The grid is inherently unreliable due to its geographical dispersion, heterogeneity and the involvement of multiple administrative domains. The most general case of failures are so-called Byzantine failures where no assumptions about the behavior of faulty components can be made. In this paper a novel system is described that allows to diagnose and tolerate byzantine faults based on service replication. We suggest, briefly describe and compare two fail-stop and two byzantine fault tolerance algorithms. Given that many scientific larger-scale grid applications have complex outputs the comparison of replica results as needed to implement byzantine fault tolerance becomes a non-trivial task. Therefore we include an automation mechanism based on a generic description language and code generation for this particualar problem. Our approach has been implemented as extension to the Otho Toolkit, a system that synthesizes tailor-made wrapper services for a given application, grid environment and resource. An analysis of performance and overheads for three real-world applications completes our work.

Classification turns chaotic knowledge into regularity by systematizing a domain and providing a common vocabulary. Currently there is a lack of systematic and comprehensive studies in organization and classification of Grid faults. We address this gap with a multi-perspective Grid fault taxonomy describing an incident using eight different characteristics. It is hard to define a taxonomy of broad validity and acceptance that satisfies the vast number of requirements of the many Grid user communities. Nevertheless we proof that our taxonomy can serve as a solid basis for defining project-specific custom classification schemes by giving a concrete example created for a state-of-the-art Grid middleware environment.

Quick and accurate identification of the root cause of failures is an important prerequisite for any reliable system. However with increasing Grid size and complexity, the manual diagnosis of application faults becomes impractical, tedious and time-consuming. So far there has been a lack of systematic and comprehensive studies in organization and classification of Grid faults. We address this gap with a multi-perspective Grid fault taxonomy precisely describing an incident using eight different characteristics. Based on that, we develop a pragmatic model-based technique for application-specific fault diagnosis using indicators, symptoms and rules. Customized wrapper services then apply this knowledge to reason about the root causes of failures. In addition to applying user-provided diagnosis models we demonstrate that given a set of past classified fault events it is possible to automatically extract new models through learning. We investigated and compared several supervised classification learning and cluster analysis algorithms for this purpose. Our approach was implemented as part of the Otho Toolkit, a framework for "service-enabling" legacy applications by synthesizing tailor-made wrapper service.

Extraordinary long lifecycles of many scientific applications commonly surpass multiple generations of Grid technologies. Therefore the smooth adaptation and migration to newer environments remains as interesting research question. This paper presents the Otho Toolkit for synthesis of application-specific Grid service wrappers based on specifications of scientific legacy programs. The services are customised and tailor-made for a specific application, service hosting environment and computational infrastructure and include source code for optional manual refinement. We demonstrate its unique combination of advanced features like support for multiple service platforms, parameter sweeping, iterative and parallel programs, progress-reporting, filestaging and security credential management. Moreover our services reliably identify program termination-causes based on programmatically evaluated post-mortem program states. We applied the Otho Toolkit recursively to itself to synthesise a sophisticated Factory service that creates application-specific Grid services on-demand^1.

Resource management is a key concern for implementing effective Grid middleware and shielding application developers from low level details. Existing resource managers concentrate mostly on physical resources. However, some advanced Grid programming environments allow application developers to specify Grid application components at high level of abstraction which then requires an effective mapping between high level application description (activity types) and actual deployed software components (activity deployments). This paper describes GLARE framework that provides dynamic registration, automatic deployment and on-demand provision of application components (activities) that can be used to build Grid applications. GLARE simplifies description and presentation of both activity types and deployments so that they can easily be located in the Grid and thus become available on-demand. GLARE has been implemented based on a super-peer model with support for activity leasing, self management, and fault tolerance. Experiments are shown to reflect the effectiveness of the GLARE.