-
Notifications
You must be signed in to change notification settings - Fork 8
Description
Problem
In §11.7.2 of the MPI standard, the text states:
MPI specifies ordering between accumulate operations from one process to the same (or overlapping) memory locations at another process on a per-datatype granularity.
This can be read to not include operations with from one process to itself (origin == target
) [1]. However, there is no reason to exclude operations to self
from the ordering guarantees (and it seems to be unintentional [2]).
Proposal
By consistently using the terminology of origin
and target
established in the RMA chapter this ambiguity can be avoided:
MPI specifies ordering between accumulate operations
from an origin process to the same (or overlapping) memory locations at
a target process on a per-datatype granularity.
This proposal fixes another instance where source
is used instead of origin
.
Changes to the Text
See above.
PR: https://github.com/mpi-forum/mpi-standard/pull/127
Impact on Implementations
None
Impact on Users
It is made clear that accumulate operations from a process to itself are ordered according to the ordering rules outlined in §11.7.2.
References
[1] Merriam Webster: https://www.merriam-webster.com/dictionary/another
Definition of another
1 : different or distinct from the one first considered
[2] RMA WG issue: mpiwg-rma/rma-issues#11