Tenth Workshop on Accelerator Programming and Directives (WACCPD 2023)

Hosted by Pawsey Supercomputing Research Centre’s Marciej Cytowski, the objective of the workshop is to share methods and case studies that demonstrate programmability, performance, and performance portability across architectures for various distributed HPC, data, and AI workloads.

As part of SC23, this workshop will be part of an unparalleled mix of thousands of scientists, engineers, researchers, educators, programmers, developers, and system administrators who intermingle to learn, share, and grow. Each year, SC provides the leading technical program for the HPC community. The Program is designed to share best practices in all areas of high performance computing.

The inclusion of accelerators in HPC systems has become well established, with many examples of successful deployments today. Looking forward, the trend is expected to continue, with accelerators becoming even more widely used and delivering a larger fraction of the system compute capability. There will also be a tighter coupling of components within a compute node. This transformation in the HPC system landscape has been facilitated by advancements in the capability and usability of accelerators, such as GPUs. These advancements include higher bandwidth memory technologies with larger capacities, hardware-managed caches, and the ability to access CPU data without explicit data management. As a result, scientific software developers now have a powerful platform to leverage the multiple levels of parallelism in their applications.

In the current HPC environment, systems with heterogeneous node architectures that offer multiple levels of parallelism are ubiquitous. Furthermore, the next generation of systems may incorporate GPU-like accelerators along with other accelerators to enhance performance for a wider range of application kernels. This would introduce additional complexity for application programmers, as different programming languages and frameworks like CUDA and HIP may be required for each architectural component within a compute node. Such specialization adds challenges to maintenance and portability across different systems. Consequently, the importance of programming approaches that can deliver performance, scalability, and portability while exploiting the available parallelism is increasing. It is highly desirable for programmers to maintain a single code base to simplify maintenance and avoid the need to debug and optimize multiple versions of the same code.

To fully exploit the available parallelism in these systems, applications need to be refactored and a programming approach that can utilize accelerators must be employed. Historically, the preferred portable approaches, which were the main focus of earlier workshops, included OpenMP offloading and OpenACC, both based on directives. However, as technology has evolved to accommodate heterogeneity, other options have emerged. Starting in 2021, the workshop extended its scope to encompass the use of standard Fortran/C++, SYCL, DPC++, Kokkos, and RAJA, among other alternatives. These alternatives offer scalable and portable solutions without compromising performance. Programmers expect the software community to provide solutions that enable the maintenance of a single code base whenever possible, thus minimizing duplicated efforts across programming models and architectures.

Software abstraction-based programming models like OpenMP and OpenACC have been serving this purpose for several years and are likely to continue being relevant. These programming models address the “X” component in a hybrid MPI+X programming approach by providing high-level directives to programmers and offloading some of the burden to the compiler. With the increasing importance of other programming models (e.g., SYCL, DPC++, Kokkos, and RAJA) in this workshop, there may be additional challenges and opportunities for efficiently distributing computations across multiple nodes.

Important Deadlines for the Workshop:

• Paper Submission Deadline: August 4, 2023
• Author Notification: September 8, 2023
• Camera Ready Deadline: September 29, 2023