Scientific computing requires an ever-increasing number of resources to deliver results for growing problem sizes in a reasonable timeframe. Few years ago, supercomputers were the only way to get enough computation power for such compute intensive tasks. In the last decade, while the largest research projects were able to afford expensive supercomputers, other projects were forced to opt for cheaper resources such as commodity clusters or more modern and challenging computational Grids.
Despite the existence of many vendors that, similar to Grid computing, aggregate a potentially unbounded number of compute resources, Cloud computing remains a domain dominated by business applications (e.g. Web hosting, database servers) and whose suitability for scientific computing remains largely unexplored. In this project we plan to research basic methods that investigate the potential of Cloud infrastructures for high- performance scientific computing and apply them for operational daily use in a domain with high computational and QoS demands: meteorological weather forecast. Today, conducting large-scale simulations for obtaining accurate forecasts is still limited by three major barriers: expensive price of operating high- performance parallel hardware, unreliability of distributed Grid resources, and lack of QoS guarantees (especially forecasting deadlines and real-time streaming bandwidth).
The results of the project are expected to have a twofold economical and regional impact:
- Universities and other research centres financing high-performance scientific computing research may achieve significant budget savings by shifting their business model from operating expensive self-owned data/supercomputing centres to renting virtualised resources from specialised Cloud providers only when and where needed;
- Tyrol, as an alpine area with very complex topography, requires correct and high resolution precipitation forecasts that are difficult to produce, but essential for public services having to warn or prepare for catastrophic events caused by floods or avalanches. The Tyrolean Avalanche Service (Lawinenwarndienst) and the Hydrographical Service (Hydrographsche Dienst) will test and drive the project research, with the ultimate goal of employing the developed product it in their daily operation.
In this project, we plan to research the potential of using Cloud infrastructures for scientific computing by:
- Investigating the computational resources offered by several (i.e. two or three) major commercial Cloud providers (e.g. (Amazon) (AppNexus) (ENKI) (FlexiScale) (GoGrid) (Joyent) (NewServers)) and device models that assess whether their performance is sufficient for scientific computing;
- Researching resource management and scheduling methods for scientific workflows on Cloud platforms by extending an existing Grid application development and computing environment;
- Researching economically-viable SLAs that encapsulate a balance between the QoS offered by the resource providers and the cost of resource use;
- Quantifying the benefits of using leased Cloud resources for scientific applications with respect to performance, reliability, and price, compared to traditional supercomputers, clusters, and Grids;
- Validating the research methods for two real scientific applications from the meteorological and astrophysics domains;
- Using the researched methods and infrastructure in operational daily use at the Tyrolean Avalanche and Hydrographical Services for obtaining precipitation forecasts in mountainous terrains with a spatial resolution of 0.5 km and with extensive information about the forecast uncertainties.
A downscaling model will exploit the strong forcing of underlying terrain on precipitation patterns and amounts in mountainous areas, and provide a means of achieving sufficiently large ensemble sizes for an adequate probabilistic forecast to account for the inherent chaotic nature of atmospheric motions and processes. It uses the general-purpose NWP models to provide conditions at boundaries away from topography over flat terrain. The fine-scale (order of 200-500 meters horizontal resolution) probabilistic precipitation forecasts over the whole Alps with the downscaling model are ideally suited for a workflow solution due to its many atomic entities. We will therefore design a workflow application called RainCloud within the ASKALON environment.