System support for object replication in distributed systems

[摘要] Distributed systems are composed of a collection of cooperating but failure prone system components. The number of components in such systems is often large and, despite low probabilities of any particular component failing, the likelihood that there will be at least a small number of failures within the system at a given time is high. Therefore, distributed systems must be able to withstand partial failures. By being resilient to partial failures, a distributed system becomes more able to offer a dependable service and therefore more useful. Replication is a well known technique used to mask partial failures and increase reliability in distributed computer systems. However, replication management requires sophisticated distributed control algorithms, and is therefore a labour intensive and error prone task. Furthermore, replication is in most cases employed due to applications' non-functional requirements for reliability, as dependability is generally an orthogonal issue to the problem domain of the application. If system level support for replication is provided, the application developer can devote more effort to application specific issues. Distributed systems are inherently more complex than centralised systems. Encapsulation and abstraction of components and services can be of paramount importance in managing their complexity. The use of object oriented techniques and languages, providing support for encapsulation and abstraction, has made development of distributed systems more manageable. In systems where applications are being developed using object-oriented techniques, system support mechanisms must recognise this, and provide support for the object-oriented approach. The architecture presented exploits object-oriented techniques to improve transparency and to reduce the application programmer involvement required to use the replication mechanisms. This dissertation describes an approach to implementing system support for object replication, which is distinct from other approaches such as replicated objects in that objects are not specially designed for replication. Additionally, object replication, in contrast to data replication, is a function-shipping approach and deals with the replication of both operations and data. Object replication is complicated by objects' encapsulation of local state and the arbitrary interaction patterns that may exist among objects. Although fully transparent object replication has not been achieved, my thesis is that partial system support for replication of program-level objects is practicable and assists the development of certain classes of reliable distributed applications. I demonstrate the usefulness of this approach by describing a prototype implementation and showing how it supports the development of an example toy application. To increase their flexibility, the system support mechanisms described are tailorable. The approach adopted in this work is to provide partial support for object replication, relying on some assistance from the application developer to supply application dependent functionality within particular collators for dealing with processing of results from object replicas. Care is taken to make the programming model as simple and concise as possible.

[发布机构] University:University of Glasgow

[关键词] Computer science