[ofiwg] slight API enhancement proposal for msg ordering

Hefty, Sean sean.hefty at intel.com
Thu May 9 09:34:50 PDT 2019


The following PR defines new ordering bits for message processing:

https://github.com/ofiwg/libfabric/pull/5019

Current message ordering captures [Read, Write, Send] after [Read, Write, Send] semantics.  E.g. RAR, RAW, RAS, WAR, WAW, etc.  The read/write ordering includes both atomic and RMA operations, meaning that atomic and RMA operations must be ordered with each other for an ordering bit to be set.

The proposed change allows specifying atomic ordering separately from RMA ordering.  The existing definitions are unmodified.  These 8 new bits are introduced:

FI_ORDER_RMA_[RAR, RAW, WAR, WAW]
FI_ORDER_ATOMIC_[RAR, RAW, WAR, WAW]

New bits were *not* added for rma/atomic specific ordering wrt sends.  This was done to limit the number of new bits that are used, and I'm not aware of a use case for them not covered by the existing bits.

- Sean


More information about the ofiwg mailing list