Solution of Multi-Physics Flow Problems in Performance Portable Frameworks

Computational fluid dynamics (CFD) has been one of the most important classes of applications targeted by high-performance computing (HPC). Use of CFD in HPC is motivated by the need for solving challenging multi-physics problems. Industries such as aerospace, nuclear power, oil and gas regularly employ these codes at scale requiring flexibility to adapt the computations to a specific problem.

With diminishing returns from Moore's Law in recent years, accelerators, in particular GPUs, have become an important technology to boost application performance. Accelerators can now be found in many top-tier supercomputers, including Summit (ORNL) and Sierra (LLNL) as well as Europe's Piz Daint (ETH Zurich). The use of accelerators raises concerns in terms of programming complexity and performance portability. This is even more noticeable in CFD applications that already have large established production code bases optimized for CPUs. Moreover, algorithmic choices made in the past, such as low order implicit methods and unstructured mesh data-layouts, lead to a substantial effort being required to adapt existing software to increasingly heterogeneous computer architectures. To address these challenges new algorithms, programming paradigms and frameworks are being actively developed in the effort to create a more flexible software ecosystem of the future.

A potentially disruptive technology that has emerged in recent years are domain-specific languages (DSLs) that enable efficient code generation for structured or unstructured mesh operations. DSLs are designed to provide a high-level abstraction with a set of directives and well-defined semantics that the engineer or scientist can use to quickly define their model. They serve as a building block, hiding boiler plate and implementation details for specific architectures. Warwick University in collaboration with the University of Oxford, PPCU Hungary and Imperial College London have made important contributions in performance portability with DSLs in the past.

IBM Research group at Daresbury Laboratory is working with several UK industries that leverage HPC, helping them to take full advantage of constantly evolving hardware and software stacks. Based on the current limitations in computer modelling, IBM-R is interested to build a prototype implementation of a flow solver that would serve as a testbed for novel algorithms and flexible data structures. Warwick and IBM-R have worked together in the past and both are committed to taking the DSL approach to the next level.

In collaboration with IBM Research, this project will target the development of multi-material accelerator-friendly algorithms and the creation of user-friendly abstractions to demonstrate them in a DSL in order to enable wider adoption. These algorithms will need to consistently expose more compute, which accelerators excel at, while reducing the movement of data and amount of collective operations which are the main bottleneck in many heterogeneous systems. As part of this project, the student will undertake a placement at IBM.