Project

General

Profile

Feature #1973

GPU Messages

Added by Michael Robson 19 days ago.

Status:
New
Priority:
Normal
Category:
GPU Support
Target version:
Start date:
09/06/2018
Due date:
10/29/2018
% Done:

0%


Description

Enable Charm++ to send RDMA-style messages to/from buffers on the GPU. Our initial plan is to use the RDMA syntax. The runtime should detect whether the source and/or target buffers are resident on a local or remote GPU and invoke the corresponding GPU Direct call. We will test this approach with benchmarks (e.g. stencil2d) against GPU-aware MPI versions on multi-GPU clusters (e.g. Bridges, Comet, etc) to ensure we have comparable performance.

Also available in: Atom PDF