surviving processes. This includes also, that the ranks of some pro- cesses in MPI COMM WORLD will change. FT-MPI requires that the sequence of surviving processes is identical before and after recov- ery. Figure 4.1 is showing an example, where two out of four processes fail, how the ranks are assigned after recovery.
M P I _ C O M M _ W O R size=4 L
M P I _ C O M M _ W O R L size = 2 D
Figure 4.1: Example FTMPI COMM MODE SHRINK.
FTMPI COMM MODE ABORT and FTMPI COMM MODE REBUILD require no changes to the MPI-1 and MPI-2 specification after recovery. The communicator modes FTMPI COMM MODE BLANK and FTMPI- COMM MODE SHRINK introduce some new aspects to MPI and are de- tailed in the following.
FTMPI COMM MODE BLANK
Point to point operations A blank process is defined to behave like MPI PROC NULL in the MPI-1 specification. This includes, that sending a message to a blank process will not raise an error, however no data is transmitted. Receive operations from a blank processes will return a null- status (see section 3.11 in MPI-1), the receiver bu er is unchanged.
Collective operations For collective operations, two di erent issues have to be taken into account for the blank mode. If the root of one of the rooted collective operations (MPI Bcast, MPI Reduce, MPI Gather(v), MPI Scatter(v)) is a blank process all processes will return immediatly. No input or output bu er is modified.