sgoverview.tex

来自「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」· TEX 代码 · 共 29 行

\section{Communication Buffer management}

 The routines in this section fill two roles.  In the implementation of
 routines such as \texttt{MPID_Isend} where the only device-specific routines are 
 the few in \texttt{MPID_CORE}, it is necessary to have the ability to pack and 
 unpack partial messages.  Consider the case where the user calls
 \texttt{MPI_Isend( buf, 1, my_huge_indexed_type, ... )}
 where the total amount of data being sent is 100 MB.  Since \texttt{MPID_CORE}
 can only move contiguous data, we need to convert the described data
 into a sequence of contiguous memory segments of some reasonable size
 (we don\texttt{}t want to have to allocate a 100 MB temporary buffer).
 Thus, to implement this operation, we need a routine that can be called 
 repeatedly to get the next m bytes from the user\texttt{}s described data buffer.
 Further, we need to be able to pause and remember where we are in processing
 a datatype.  The MPI unpack routine does not have this flexibility.

 The other place where these routines are needed is in the implementation
 of efficient versions of the MPI collective communication routines.  
 These algorithms often need to look at the message as a range of bytes
 from which segments are extracted and moved.  The implementation of the
 MPI collective routines provided with MPICH will use these routines
 express the algorithms.

 These routines also provide a way to specify a data area that may be
 used in store-and-forward algorithms, without requiring copies to and
 from an intermediate device buffer.  

 Note that \texttt{MPID_Request}s include a segment descriptoin within them.