Implementation Issues

• Input to this program

  4 files(file0, file1, file2, file3). The whole data matrix is distributed on these 4 files. Each of 4 processors reads in one of files. This idea is about those cases we have larger data set than one processor can handle. If we break up data into files, we can scale better and handle more data.

  Just in case it is not clearly stated yet, the reason why we divide the data matrix into 4 files is that we assume each processor can handle only that one file at a time because of the limitation of memory size.

• Output of this program

  4 files(file0.out, file1.out, file2.out, file3.out). Those are results of computation.

• Transpose step

      p0        p1         p2        p3
    --------  --------  --------  --------
    |  X   |  |      |  |	     |  |      |
    |      |  |      |  |	     |	|      |
    --------  --------  --------  --------
    --------  --------  --------  --------
    |      |  |  X   |  |      |  |      |
    |      |  |      |  |      |  |      |
    --------  --------  --------  --------
    --------  --------  --------  --------
    |      |  |      |  |  X   |  |      |
    |      |  |      |  |      |  |      |
    --------  --------  --------  --------
    --------  --------  --------  --------
    |      |  |      |  |      |  | X    |
    |      |  |      |  |      |  |      |
    --------  --------  --------  --------
  
  

  Suppose there are 4 processors. At the above picture, the outer square means the whole data matrix. It is divided by 4 files and each file is charged by each processors. When we do transpose, each processor can divide it's own data set by 4 (number of processors) modules. Those modules marked 'X' stay, which means "do not involve MPI communication." Other 3 modules should be sent to other processors to achieve transpose.

• Compile and run

  mpicc trans.c -lm -o trans

  mpirun -np 4 -machinefile me trans