%
% This file was automatically produced at Feb 27 2003, 22:51:56 by
% c2latex -c mpithreads_mpi.c
%
\documentclass[11pt,a4paper]{article}
\setlength{\textwidth}{15cm}
\setlength{\textheight}{22.5cm}
\setlength{\hoffset}{-2cm}
\setlength{\voffset}{-2cm}
\begin{document}
\expandafter\ifx\csname indentation\endcsname\relax%
\newlength{\indentation}\fi
\setlength{\indentation}{0.5em}
\begin{flushleft}
{$/\ast$\it{}$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$\mbox{}\\
$\ast$ FILE: mpithreads\_mpi.c\mbox{}\\
$\ast$ DESCRIPTION:\mbox{}\\
$\ast$   This simple program illustrates the use of MPI in a program obtained \mbox{}\\
$\ast$   by modifying a serial code that performs a dot product. It is the third \mbox{}\\
$\ast$   of four codes used to show the progression from a serial program to a \mbox{}\\
$\ast$   hybrid MPI/Pthreads program.  The other relevant codes are:\mbox{}\\
$\ast$      $-$ mpithreads\_serial.c   $-$ The serial version\mbox{}\\
$\ast$      $-$ mpithreads\_threads.c  $-$ A shared memory programming model using\mbox{}\\
$\ast$          Pthreads\mbox{}\\
$\ast$      $-$ mpithreads\_both.c $-$ A hybrid model that utilizes both MPI and\mbox{}\\
$\ast$          Pthreads to execute on systems that are comprised of clusters\mbox{}\\
$\ast$          of SMP's.\mbox{}\\
$\ast$\mbox{}\\
$\ast$   Use of the SPMD model was chosen and for convenience, with replication \mbox{}\\
$\ast$   of the main data on all nodes. A more memory efficient implementation \mbox{}\\
$\ast$   would be advisable for larger data sets.\mbox{}\\
$\ast$\mbox{}\\
$\ast$ SOURCE: Vijay Sonnad, IBM\mbox{}\\
$\ast$ LAST REVISED:  10/8/99 Blaise Barney\mbox{}\\
$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast$$\ast/$}\mbox{}\\
\mbox{}\\
{\tt \#include} $<${\tt{}mpi.h}$>$\mbox{}\\
{\tt \#include} $<${\tt{}stdio.h}$>$\mbox{}\\
{\tt \#include} $<${\tt{}malloc.h}$>$\mbox{}\\
\mbox{}\\
{$/\ast$\it{}   \mbox{}\\
The following structure contains the necessary information to allow the \mbox{}\\
function "dotprod" to access its input data and place its output into \mbox{}\\
the structure.  Note that this structure is unchanged from the sequential \mbox{}\\
version.\mbox{}\\
$\ast/$}\mbox{}\\
\mbox{}\\
{\bf typedef} {\bf struct} \mbox{}\\
\hspace*{1\indentation}\{\mbox{}\\
\hspace*{3\indentation}{\bf double}      $\ast$a;\mbox{}\\
\hspace*{3\indentation}{\bf double}      $\ast$b;\mbox{}\\
\hspace*{3\indentation}{\bf double}     sum; \mbox{}\\
\hspace*{3\indentation}{\bf int}     veclen; \mbox{}\\
\hspace*{1\indentation}\} DOTDATA;\mbox{}\\
\mbox{}\\
{$/\ast$\it{} Define globally accessible variables $\ast/$}\mbox{}\\
\mbox{}\\
{\tt \#define} VECLEN 100\mbox{}\\
DOTDATA dotstr; \mbox{}\\
\mbox{}\\
{$/\ast$\it{}\mbox{}\\
The function dotprod is very similar to the sequential version except that \mbox{}\\
we now have each node working on a different part of the data. As before, \mbox{}\\
all access to the input is through a structure of type DOTDATA and all \mbox{}\\
output from this function is written into this same structure. \mbox{}\\
$\ast/$}\mbox{}\\
\mbox{}\\
{\bf void} $\ast$dotprod()\mbox{}\\
\{\mbox{}\\
\mbox{}\\
\hspace*{3\indentation}{$/\ast$\it{} Define and use local variables for convenience $\ast/$}\mbox{}\\
\mbox{}\\
\hspace*{3\indentation}{\bf int} i, start, end, myid, len;\mbox{}\\
\hspace*{3\indentation}{\bf double} mysum, $\ast$x, $\ast$y;\mbox{}\\
\hspace*{3\indentation}\mbox{}\\
\hspace*{3\indentation}{$/\ast$\it{} Obtain rank of this node $\ast/$}\mbox{}\\
\mbox{}\\
