Prototyping for Exascale

Prototyping is a common approach to developing new scientific codes. Through prototypes a scientist builds confidence that an idea can in fact work for a scientific challenge, and the prototype also acts as the definitive design for a final implementation. As supercomputing approaches exaflops performance, the teraflops platforms that are available for prototyping becomes increasingly distant from the target performance, and new tools are needed to help close the performance gap between high productivity prototyping and high performance end-solutions. In this work we propose Bohrium, an open-source just-in-time compiler, as a possible solution to the prototyping problem. We will show how the same Python program can execute seamlessly on single-core, multi-cores, GPGPU and cluster architectures, and thus eliminating the need for parallel programming in the prototype stage. We will show how the same, unmodified, Python implementation of a Jacobi solver, a Black-Scholes pricing, an O(n2) complexity n-body simulation, and a Shallow-Water simulation scales to a 32-core machine with 50.1, 29.8, 17.3, and 44.4 speedups compared to the NumPy execution, while the same Python benchmarks run on a NVidia GTX 680, achieves speedups of 55.7, 43.0, 77.1, and 140.2, and a eight node cluster with gb-ethernet interconnect (256 cores in total), obtain speedups of 4.1, 7.9, 6.6 and 6.4, compared to a single 32 core node.