High-level Abstractions for Performance, Portability and Continuity of Scientific Software on Future Computing Systems- CloverLeaf 3D

In this report we present research in re-engineering the 3D version of the CloverLeaf hydrodynamics application to utilize the OPS domain specific high-level abstraction framework. This work follows on from our previous work on the 2D version of CloverLeaf, but is aimed at extending our findings to a more representative application. We present (1) the lessons learnt in the conversion of CloverLeaf 3D to utilize OPS, (2) the lessons learnt in the development of two new parallelizations for OPS, namely OpenCL and OpenACC and applying them for parallelizing CloverLeaf 3D and (3) performance of the application on a range of modern parallel systems including multi-core CPUs, NVIDIA GPUs, a Cray XC30 distributed memory CPU system and a Cray XK7 distributed memory GPU system. For the purposes of documentation we also detail the recent optimisations carried out for OPS’s distributed memory parallelization, which includes message aggregation to reduce communication times. Our results re-confirm the conclusions of our previous report that the development of parallel HPC applications through the careful factorization of a parallel program’s functionality and imple-mentation, through a high-level framework such as OPS is no more time consuming nor difficult than writing a one-off parallel program targeting only a single parallel implementation. Similar to the 2