Exercises

1

What is the largest size of SIZE, for which MPI_Type_construct_darray is still going to work? Test it using the program discussed in section 5.3.4.

2

Would it help if we replaced MPI_INT with a more elaborate structure that would contain more data?

3

The restriction on the extent of file_type in section 5.3.4 does not imply the restriction on the size of the file. It is only a restriction on the amount of data that can be written on the file in a single MPI_File_write operation. But this operation can be repeated. Rewrite the program discussed in section 5.3.4 in the following way:

• Instead of one array write_buffer and read_buffer define and malloc several, e.g., write_buffer_1, write_buffer_2, etc., of identical size, but make sure that you have enough memory for them all. Aim for not more than a GB of memory per process.
• Populate the write_buffer_? arrays with random integers, using a different random number generator seed for every process.
• Write them all on the file in multiple MPI_File_write_all operations, one after another. Time the whole data transfer and print MB/s obtained on standard output.
• Read them back from the file on the read_buffer_? arrays and compare every entry with the corresponding entry in the write_buffer_? arrays. Time the data transfer and print MB/s obtained on standard output.
• Replace MPI_DISTRIBUTE_BLOCK with MPI_DISTRIBUTE_CYCLIC  and compare the data transfer rates.
• Run the program on up to 32 nodes under PBS, generating files up to 16GB long. Remember that half of the 1GB memory per process will have to be reserved for read_buffer_?.