred level of component and network granularity, accounting for cascading infrastructure and communication faults, including the individual replication semantics of services. Moreover, this work introduces scalable Bayesian network structures to enable the modeling of FaaS offerings, or large-scale replicated services, with many instances. The presented service recommendation system provides a generic approach of utilizing meta-heuristics to exploit a component-based architectural description of a cloud application. Cloud computing patterns are used as architectural placeholders while at the same time encoding the solution space of concrete services. Combining the service recommendation system with the availability model, this work demonstrates its results by implementing a system that suggests services with minimal operational cost, while adhering to availability constraints. To show the feasibility of the modeling concepts, this thesis analyses a set of thirty-one real-life architectural examples. Performance evaluations show that the service recommendation system can return a near-to-optimal solution in a feasible time.