Management of Long-Running High-Performance Persistent Object Stores

The popularity of object-oriented programming languages, such as Java and C++, for large application development has stirred an interest in improved technologies for high-performance, reliable, and scalable object storage. Such storage systems are typically referred to as Persistent Object Stores. This thesis describes the design and implementation of Sphere, a new persistent object store developed at the University of Glasgow, Scotland. The requirements for Sphere included high performance, support for transactional multi-threaded loads, scalability, extensibility, portability, reliability, referential integrity via the use of disk garbage collection, provision for flexible schema evolution, and minimised interaction with the mutator. The Sphere architecture is split into two parts: the core and the application-specific customisations. The core was designed to be modular, in order to encourage research and experimentation, and to be as light-weight as possible, in an attempt to achieve high performance through simplicity. The customisation part includes the code that deals with and is optimised for the specific load of the application that Sphere has to support: object formats, free-space management, etc. Even though specialising this part of the store is not trivial, it has the benefit that the interaction between the mutator and Sphere is direct and more efficient, as translation layers are not necessary. Major design decisions for Sphere included (i) splitting the store into partitions, to facilitate incremental disk garbage collection and schema evolution, (ii) using a flexible two-level free-space management, (Hi) introducing a three-dimensional method-dispatch matrix to invoke store operations, which contributes to Sphere's ease-of-extensibility, (iv) adopting a logical addressing scheme, to allow straightforward object and partition relocation, (v) requiring that Sphere can identify reference fields inside objects, so that it does not have to interact with the mutator in order to do so, and (vi) adopting the well-known ARIES recovery algorithm to ensure fault-tolerance. The thesis contains a detailed overview of Sphere and the context in which it was developed. Then, it concentrates on two areas that were explored using Sphere as the implementation platform. First, bulk object-loading issues are discussed and the Ghosted Allocation promotion algorithm is described. This algorithm was designed to allocate large numbers of objects to a store efficiently and with minimal log traffic and was evaluated using large-scale experiments. Second, the disk garbage collection framework of Sphere is overviewed and the implemented compacting, relocating garbage collector is described, along with the model of synchronisation with the mutator