Does GDRcopy support the HPE/Cray "SlingShot" backbone?
I'm looking into a performance issue with an app. If you could tell me up-front whether you support this kind of cluster, it would save some troubleshooting time.
Carl, the libfabric plugin and the NCCL plugin have both been able to use GDRCopy on a SlingShot based machine.
See Jim's patch: https://github.com/aws/aws-ofi-nccl/pull/146
Do you know if UCX can use it? I'll check with the UCX people...
Do you mean "can UCX use GDRCopy?" ? I believe that UCX will use GDRCopy if the compile-time options / runtime environments are satisfied. The code is here: https://github.com/openucx/ucx/tree/master/src/uct/cuda/gdr_copy.