D12.1: Heterogeneous and Auto-tuned Runtime System

Work Package 12 (WP12) “Novel Programming Techniques” performs research and development in four key areas for future multi-petascale and exascale systems. The work in WP12 focuses on auto tuned and automatic techniques to be applied in parallel programming model runtimes (Task 12.1: “Auto-tuned runtime Environments”), performance tools (Task 12.3: “Development environments and tools”) and file systems (Task 12.4: “File system optimization”). Furthermore, as it is widely accepted that the key to exploiting future high-end systems will be based on research on new numerical algorithms as well as advancing the parallel processing technology used for higher scalability in numerical applications; consequently WP12 also focuses on research studies exposing more scalability for numerical algorithms (Task 12.2: “Scalable numerical algorithms”). Task 12.1 contributes to improve the support of auto-tuning methods to face the complexity of existing and future large scale systems. It impacts parallel languages, runtime, generic and kernel specific auto-tuning algorithms, multi-core, many-core and multi-node sytems, as well as batch systems and energy consumption measurement methods. In total, five areas organized as individual projects were covered. This document is a summary of the projects’ results. It contains a brief description of covered topics. Links to PRACE white papers are given for those readers that are interested in more detailed information. The key research topics investigated in Task 12.1, related to auto-tuned runtime environnement, are the following: Language and runtime support for code variant selection Kernel specific auto-tuning algorithm (FFTW, 2D stencil, and integral graph) Auto-tuning collective operations in MPI Energy efficiency GPU support in batch schedule