mpid_abort.tex

刚才是说明 现在是安装程序在 LINUX环境下进行编程的MPICH安装文件

\startmanpage
\mantitle{MPID{\tt \char`\_}Abort}{tex}{10/7/2002}
\manname{MPID{\tt \char`\_}Abort}
--- Abort at least the processes in the specified communicator. 
\subhead{Synopsis}
\startvb\begin{verbatim}
int MPID_Abort( MPID_Comm *comm, int return_code )

\end{verbatim}
\endvb

\subhead{Input Parameters}
\startarg{comm        }{Communicator of processes to abort
}
\startarg{return{\tt \char`\_}code }{Return code to return to the calling environment.  See notes.
}
\par
\subhead{Return value}
{\tt MPI{\tt \char`\_}SUCCESS} or an MPI error code.  Normally, this routine should not
return, since the calling process must be a member of the communicator.
However, under some circumstances, the {\tt MPID{\tt \char`\_}Abort} might fail; in this
case, returning an error indication is appropriate.
\par
\subhead{Notes}
\par
In a fault-tolerant MPI implementation, this operation should abort {\em only
}the processes in the specified communicator.  Any communicator that shares
processes with the aborted communicator becomes invalid.  For more
details, see (paper not yet written on fault-tolerant MPI).
\par
In particular, if the communicator is {\tt MPI{\tt \char`\_}COMM{\tt \char`\_}SELF}, only the calling
process should be aborted.
\par
The {\tt return{\tt \char`\_}code} is the return code that this particular process will
attempt to provide to the {\tt mpiexec} or other program invocation
environment.  See {\tt mpiexec} for a discussion of how return codes from
many processes may be combined.
\par
An external agent that is aborting processes can invoke this with either
{\tt MPI{\tt \char`\_}COMM{\tt \char`\_}WORLD} or {\tt MPI{\tt \char`\_}COMM{\tt \char`\_}SELF}.  For example, if the process manager
wishes to abort a group of processes, it should cause {\tt MPID{\tt \char`\_}Abort} to
be invoked with {\tt MPI{\tt \char`\_}COMM{\tt \char`\_}SELF} on each process in the group.
\par
\subhead{Question}
An alternative design is to provide an {\tt MPID{\tt \char`\_}Group} instead of a
communicator.  This would allow a process manager to ask the ADI
to kill an entire group of processes without needing a communicator.
However, the implementation of {\tt MPID{\tt \char`\_}Abort} will either do this by
communicating with other processes or by requesting the process manager
to kill the processes.  That brings up this question: should
{\tt MPID{\tt \char`\_}Abort} use {\tt BNR} to kill processes?  Should it be required to
notify the process manager?  What about persistent resources (such
as SYSV segments or forked processes)?
\par
This suggests that for any persistent resource, an exit handler be
defined.  These would be executed by {\tt MPID{\tt \char`\_}Abort} or {\tt MPID{\tt \char`\_}Finalize}.
See the implementation of {\tt MPI{\tt \char`\_}Finalize} for an example of exit callbacks.
In addition, code that registered persistent resources could use persistent
storage (i.e., a file) to record that information, allowing cleanup
utilities (such as {\tt mpiexec}) to remove any resources left after the
process exits.
\par
{\tt MPI{\tt \char`\_}Finalize} requires that attributes on {\tt MPI{\tt \char`\_}COMM{\tt \char`\_}SELF} be deleted
before anything else happens; this allows libraries to attach end-of-job
actions to {\tt MPI{\tt \char`\_}Finalize}.  It is valuable to have a similar
capability on {\tt MPI{\tt \char`\_}Abort}, with the caveat that {\tt MPI{\tt \char`\_}Abort} may not
guarantee that the run-on-abort routines were called.  This provides a
consistent way for the MPICH implementation to handle freeing any
persistent resources.  However, such callbacks must be limited since
communication may not be possible once {\tt MPI{\tt \char`\_}Abort} is called.  Further,
any callbacks must guarantee that they have finite termination.
\par
One possible extension would be to allow {\em users} to add actions to be
run when {\tt MPI{\tt \char`\_}Abort} is called, perhaps through a special attribute value
applied to {\tt MPI{\tt \char`\_}COMM{\tt \char`\_}SELF}.  Note that is is incorrect to call the delete
functions for the normal attributes on {\tt MPI{\tt \char`\_}COMM{\tt \char`\_}SELF} because MPI
only specifies that those are run on {\tt MPI{\tt \char`\_}Finalize} (i.e., normal
termination).
\par
\subhead{Module}
MPID{\tt \char`\_}CORE
\endmanpage