Argonne science journalist Nils Heinohen wrote this article about the work of JaeHyuk Kwack, a computer scientist at the Argonne Leadership Computing Facility (ALCF), who is responsible for the readiness of scientific applications for high-performance use in the upcoming exascale of the US Department of Energy (DOE) systems .
For several years now, Intel has offered the performance profiling software Advisor and VTune for CPU-based Intel architectures. Before ALCF shipped the Intel HPE Exascale System Aurora in 2022, Intel extended these tools to its Xe GPU architecture. Because the Aurora Software Development Kit (SDK) contains several key programming models, such as: B. DPC ++ / SYCL, OpenMP Target Offloading (for C, C ++ and Fortran), OpenCL, Kokkos and RAJA, these tools must be tested for each combination of the two programming model. Many key applications being developed as part of the DOE’s Exascale Computing Project (ECP) incorporate Intel’s optimized math libraries to maximize their performance. It is therefore critical that the Advisor and VTune can seamlessly capture their performance characteristics.
ALCF’s Kwack has validated key Advisor and VTune performance data for Aurora testbed systems, leveraging key ECP and ECP proxy applications in a wide range of scientific fields ranging from molecular structure systems to atmospheric boundary layer flow simulations. By working with ECP application developers and ALCF performance engineers, he has also provided feedback to Intel teams on functional requirements, critical errors, and user interfaces.
Roof line analysis
Kwack recently focused on promoting a roofline analysis feature from Advisor. The roof line analysis records the performance characteristics of the application and then determines the achievable peak performance. Knowing these characteristics, application developers can identify performance bottlenecks in their applications and optimize the code for Aurora testbed systems (and ultimately for the final Aurora system itself). Kwack leads a variety of conferences including SC, Aurora User Training Events, and the ECP Annual Meeting, roofline analysis tutorials that are regularly updated to reflect the latest technology and trends.
Kwack’s recent endeavors include working on the Advisor’s performance projection feature. Ponte Vecchio GPUs – the processors that are expected to power Aurora – are currently unavailable. However, application developers still need estimates of how application performance will develop under the architecture of the exascale system. Using existing testbed systems to generate ratings for the Ponte Vecchio GPUs, Advisor offers a systematic approach to estimating application performance for Aurora. Kwack has used a number of applications to provide multiple cases for performance projection function validation. In this area, he continues to work with the Intel team to develop improved and more reliable features.
Working with Intel tool developers also has the benefit of informing partnerships with peers at institutions like Rice University and the University of Oregon to prepare applications like HPCToolKit and TAU for production science once Aurora is rolled out to users.