2024/09/30 by Alex Brooks, Philip Marshall, Brooks, Alex +9 · 1 citation
Computer Science · Physics and Astronomy · #Parallel Computing and Optimization Techniques #Advanced Data Storage Technologies #Model Reduction and Neural Networks
paper · pdf · doi:10.48550/arxiv.2409.20476
Modern high-end systems are increasingly becoming heterogeneous, providing users options to use general purpose Graphics Processing Units (GPU) and other accelerators for additional performance. High Performance Computing (HPC) and Artificial Intelligence (AI) applications are often carefully arranged to overlap communications and computation for increased efficiency on such platforms. This has led to efforts to extend popular communication libraries to support GPU awareness and more recently, GPU-initiated operations. In this paper, we present Intel SHMEM, a library that enables users to write programs that are GPU aware, in that API calls support GPU memory, and also support GPU-initiated communication operations by embedding OpenSHMEM style calls within GPU kernels. We also propose thread-collaborative extensions to the OpenSHMEM standard that can enable users to better exploit the strengths of GPUs. Our implementation adapts to choose between direct load/store from GPU and the GPU copy engine based transfer to optimize performance on different configurations.