February 2019 OLCF User Conference Call: Advanced On-Node GPU Communication
The OLCF hosts monthly User Conference Calls. These calls are your opportunity to speak with center personnel to get the latest updates, express any concerns you may have, etc. No registration is required for this event.
Monthly Topic: Advanced On-Node GPU Communication
Speaker: Stephen Abbott, NVIDIA
Abstract: One strength of a Summit node is the NVLink2 between the V100 GPUs, allowing fast data movement between GPUs. In this session I’ll show how to exploit these connections using CUDA. We’ll begin with a brief review of GPU Direct technology and the Summit node topology. With that topology in mind, we’ll show how to enable and perform peer-to-peer copies in the single-process multi-device model. Finally, I’ll introduce the CUDA Interprocess communicators (IPC) and show how they enable GPU to GPU communication between processes.