mpid_comm_thread_lock.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`\_}Comm{\tt \char`\_}thread{\tt \char`\_}lock}{tex}{10/7/2002}
\manname{MPID{\tt \char`\_}Comm{\tt \char`\_}thread{\tt \char`\_}lock}
--- Acquire a thread lock for a communicator 
\subhead{Synopsis}
\startvb\begin{verbatim}
void MPID_Comm_thread_lock( MPID_Comm *comm )

\end{verbatim}
\endvb

\subhead{Input Parameter}
\startarg{comm }{Communicator to lock
}
\par
\subhead{Notes}
This routine acquires a lock among threads in the same MPI process that
may use this communicator.  In all MPI thread modes except for
{\tt MPI{\tt \char`\_}THREAD{\tt \char`\_}MULTIPLE}, this can be a no-op.  In an MPI implementation
that does not provide {\tt MPI{\tt \char`\_}THREAD{\tt \char`\_}MULTIPLE}, this may be a macro.
\par
It is invalid for a thread that has acquired the lock to attempt to
acquire it again.  The lock must be released by {\tt MPID{\tt \char`\_}Comm{\tt \char`\_}thread{\tt \char`\_}unlock}.
\par
Note that there is also a common per-process lock ({\tt common{\tt \char`\_}lock}).
That lock should be used instead of a lock on lock on {\tt MPI{\tt \char`\_}COMM{\tt \char`\_}WORLD} when
a lock across all threads is required.
\par
A high-quality implementation may wish to provide fair access to the lock.
\par
In general, the MPICH implementation tries to avoid using locks because
they can cause problems such as livelock and deadlock, particularly when
an error occurs.  However, the semantics of MPI collective routines make
it difficult to avoid using locks.  Further, good programming practice by
MPI programmers should be to avoid having multiple threads using the
same communicator.
\par
\subhead{Module}
Communicator
\par
\subhead{See Also}
{\tt MPID{\tt \char`\_}Comm{\tt \char`\_}thread{\tt \char`\_}unlock
}
\par
\subhead{Questions}
Do we also need versions of this for datatypes and window objects?
For example, communicators, datatypes, and window objects all have
attributes; do we need a thread lock for each type?  Should we instead have
an MPI Object, on which some common operations, such as thread lock,
reference count, and name are implemented?
\endmanpage