At the recent International Conference on High Performance Computing, Networking, Storage, and Analysis (SC’17), Michael LeBeane from AMD presented a research paper entitled “GPU Triggered Networking for Intra-Kernel Communications.” The paper describes an approach using triggered operations to improve networking performance for systems, such as the Titan machine at Oak Ridge National Laboratory, that employ both traditional processors and graphics processing units (GPUs). Triggered operations are an event-driven technique for initiating data transfers directly from a network interface controller (NIC). Using triggered operations for GPUs is shown to reduce latency performance by 35% compared to other methods. Triggered operations are a key capability of Sandia’s Portals 4 network programming interface, which is intended to enable the design and development of high-performance interconnect hardware. In addition to leveraging Portals 4 triggered operations to improve GPU communication, the simulation environment used in the study is also based on a model of a Portals 4 NIC.
November 1, 2017