To: J3                                                     J3/23-245
From: Jeff Hammond
Subject: Asynchronous collective operations
Date: 2023-October-25

1. Introduction
===============

The current set of collective operations do not return until complete,
which does not permit optimizations that are possible with MPI,
OpenSHMEM or other lower-level software that can be used to implement
them.

Providing a way to make these operations asynchronous enables better
performance in some use cases, such as a sequence of collectives,
where the Fortran program only needs to observe the results at the
end of the sequence.  Another use case is overlapping collective
operations with computations, which is common in many scientific
simulation workloads.

2. Proposal
============

Add an optional argument to CO_{BROADCAST,MAX,MIN,REDUCE,SUM}
that indicates that an asynchronous collective is desired,
which is used to synchronize the operation.
The optional argument should be a scalar variable of event_type.
Completion of a collective operation increments the effective
value of the notify_type.  The program uses EVENT_QUERY to
determine that operations have completed.
Multiple collective operations may use the same event variable.
The event argument may not be coindexed.

3. Discussion
==============

Using event variables for synchronization allows the user to
test for completion without blocking, which is similar to what
MPI supports.  Using a notify type would not permit this usage.

4. Example
===========

program main
    use iso_fortran_env
    implicit none
    integer :: i, count
    type(event_type) :: broadcast_sync[*]
    integer :: x(100)
    x = this_image()
    do i = 1,size(x)
        call co_broadcast( x(i), source_image=1, event=broadcast_sync )
    end do
    event wait ( broadcast_sync, until_count = size(x) )
    print*,this_image(),' is finished'
end program main

5. References
==============

===END===