As large-scale scientific computing platforms increase in size and capability, their complexity also grows. These systems require great care and attention, much of which is due to the rise in failures from increased node/ component counts. Fault tolerance, or resilience, is a key challenge for computing and a major factor in the successful utilization of high-end scientific computing platforms. As the importance of fault tolerance increases, methods for experimentation into new mechanisms and policies are critical. The methodical investigation of failure in these systems is hampered by their scale, and a lack of tools for controlled experimentation. The focus of this research is to provide a versatile: low-overhead platform for fault tolerance/ resilience experimentation in a high-performance computing (HPC) environment. The objective is to extend the HPC workflow and toolkit to provide ways for studying largescale scientific applications at extreme scales with synthetic faults (errors) in a controlled environment. As part of this research we leverage prior work in the areas of HPC system software and performance evaluation tools to enable controlled experimentation through fault injection, while maintaining acceptable performance for scientific workloads. The research identifies two crucial characteristics that are balanced for fault-injection experiments: (i) integration (context), and (ii) isolation (protection). The result of this research is a Resilience Investigation Framework (RIF) that provides HPC users and developers a versatile experimental framework that balances integration and isolation when exploring resilience methods and policies in large-scale systems

The aim of this thesis is to develop high-level approaches for constructing efficient on-demand HPC applications in the cloud. The orchestration language Orc is used to express a number of service-based software designs, at the Software, Platform and Infrastructure levels of the cloud. A partial-correctness framework is developed to reason about orchestration software. Partial-correctness is used because cloud-based services have the potential to fail. A cloud based implementation of Block Matrix Multiplication has been implemented and experimental results from the Amazon cloud have been generated and analysed. The problem of discovering appropriate cloud resources for deploying an application is tackled firstly by specifying a set of application requirements; these are subsequently used to drive a search for appropriate cloud resources for hosting the application. The approach is flexible in that resources can be discovered in a multi-provider marketplace. Typically orchestrations are evaluated in business environments where demand fluctuates. In periods of high demand, service performance can be degraded, perhaps even to the point of failure. Elasticity can be used to counteract performance problems by supplying extra compute resources, as necessary. Game theory is used to analyse the performance of block matrix multiplication in a number of stressed cloud scenarios.

No description available.

No description available.

No description available.

No description available.

No description available.

No description available.

No description available.

No description available.