An architecture for flexible auto-tuning: The Periscope Tuning Framework 2.0

Abstract

Due to the complexity and diversity of new parallel architectures, automatic tuning of parallel applications has become increasingly important for achieving acceptable performance levels, as well as performance portability. The European AutoTune project developed a tuning framework that closely integrates and automates performance analysis and performance tuning. The Periscope Tuning Framework (PTF) relies on a flexible plugin mechanism and provides tuning plugins for various different tuning aspects. Each plugin provides codified expert knowledge for performance or energy efficiency tuning. PTF is able to tune serial and parallel codes for homogeneous and heterogeneous target hardware. The output of the framework is tuning recommendations that can be integrated into the production version of the code. In this paper, we present the latest development in the design of PTF aiming at (1) achieving higher portability and scalability by using the Score-P measurement infrastructure, (2) extending Score-P with tuning capabilities, (3) increasing analysis capabilities by providing new analysis strategies, and (4) increasing tuning capabilities by providing new plugins.