Automatic tuning of HPC applications

Más información sobre el libro

High Performance Computing (HPC) is crucial for various applications, including scientific research in weather prediction, climate simulations, genomics, and automotive and aerospace simulations. Programmers aim to maximize performance on highly parallel and heterogeneous systems, which consist of numerous nodes with multicore processors and optional accelerators like GPGPUs and Intel® Xeon Phi™ processors. Tuning applications for these systems involves co-optimizing aspects such as single-core performance, memory access, load balancing, inter-process communication, and energy consumption. Performance analysis tools like Periscope assist programmers in identifying performance bottlenecks through online analysis during application execution. Periscope's scalable distributed structure allows it to automatically analyze performance. The European FP7 project AutoTune enhanced Periscope for automatic tuning of HPC applications, resulting in the Periscope Tuning Framework (PTF). PTF includes aspect-specific tuning plugins that encapsulate expert knowledge for effective tuning, targeting interprocess communication, OpenMP scalability, single-core execution, and energy efficiency. It supports dynamic performance data access and evaluates multi-dimensional search spaces using advanced algorithms. The framework automates experiment execution and parallel scenario evaluation to expedite tuning. Extensive documentation and an online tutor