On Wed, Aug 16, 2017 at 2:24 AM, Simon Schwitanski <simon.schwitanski@rwth-aachen.de> wrote:

Hello,

I have got a question about the atomicity of accumulate operations as
described in 11.7.1. It says (page 461, lines 8 - 11):

> The previous correctness conditions imply that a location updated by a
> call to an accumulate operation cannot be accessed by a load or an RMA
> call other than accumulate until the accumulate operation has completed
> (at the target).

I am a little bit confused about what this actually means: Does it mean
that local load accesses on the target are (somehow) delayed if there is
an ongoing accumulate operation to the same location? Or does it just
mean that concurrent load accesses are forbidden and lead to undefined
behavior? And why does this not hold for store accesses?

The MPI standard does not specify how atomicity is implemented. It can use hardware atomics or regular instructions with mutual exclusion. Both of these serialize, but at different granularity.

The behavior of concurrency MPI_Accumulate and direct local access is not defined, although it is allowed. If you want to read memory concurrently with an MPI_Accumulate, use MPI_Get_accumulate(MPI_NO_OP) or MPI_Fetch_op(MPI_NO_OP).

Jeff

Thanks!

Simon
_______________________________________________
mpiwg-rma mailing list
mpiwg-rma@lists.mpi-forum.org
https://lists.mpi-forum.org/mailman/listinfo/mpiwg-rma

Jeff Hammond
jeff.science@gmail.com
http://jeffhammond.github.io/