Do All-to-All collectives *must* synchronize?

[mknobi]

Problem

Remark 18 in Appendix A.2 (Summary of the Semantics of all Operation-Related MPI Procedures) states

Based on their semantics, when called using an intra-communicator, MPI_ALLGATHER,
MPI_ALLTOALL, and their V and W variants, MPI_ALLREDUCE, MPI_REDUCE_SCATTER,
and MPI_REDUCE_SCATTER_BLOCK must synchronize (i.e., S1/S2 instead of W1/W2) provided that all counts and the size of all datatypes are larger than zero

In the March Forum meeting the question came up whether the hard guarantee (must synchronize) is actually true or if there are cases which can be implemented (in some smart way) that doesn't require synchronization.

Proposal

Check whether that statement is actually true replace it if not.

Changes to the Text

Remove the hard requirement if not necessary

Impact on Implementations

All-to-all collectives could be implemented in a non-synchronizing way

Impact on Users

Users can't rely on synchronization properties of collective routines

References and Pull Requests

[jprotze]

The formulation "must synchronize" is probably too much, but it will not be weaker than "Based on their semantics ... will synchronize (i.e., S1/S2 instead of W1/W2) provided that all counts and the size of all datatypes are larger than zero.

Under the given condition (no empty messages !), all participating processes need to contribute something to the resulting buffer content of each process. This means, that all processes need to start the collective operation before any process can complete its operation. So, from Users perspective the function behaves as synchronizing call.

[jeffhammond]

All-to-all collectives could be implemented in a non-synchronizing way

False, assuming no zero counts.

[jeffhammond]

Because zero counts are allowed, absolutely no MPI collectives are guaranteed to synchronize except Barrier.

[mknobi]

It states

must synchronize (i.e., S1/S2 instead of W1/W2) provided that all counts and the size of all datatypes are larger than zero

The question came up - and I doubt that's possible - if there are possible implementations of non-synchronizing "complete" all-to-all collectives.

[RolfRabenseifner]

Alltoall, allgather and allreduce can return (i.e. complete) in a process only, if all processes of the communicator have provided their input data (provided that count > 0, which is clearly defined in the Remark 18), i.e. have started the routine. That a collective rutine can return in a process only when all processes of the communicator have started the corresponding call, is exactly the defineion of a barrier synchronization.
Therefore Remark 18 is correct.
@mknobi I recommend that you close this issue if you agree with my proof.