sean.hefty at intel.com
Tue Feb 15 14:33:32 PST 2022
> Rxm will layer over verbs and tcp, but not gni. Rxm layers over msg endpoints (this is
> what the 'm' stands for).
> You should be able to test FI_COLLECTIVE over either of those two.
> why not gni? it supports FI_EP_MSG doesn't it? How does one know if rxm is supported or
> not if we ignore the FI_EP_MSG entry?
It may work, but I doubt that path has ever been tested. It's not an expected combination.
The main purpose behind rxm is to add reliable datagram semantics over providers that support msg endpoints. Since gni already supports reliable datagram endpoints, rxm does not target it. But it shouldn't exclude it from being used, unless there's some check buried in the code to prevent it.
More information about the Libfabric-users