Reliable Cloud Services with Byzantine Fault Tolerance

Cloud computing has developed rapidly in the last decade, offering different types of products to customers. Compared to local servers in the data centers, the use of clouds can bring significant benefits to customers, such as flexible scalability, high availability, and lower infrastructure maintenance costs. For this reason, more and more Internet service providers prefer to run their services such as social networking and online streaming on a cloud infrastructure rather than on local servers.
In addition to these commodity services, there is also such a trend in traditional infrastructure-related services. For example, traffic control and railway network management are gradually being integrated into cloud infrastructures, which can be critical as they usually have an incredible responsibility to society, meaning that any failure can have fatal consequences. However, since most existing cloud infrastructures do not guarantee fault tolerance for dependable stateful services, they need to be strengthened to provide higher reliability and availability.
In this thesis, we present the design for building a cloud environment that provides reliable services with Byzantine Fault Tolerance (BFT) that tolerates arbitrary failures of system components and guarantees good performance, including a set of solutions that address the problems from different angles.
First, we introduce a parallel ordering framework that relies on multiple leaders to improve the performance and scalability of BFT systems. Under this framework, multiple BFT protocol instances run within each replica for concurrent ordering and execution of independent requests. Second, we present a prototype that leverages trusted execution environments to provide transparent access to BFT systems. It implements a replacement for the library required for client access within the trusted environment to allow legacy clients to access the replicated services without modification. We also introduce a novel read optimization to further accelerate the processing of read-heavy workloads while providing strong consistency guarantees. Third, we implement a framework that can automate the deployment and evaluation of BFT systems with different software configurations. By using this framework, the error-prone manual management and orchestration of replicated services in the cloud environment can be avoided, and the comparison of different fault-tolerant systems can be greatly facilitated.