localpid.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{LocalPID}{tex}{10/7/2002}
\manname{LocalPID}
--- Description of the local process ids MPI specifies communication among processes.  The MPI object that
describes collections of processes is the {\tt MPI{\tt \char`\_}Group}.  In MPI-1,
processes are uniquely identified by their rank in {\tt MPI{\tt \char`\_}COMM{\tt \char`\_}WORLD}.
In MPI-2, the situation is more complex, since processes (beyond
{\tt MPI{\tt \char`\_}COMM{\tt \char`\_}WORLD}) may come and go as a result of calls to
routines such as {\tt MPI{\tt \char`\_}Comm{\tt \char`\_}spawn} and {\tt MPI{\tt \char`\_}Comm{\tt \char`\_}detach}.  To
implement the various operations on MPI groups, it is very helpful to
have an enumeration of the processes.  This enumeration need not be
consequtive (gaps are acceptable) and it need not correspond to the
enumeration of the same processes on a different MPI process.  That is,
such an enumeration need only be local.  These are called {\tt LocalPID}s,
and are allocated by the MPI routines that change the number of
processes that are connected (for most applications, a simple counter
could be used).
\par
Local process ids are used only by the {\tt MPI{\tt \char`\_}Group} routines and for caching
datatypes in the RMA routines.  Communication
uses {\em virtual connections} to select the appropriate communication path
from one process to another.  This is different from the implementation
of MPICH using ADI-1 and ADI-2 where the mapping from communicator rank
to rank in {\tt MPI{\tt \char`\_}COMM{\tt \char`\_}WORLD} (through the {\tt MPI{\tt \char`\_}Group} associated with
a communicator) was used to identify the process with which to communicate.
See the discussion of the {\tt MPI{\tt \char`\_}Group} routines in the MPICH Design
Document for more details on the use of local process ids.
\endmanpage