Evaluating technologies and techniques for transitioning hydrodynamics applications to future generations of supercomputers

Current supercomputer development trends present severe challenges for scientific codebases. Moore’s law continues to hold, however, power constraints have brought an end to Dennard scaling, forcing significant increases in overall concurrency. The performance imbalance between the processor and memory sub-systems is also increasing and architectures are becoming significantly more complex. Scientific computing centres need to harness more computational resources in order to facilitate new scientific insights and maintaining their codebases requires significant investments. Centres therefore have to decide how best to develop their applications to take advantage of future architectures. To prevent vendor "lock-in" and maximise investments, achieving portableperformance across multiple architectures is also a significant concern. Efficiently scaling applications will be essential for achieving improvements in science and the MPI (Message Passing Interface) only model is reaching its scalability limits. Hybrid approaches which utilise shared memory programming models are a promising approach for improving scalability. Additionally PGAS (Partitioned Global Address Space) models have the potential to address productivity and scalability concerns. Furthermore, OpenCL has been developed with the aim of enabling applications to achieve portable-performance across a range of heterogeneous architectures. This research examines approaches for achieving greater levels of performance for hydrodynamics applications on future supercomputer architectures. The development of a Lagrangian-Eulerian hydrodynamics application is presented together with its utility for conducting such research. Strategies for improving application performance, including PGAS- and hybrid-based approaches are evaluated at large node-counts on several state-of-the-art architectures. Techniques to maximise the performance and scalability of OpenMP-based hybrid implementations are presented together with an assessment of how these constructs should be combined with existing approaches. OpenCL is evaluated as an additional technology for implementing a hybrid programming model and improving performance-portability. To enhance productivity several tools for automatically hybridising applications and improving process-to-topology mappings are evaluated. Power constraints are starting to limit supercomputer deployments, potentially necessitating the use of more energy efficient technologies. Advanced processor architectures are therefore evaluated as future candidate technologies, together with several application optimisations which will likely be necessary. An FPGA-based solution is examined, including an analysis of how effectively it can be utilised via a high-level programming model, as an alternative to the specialist approaches which currently limit the applicability of this technology