Cloud computing provides a great opportunity for scientists, as it enables large-scale experiments that cannot are too long to run on local desktop machines. Cloud-based computations can be highly parallel, long running and data-intensive, which is desirable for many kinds of scientific experiments. However, to unlock this power, we need a user-friendly interface and an easy-to-use methodology for conducting these experiments. For this reason, we introduce here a formal model of a cloud-based platform and the corresponding opensource implementation. The proposed solution allows to conduct experiments without having a deep technical understanding of cloud-computing, HPC, fault tolerance, or data management in order to leverage the benefits of cloud computing. In the current version, we have focused on biophysics and structural chemistry experiments, based on the analysis of big data from synchrotrons and atomic force microscopy. The domain experts noted the time savings for co mputing and data management, as well as user-friendly interface.

This paper presents a formal model for science clouds, capable of predicting and controlling resources scalably, as well as its implementation as an open source solution, called Chiminey. The feasibility of Chiminey is shown using case studies on biophysics and structural chemistry computations. Big data is acquired from scientific instruments such as synchrotrons and atomic force microscopes. The model takes into account the architecture of the overall parallel and distributed system including large-scale data sources; data sinks, for example petabyte research data stores; and cluster or cloud virtual resources and infrastructures characterised by users in simple parameters upfront. Chiminey is developed to control large numbers of processes and to provide a reliable computing and data management, which can be used by researchers without having to learn extensive infrastructure concepts and technologies.

The enabling of scientific experiments that are embarrassingly parallel, long running and data-intensive into a cloud-based execution environment is a desirable, though complex undertaking for many researchers. The management of such virtual environments is cumbersome and not necessarily within the core skill set for scientists and engineers. We present here Chiminey, a software platform that enables researchers to (i) run applications on both traditional high-performance computing and cloud-based computing infrastructures, (ii) handle failure during execution, (iii) curate and visualise execution outputs, (iv) share such data with collaborators or the public, and (v) search for publicly available data.

Failure in grids is costly and inevitable. Existing fault tolerance (FT) mechanisms are typically defensive and reactive, thus unnecessarily costly. In this paper we propose a hybrid FT approach, recovery aware component (RAC), combining reactive and proactive FT, with failure recovery or aversion of user-defined granularity, by component-orientation and architecture-level reasoning about FT, to increase reliability and availability without needless performance sacrifices. We model and analyse a parameterised RAC implementation combining prediction, proactive rejuvenation and reactive restarting to varying extents, calculating cost savings, reliability improvements and cost-benefit, under parameters such as prediction frequency and accuracy.