Constructing fail-controlled nodes for distributed systems : a software approach

• Main Author: Brasileiro, Francisco Vilar
• Published: University of Newcastle Upon Tyne 1995
• Subjects: 005; Fault tolerance
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.260827

Designing and implementing distributed systems which continue to provide specified services in the presence of processing site and communication failures is a difficult task. To facilitate their development, distributed systems have been built assuming that their underlying hardware components are Jail-controlled, i.e. present a well defined failure mode. However, if conventional hardware cannot provide the assumed failure mode, there is a need to build processing sites or nodes, and communication infra-structure that present the fail-controlled behaviour assumed. Coupling a number of redundant processors within a replicated node is a well known way of constructing fail-controlled nodes. Computation is replicated and executed simultaneously at each processor, and by employing suitable validation techniques to the outputs generated by processors (e.g. majority voting, comparison), outputs from faulty processors can be prevented from appearing at the application level. One way of constructing replicated nodes is by introducing hardwired mechanisms to couple replicated processors with specialised validation hardware circuits. Processors are tightly synchronised at the clock cycle level, and have their outputs validated by a reliable validation hardware. Another approach is to use software mechanisms to perform synchronisation of processors and validation of the outputs. The main advantage of hardware based nodes is the minimum performance overhead incurred. However, the introduction of special circuits may increase the complexity of the design tremendously. Further, every new microprocessor architecture requires considerable redesign overhead. Software based nodes do not present these problems, on the other hand, they introduce much bigger performance overheads to the system. In this thesis we investigate alternative ways of constructing efficient fail-controlled, software based replicated nodes. In particular, we present much more efficient order protocols, which are necessary for the implementation of these nodes. Our protocols, unlike others published to date, do not require processors' physical clocks to be explicitly synchronised. The main contribution of this thesis is the precise definition of the semantics of a software based Jail-silent node, along with its efficient design, implementation and performance evaluation.