considered to be local objects list of participating processes.
Since this list might have changed during recovery, all user fined groups are considered to be potentially out of date.
After the recovery operation, the user has access to the same non-local operations like after MPI Init. These are:
Communicators: MPI COMM WORLD and MPI COMM SELF.
Rationale It would be theoretically possible to modify non-local objects on the surviving processes such, that they contain the up-to-date information of the run-time environment. However, assuming that failed processes are replaced by the run-time envi- ronment (see the following section) there is no MPI function call to pass the additional handles to the re-spawned processes in a portable, MPI conforming manner.
Groups and Communicators can have di erent formats after the recovery procedure, depending on the communicator mode. The communicator mode specifies, how the run-time environment should treat failed processes. Four modes are currently defined:
1. FTMPI COMM MODE ABORT: like in MPI-1 and MPI-2, the MPI library will abort the execution if one or several processes have failed. This mode is available for backward compatibility.
2. FTMPI COMM MODE REBUILD: failed processes will be replaced by the run-time environment. Surviving processes will retain their rank in MPI COMM WORLD. No assumptions are made within the FT-MPI specification where the new processes are placed.
3. FTMPI COMM MODE BLANK: failed processes will not be replaced, the size of MPI COMM WORLD will remain unchanged. However, the failed processes are blanked out and treated similarly to MPI- PROC NULL. Detailed specifications about operations using blank processes can be found in the next subsections.
4. FTMPI COMM MODE SHRINK: failed processes will not be replaced. The size of MPI COMM WORLD will be adjusted to the number of