Hi all,
The Fault Tolerance Working Group will be reading the following ticket at the September 2019 meeting:
- Create MPI_INFO_ENV before MPI_INIT
- Issue #143
- Pull Request #126
- Add a new API function for constructing an MPI_INFO_ENV object when one might not exist (or might be wrong). This is sort of in conjunction with the Tools and Sessions working groups.
We will also hold a no-no and first vote on the following ticket:
Finally, we will have a second vote on these two issues:
- Clarify what MPI_SUCCESS guarantees, and what is left undefined when MPI_ERR_XXX happens
- Issue #119
- Pull Request #95
- The definition of MPI_SUCCESS is a terse "no error returned". When we operate under the assumption that some failure (process, resource, transport, otherwise) may happen, we need to clarify in which cases MPI_SUCCESS is possible, and what it guarantees. For example an implementation may mask errors that do not have a direct consequence on the call semantic.
Conversely, we need to clarify what arguments and buffers are valid or undefined after a call has raised an error.
- Define error/failure behavior in MPI_INIT/FINALIZE
- Issue #102
- Pull Request #50
- Errors before or during MPI_INIT may kill the entire application, which makes initializing MPI a possibly risky proposition in a mixed-mode application.