summaryrefslogtreecommitdiffstats
path: root/testpar/t_mpi.c
blob: dbcfadae70a6bf42b55de66259c10c509a486e51 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
/*
 * MPIO independent overlapping writes.
 *
 * First n-1 processes open 1 file.
 * Each of the n-1 process writes chunks of data to the file in round-robin
 * fashion, in a interleaved but not overlapped fashion.  Using increasing
 * chunk sizes for the benefits of testing different write sizes and also
 * reducing the numbers of writes.
 *
 * Last process (n-1) just waits.
 * First n-1 processes finish writing and cloose the file.
 * Last process opens the same file and verifies the data.
 */

#include <testphdf5.h>

#define MPIO_TEST_WRITE_SIZE 1024*1024     /* 1 MB */

void
test_mpio_overlap_writes(char *filename[])
{
    int mpi_size, mpi_rank;
    MPI_Comm comm;
    MPI_Info info = MPI_INFO_NULL;
    int color, mrc;
    MPI_File	fh;
    int newrank, newprocs;
    hid_t fid;			/* file IDs */
    hid_t acc_tpl;		/* File access properties */
    herr_t ret;			/* generic return value */
    int i;
    char  buf[4093];		/* use some prime number for size */
    int bufsize = sizeof(buf);
    int stride;
    MPI_Offset  mpi_off;
    MPI_Status  mpi_stat;


    if (verbose)
	printf("MPIO independent overlapping writes test on file %s\n",
	    filename[0]);

    /* set up MPI parameters */
    MPI_Comm_size(MPI_COMM_WORLD,&mpi_size);
    MPI_Comm_rank(MPI_COMM_WORLD,&mpi_rank);

    /* Need at least 2 processes */
    if (mpi_size < 2) {
	if (MAINPROCESS)
	    printf("Need at least 2 processes to run MPIO test.\n");
	    printf(" -SKIP- \n");
	return;
    }

    /* splits processes 0 to n-2 into one comm. and the last one into another */
    color = ((mpi_rank < (mpi_size - 1)) ? 0 : 1);
    mrc = MPI_Comm_split (MPI_COMM_WORLD, color, mpi_rank, &comm);
    VRFY((mrc==MPI_SUCCESS), "Comm_split succeeded");

    if (color==0){
	/* First n-1 processes (color==0) open a file and write it */
	mrc = MPI_File_open(comm, filename[0], MPI_MODE_CREATE|MPI_MODE_RDWR,
		info, &fh);
	VRFY((mrc==MPI_SUCCESS), "");

	stride = 1;
	mpi_off = mpi_rank*stride;
	while (mpi_off < MPIO_TEST_WRITE_SIZE){
	    /* make sure the write does not exceed the TEST_WRITE_SIZE */
	    if (mpi_off+stride > MPIO_TEST_WRITE_SIZE)
		stride = MPIO_TEST_WRITE_SIZE - mpi_off;

	    /* set data to some trivial pattern for easy verification */
	    for (i=0; i<stride; i++)
		buf[i] = (mpi_off+i) & 0x7f;
	    mrc = MPI_File_write_at(fh, mpi_off, buf, stride, MPI_BYTE,
		    &mpi_stat);
	    VRFY((mrc==MPI_SUCCESS), "");
	    
	    /* move the offset pointer to last byte written by all processes */
	    mpi_off += (mpi_size - 1 - mpi_rank) * stride;

	    /* Increase chunk size without exceeding buffer size. */
	    /* Then move the starting offset for next write. */
	    stride *= 2;
	    if (stride > bufsize)
		stride = bufsize;
	    mpi_off += mpi_rank*stride;
	}

	/* close file and free the communicator */
	mrc = MPI_File_close(&fh);
	VRFY((mrc==MPI_SUCCESS), "MPI_FILE_CLOSE");
	mrc = MPI_Comm_free(&comm);
	VRFY((mrc==MPI_SUCCESS), "MPI_Comm_free");
    
	/* sync with the other waiting processes */
	mrc = MPI_Barrier(MPI_COMM_WORLD);
	VRFY((mrc==MPI_SUCCESS), "Sync after writes");
    }else{
	/* last process waits till writes are done,
	 * then opens file to verify data.
	 */
	mrc = MPI_Barrier(MPI_COMM_WORLD);
	VRFY((mrc==MPI_SUCCESS), "Sync after writes");

	mrc = MPI_File_open(comm, filename[0], MPI_MODE_RDONLY,
		info, &fh);
	VRFY((mrc==MPI_SUCCESS), "");

	stride = bufsize;
	for (mpi_off=0; mpi_off < MPIO_TEST_WRITE_SIZE; mpi_off += bufsize){
	    /* make sure it does not read beyond end of data */
	    if (mpi_off+stride > MPIO_TEST_WRITE_SIZE)
		stride = MPIO_TEST_WRITE_SIZE - mpi_off;
	    mrc = MPI_File_read_at(fh, mpi_off, buf, stride, MPI_BYTE,
		    &mpi_stat);
	    VRFY((mrc==MPI_SUCCESS), "");
	    for (i=0; i<stride; i++){
		char expected;
		expected = (mpi_off+i) & 0x7f;
		if (buf[i] != expected)
		    printf("proc %d: found data error at [%ld], expect %d, got %d\n",
		    mpi_rank, mpi_off+i, expected, buf[i]);
	    }
	}

	/* close file and free the communicator */
	mrc = MPI_File_close(&fh);
	VRFY((mrc==MPI_SUCCESS), "MPI_FILE_CLOSE");
	mrc = MPI_Comm_free(&comm);
	VRFY((mrc==MPI_SUCCESS), "MPI_Comm_free");
    }

    /*
     * one more sync to ensure all processes have done reading
     * before ending this test.
     */
    mrc = MPI_Barrier(MPI_COMM_WORLD);
    VRFY((mrc==MPI_SUCCESS), "Sync before leaving test");
}