\hspace*{3\indentation}MPI\_Comm\_rank (MPI\_COMM\_WORLD, \&myid);\mbox{}\\
\hspace*{5\indentation}\mbox{}\\
\hspace*{3\indentation}len = dotstr.veclen;\mbox{}\\
\hspace*{3\indentation}start = myid$\ast$len;\mbox{}\\
\hspace*{3\indentation}end   = start + len;\mbox{}\\
\hspace*{3\indentation}x = dotstr.a;\mbox{}\\
\hspace*{3\indentation}y = dotstr.b;\mbox{}\\
\mbox{}\\
\hspace*{3\indentation}{$/\ast$\it{}\mbox{}\\
\hspace*{3\indentation}Perform the dot product and assign result to the appropriate variable in \mbox{}\\
\hspace*{3\indentation}the structure. \mbox{}\\
\hspace*{3\indentation}$\ast/$}\mbox{}\\
\mbox{}\\
\hspace*{3\indentation}mysum = 0;\mbox{}\\
\hspace*{3\indentation}{\bf for} (i=start; i$<$end ; i++) \mbox{}\\
\hspace*{4\indentation}\{\mbox{}\\
\hspace*{6\indentation}mysum += (x[i] $\ast$ y[i]);\mbox{}\\
\hspace*{4\indentation}\}\mbox{}\\
\mbox{}\\
\hspace*{3\indentation}dotstr.sum += mysum;\mbox{}\\
\mbox{}\\
\}\mbox{}\\
\mbox{}\\
{$/\ast$\it{} \mbox{}\\
As before,the main program does very little computation. It does however make \mbox{}\\
all the calls to the MPI routines. This is not a master$-$worker arrangement \mbox{}\\
and all nodes participate equally in the work. \mbox{}\\
$\ast/$}\mbox{}\\
\mbox{}\\
{\bf int} main ({\bf int} argc, {\bf char}$\ast$ argv[])\mbox{}\\
\{\mbox{}\\
\hspace*{3\indentation}{\bf int} i,len=VECLEN;\mbox{}\\
\hspace*{3\indentation}{\bf int} myid, numprocs;\mbox{}\\
\hspace*{3\indentation}{\bf double} $\ast$a, $\ast$b;\mbox{}\\
\hspace*{3\indentation}{\bf double} mysum, allsum;\mbox{}\\
\mbox{}\\
\hspace*{3\indentation}{$/\ast$\it{} MPI Initialization $\ast/$}\mbox{}\\
\mbox{}\\
\hspace*{3\indentation}MPI\_Init (\&argc, \&argv);\mbox{}\\
\hspace*{3\indentation}MPI\_Comm\_size (MPI\_COMM\_WORLD, \&numprocs);\mbox{}\\
\hspace*{3\indentation}MPI\_Comm\_rank (MPI\_COMM\_WORLD, \&myid);\mbox{}\\
\mbox{}\\
\hspace*{3\indentation}{$/\ast$\it{} Assign storage and initialize values $\ast/$}\mbox{}\\
\mbox{}\\
\hspace*{3\indentation}a = ({\bf double}$\ast$) malloc (numprocs$\ast$len$\ast${\bf sizeof}({\bf double}));\mbox{}\\
\hspace*{3\indentation}b = ({\bf double}$\ast$) malloc (numprocs$\ast$len$\ast${\bf sizeof}({\bf double}));\mbox{}\\
\hspace*{2\indentation}\mbox{}\\
\hspace*{3\indentation}{\bf for} (i=0; i$<$len$\ast$numprocs; i++)\mbox{}\\
\hspace*{4\indentation}\{\mbox{}\\
\hspace*{5\indentation}a[i]=1;\mbox{}\\
\hspace*{5\indentation}b[i]=a[i];\mbox{}\\
\hspace*{4\indentation}\}\mbox{}\\
\mbox{}\\
\hspace*{3\indentation}dotstr.veclen = len; \mbox{}\\
\hspace*{3\indentation}dotstr.a = a; \mbox{}\\
\hspace*{3\indentation}dotstr.b = b; \mbox{}\\
\hspace*{3\indentation}dotstr.sum=0;\mbox{}\\
\mbox{}\\
\hspace*{3\indentation}{$/\ast$\it{} Call  the  dot product routine $\ast/$}\mbox{}\\
\mbox{}\\
\hspace*{3\indentation}dotprod();\mbox{}\\
\hspace*{3\indentation}mysum = dotstr.sum;\mbox{}\\
\hspace*{3\indentation}printf({\tt"Task \%d partial sum is \%f$\backslash$n"},myid, mysum);\mbox{}\\
\mbox{}\\
\hspace*{3\indentation}{$/\ast$\it{} After the dot product, perform a summation of results on each node $\ast/$}\mbox{}\\
\hspace*{3\indentation}MPI\_Reduce (\&mysum, \&allsum, 1, MPI\_DOUBLE, MPI\_SUM, 0, MPI\_COMM\_WORLD);\mbox{}\\
\mbox{}\\
\hspace*{3\indentation}{\bf if} (myid == 0) \mbox{}\\
\hspace*{3\indentation}printf ({\tt"Done. MPI version: sum  $=$  \%f $\backslash$n"}, allsum);\mbox{}\\
\hspace*{3\indentation}free (a);\mbox{}\\
\hspace*{3\indentation}free (b);\mbox{}\\
\hspace*{3\indentation}MPI\_Finalize();\mbox{}\\
\}   \mbox{}\\
\end{flushleft}
\end{document}