PAPI on Emerging Architectures: Challenges and Directions

The Performance API (PAPI) has long served as a foundational tool for accessing hardware performance counters across diverse architectures. As HPC systems continue to evolve toward increasingly heterogeneous and accelerator-dominated platforms, PAPI must adapt to support new devices, software stacks, and programming models while maintaining portability and usability for application developers, tool builders, and facilities.

This BoF session will provide an overview of current PAPI activities and future directions, with a particular focus on ongoing efforts to modernize the codebase and expand C++ support. We will highlight recent development priorities and discuss how these changes aim to better support emerging architectures and programming paradigms.

Several short talks will then focus on PAPI support for accelerators and specialized hardware. Topics include PAPI integration with AMD technologies such as ROCm, ROCprofiler-SDK, and AMD SMI, experiences and challenges supporting CUDA-based systems, and early efforts to enable performance monitoring for AI-focused architectures. These talks are intended to provide the most current updates and to frame key challenges, rather than provide exhaustive technical detail.

The remainder of the session will be dedicated to interactive discussion with the community. We invite feedback on current PAPI capabilities, gaps users encounter on modern systems, and priorities for future development. In particular, we seek input from application developers, tool developers, vendors, and facility staff on requirements for performance monitoring across GPUs, AI accelerators, and heterogeneous platforms.

This BoF aims to foster an open dialogue around the future of performance measurement in HPC and to help guide PAPI’s evolution in response to the needs of current and next-generation systems.

Presenters

• Heike Jagode (Innovative Computing Lab (ICL), University of Tennessee Knoxville)
• Daniel Barry (Innovative Computing Lab (ICL), University of Tennessee Knoxville)
• Dong Jun Woun (Innovative Computing Lab (ICL), University of Tennessee Knoxville)
• Treece Burgess (Innovative Computing Lab (ICL), University of Tennessee Knoxville)
• Tokey Tahmid (Innovative Computing Lab (ICL), University of Tennessee Knoxville)