User-friendly programming of GPU-enhanced clusters via automated code translation and optimization

Målet med prosjektet er å la forskerne kunne bedre og lettere ta i bruk morgendagens supercomputer, som vil bestå av både standarde CPUer og såkalte akseleratorer (f.eks. GPUer og many-core prosessorer). Prosjektets aktiviteter er av tre typer. Først studerer vi nye programmerings metodikker som tilpasser seg for fremtidige heterogen supercomputer. Videre utvikler vi programmeringsverktøy som andre forskere kan bruke. Til nå har vi laget en prototype av en automatisk kilde-til-kilde kompilator som kan generere hybrid MPI-OpenMP-CUDA kode, med vanlig C kode som input. Sist og ikke minst anvender vi vår nye kunnskap på real-world applikasjoner. Den mest suksessrike applikasjonen har vært å simulere hvordan kalsium sprer seg i en del av en hjertecelle med en beregnings oppløsning ned mot nanometer. Applikasjonen har klart å bruke verdens mest kraftige datamaskin: Tianhe-2.

By developing a simple directive-based programming model and its accompanying fully automated source-to-source code translator and domain-specific optimizer, we aim to greatly simplify the task of programming scientific codes that can run efficiently on a ccelerator-enhanced computer clusters. This project is motivated by an urgent need from the community of computational scientists for programming methodologies that are easy to use, while capable of harnessing especially the non-conventional computing res ources, such as GPUs, that dominate today's HPC field. Based on a proof-of-concept work that has already successfully automated C-to-CUDA translation and optimization restricted to the single-GPU scenario and stencil methods, the proposed project aims to greatly enhance the success by extending to the following topics: (1) improving the newly developed directive-based programming model and its accompanying framework of automated code translation and optimization, (2) including finite element methods and p article methods as two new application domains, (3) extending to the scenario of multiple GPUs, (4) extending to the scenario of GPU-accelerated CPU clusters, (5) tackling a number of real-world scientific codes. The project has the potential of considera bly enhancing the productivity of computational scientists, to let them focus more on their scientific investigations at hand, instead of spending precious time on painstakingly writing complex codes.