[libfabric-users] Cannot establish a connection with verbs; ofi_rxm

Hefty, Sean sean.hefty at intel.com
Mon Jun 22 10:41:16 PDT 2020


> 3. The client sends its socket address in a message to the server and
> waits for a completion.

If the client sends its address to the server, you should not need FI_SOURCE.

> However, it seems the client-side `fi_cq_read` is always returning
> `FI_EAGAIN`.  Also the `FI_CONTEXT` mode doesn't appear in the output of
> `fi_tostr(pv, FI_TYPE_INFO)`.

Mode bits that are not needed by the provider can be cleared.  I'm not sure what's happening on line 299 in the log output.

> The logs[0] show at line #178-187 that no verbs provider have been
> found, however `fi_getinfo` still returns a valid verb provider.  I am
> confused, could you please shed some light on this?

I'm confused looking at the logs as well.  What version of libfabric are you using?

I believe verbs builds a list of fi_info structs to match against hints.  Some of those will not match, which could be what you're seeing.  However, the logs really need more detail to help here, possibly adding some messages for positive matches.

- Sean


More information about the Libfabric-users mailing list