[ofiwg] completion flags as actually defined by OFI

Weiny, Ira ira.weiny at intel.com
Tue Apr 14 15:12:49 PDT 2015

> > You mean the libfabric man page right?
> Yes - my original post was a copy-paste from the libfabric man page.  We're
> arguing over what OFI needs to do, and I'm guessing maybe four people have
> actually taken the time to read the documented behavior.  I need to know if the
> wording in the man pages is correct, or why it's not sufficient.  And if it's not
> sufficient, why it makes sense for a completion flag to describe the construction
> of a remote system's memory/cache/filesystem hierarchy.

I agree.  I tried to voice this on the call today but as some noticed I'm new to webex...  ;-)

IMO the fabric can indicate the following progress of a transaction (and potentially some other smaller sub-states which are not useful).

Initiator  -> provider -> wire -> provider -> data landed in a "posted" buffer

Where the buffer on the other end is (and if it is persistent or not) is outside the purview of libfabric.

It seems Jason is debating what the completions mean and where the data is (or the progress of a transaction is) when a particular completion is returned to the initiator.  I think that is a separate issue.


More information about the ofiwg mailing list