Authors: Stephen Poole (Los Alamos National Laboratory (LANL)), Oscar Hernandez (Oak Ridge National Laboratory (ORNL)), Michael Beebe (Texas Tech University), Naveen Namashivayam (HPE), Benjamin Michalowicz (The Ohio State University), Manjunath Gorentla Venkata (NVIDIA Corporation), Aurelien Bouteiller (Advanced Micro Devices (AMD)), Mark Brown (Intel Corporation), Aaron Welch (Oak Ridge National Laboratory (ORNL)), Nathan Wichmann (HPE)
Abstract: OpenSHMEM is a community-driven partitioned global address space (PGAS) API for low-overhead, one-sided asynchronous communication across diverse HPC architectures. This BoF will present progress since the 1.6 specification and discuss ongoing work toward OpenSHMEM 1.7, the proposed GPU auxiliary specification for accelerator support, and the longer-term OpenSHMEM 2.0 roadmap. As heterogeneous systems become the norm, the session will bring together users, implementers, middleware and tool developers, hardware vendors, and HPC facilities to share experiences, discuss evolving requirements, and help shape a simple, portable, and vendor-neutral future for OpenSHMEM.
Long Description: OpenSHMEM is a community-driven implementation of the Partitioned Global Address Space (PGAS) programming model that provides a simple, efficient API for low-overhead one-sided communication across diverse HPC architectures. Supported by multiple vendors, national laboratories, and open-source implementations, OpenSHMEM enables scalable asynchronous communication through remote memory operations, atomics, synchronization primitives, and collective operations while maintaining a compact programming model. These characteristics have made OpenSHMEM attractive for applications requiring fine-grained communication, irregular communication patterns, graph analytics, sparse algorithms, AI workloads, and other communication-intensive scientific applications. As HPC systems continue to evolve toward increasingly heterogeneous architectures, OpenSHMEM provides a lightweight programming model capable of exploiting modern RDMA-capable interconnects while minimizing software overhead.
Since SC08, the OpenSHMEM BoF has served as the primary community forum for users, implementers, middleware developers, tool providers, hardware vendors, and researchers to discuss the evolution of the OpenSHMEM specification and ecosystem. The BoF provides a venue for sharing implementation experiences, presenting new capabilities, discussing application success stories, identifying challenges, and gathering community feedback that helps shape future versions of the specification.
A major focus of this year’s BoF is the evolution of OpenSHMEM for accelerator-based systems. Leadership-class supercomputers are now accelerator-centric, and applications increasingly keep data resident on GPUs while requiring efficient communication among accelerator memories. Existing implementations—including NVSHMEM, rocSHMEM, and Intel SHMEM—demonstrate the growing demand for GPU-aware SHMEM programming models while highlighting differences in memory allocation, completion semantics, execution contexts, progress models, and device capabilities. These experiences motivate the need for a common, vendor-neutral approach that preserves portability while allowing implementations to exploit hardware-specific capabilities.
The OpenSHMEM community is actively developing a GPU Auxiliary Specification as a practical, backward-compatible extension that addresses today’s GPU programming requirements while maintaining compatibility with existing OpenSHMEM applications. The Auxiliary Specification provides a common foundation for GPU-resident symmetric memory, accelerator communication, ordering and completion semantics, and capability discovery, while creating a path toward future support for advanced features such as device-initiated communication. The BoF will discuss the motivation, current status, implementation experience, and how this effort fits into the roadmap toward OpenSHMEM 1.7 and the longer-term vision for OpenSHMEM 2.0.
The session will feature lightning talks from users, implementers, hardware vendors, middleware developers, and tool providers highlighting recent accomplishments, implementation updates, ongoing research, and future plans for the OpenSHMEM ecosystem. The BoF will conclude with an open community discussion on priorities for future specifications, interoperability across implementations, heterogeneous memory systems, testing and conformance, developer tools, and a shared roadmap that will help ensure OpenSHMEM continues to meet the needs of future HPC and AI applications.