71. DSL and Autotuning Tools for Code Optimization on HPC Inspired by Navigation Use Case
Authors: Jan Martinovic (Technical University of Ostrava)Katerina Slaninova (Technical University of Ostrava)Martin Golasowski (Technical University of Ostrava)Radim Cmar (Sygic)Joao M. P. Cardoso (University of Porto)Joao Bispo (University of Porto)Gianluca Palermo (Polytechnic University of Milan)Davide Gadioli (Polytechnic University of Milan)Cristina Silvano (Polytechnic University of Milan)
Abstract: Improving performance and scalability of a source code for the HPC platform is a tedious and time consuming task. The code has to be executed many times under different conditions in order to observe its behavior. This optimization task is best performed automatically, but due to the heterogeneous nature of the source codes for the HPC platform, full automation is often hard to implement. We present a novel approach to this problem which uses a domain specific language (DSL) and a custom compiler toolset LARA proposed in the ANTAREX project. The DSL allows to specify strategies for code transformations, including required instrumentation for various tools and libraries. Code used in our example is part of a extensive code base of a server-side navigation software. In our poster, we present LARA strategies for scalability and performance measurements of a source code and for integration of the autotuning framework.
Two-page extended abstract: pdf