The objective of this research activity is the efficient implementation of numerical methods in Work Package's 2 discontinuous multidomain problems, on many-core architectures.
In the beginning, the implementation will take place in a cluster, using the Message Passing Interface programming model, and will concern both uniform, symmetric and non-uniform schemes. Since the goal is to achieve multi-level parallelism and fully exploit multicore processors in each cluster node, a combination of MPI and other shared-memory programming models (OpenMP, Pthreads) will be used. Moreover, in case there are GPUs available in the cluster, the software will be able to distribute the computation load using appropriate interfaces (CUDA, OpenCL).
During the next stage of the action, the simulations will be carried out in a grid computing environment, formed by interconnected, geographically dispersed clusters. At this stage, the parts of the software that will be developed, will allow an efficient and flexible resource selection during the simulation. The ultimate goal is the further reduction of the numerical methods total execution time, therefore, the software that was developed in the first part of the action will be incorporated in a proper grid software (e.g Globus, gLite)
At the final stage of the action, proper interfaces will be developed in order to make the implementation of computationally demanding simulations possible through web services in the Cloud. That way, the developer will have less work to do as far as non trivial parallelization aspects are concerned and, the software will be able to take full advantage of the computational resources available, especially when more than one users need access.
The University of Patras Research Team is the main work group that will carry out the greater part of the action, will author and publish the results and edit a technical report with the description of the scientific activities that took place during the action and the research results that emerged.
