lustre/tests/mpi/write_disjoint.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/tests/write_disjoint.c
  37  *
  38  * Each loop does 3 things:
  39  *   - rank 0 truncates to 0
  40  *   - all ranks agree on a random chunk size
  41  *   - all ranks race to write their pattern to their chunk of the file
  42  *   - rank 0 makes sure that the resulting file size is ranks * chunk size
  43  *   - rank 0 makes sure that everyone's patterns went to the right place
  44  *
  45  * compile: mpicc -g -Wall -o write_disjoint write_disjoint.c
  46  * run:     mpirun -np N -machlist <hostlist file> write_disjoint
  47  *  or:     pdsh -w <N hosts> write_disjoint
  48  *  or:     prun -n N [-N M] write_disjoint
  49  */
  50
  51 #include <stdlib.h>
  52 #include <stdio.h>
  53 #include <sys/types.h>
  54 #include <sys/stat.h>
  55 #include <fcntl.h>
  56 #include <string.h>
  57 #include <errno.h>
  58 #include <unistd.h>
  59 #include <stdarg.h>
  60 #include "mpi.h"
  61
  62 #define CHUNK_MAX_SIZE 123456
  63
  64 void rprintf(int rank, int loop, const char *fmt, ...)
  65 {
  66         va_list       ap;
  67
  68         printf("rank %d, loop %d: ", rank, loop);
  69
  70         va_start(ap, fmt);
  71
  72         vprintf(fmt, ap);
  73
  74         MPI_Abort(MPI_COMM_WORLD, -1); /* This will exit() according to man */
  75 }
  76
  77 #define CHUNK_SIZE(n) chunk_size[(n) % 2]
  78
  79 int main (int argc, char *argv[]) {
  80         int i, n, fd, c;
  81         unsigned long chunk_size[2];
  82         int rank, noProcessors, done;
  83         int error;
  84         off_t offset;
  85         char **chunk_buf;
  86         char *read_buf;
  87         struct stat stat_buf;
  88         ssize_t ret;
  89         char *filename = "/mnt/lustre/write_disjoint";
  90         int numloops = 1000;
  91         int random = 0;
  92
  93         error = MPI_Init(&argc, &argv);
  94         if (error != MPI_SUCCESS)
  95                 rprintf(-1, -1, "MPI_Init failed: %d\n", error);
  96         /* Parse command line options */
  97         while ((c = getopt(argc, argv, "f:n:")) != EOF) {
  98                 switch (c) {
  99                 case 'f':
 100                         filename = optarg;
 101                         break;
 102                 case 'n':
 103                         numloops = strtoul(optarg, NULL, 0);
 104                         break;
 105                 }
 106         }
 107
 108         MPI_Comm_size(MPI_COMM_WORLD, &noProcessors);
 109         MPI_Comm_rank(MPI_COMM_WORLD, &rank);
 110
 111         chunk_buf = malloc(noProcessors * sizeof(chunk_buf[0]));
 112         for (i=0; i < noProcessors; i++) {
 113                 chunk_buf[i] = malloc(CHUNK_MAX_SIZE);
 114                 memset(chunk_buf[i], 'A'+ i, CHUNK_MAX_SIZE);
 115         }
 116         read_buf = malloc(noProcessors * CHUNK_MAX_SIZE);
 117
 118         if (rank == 0) {
 119                 fd = open(filename, O_WRONLY|O_CREAT|O_TRUNC, 0666);
 120                 if (fd < 0)
 121                         rprintf(rank, -1, "open() returned %s\n",
 122                                 strerror(errno));
 123         }
 124         MPI_Barrier(MPI_COMM_WORLD);
 125
 126         fd = open(filename, O_RDWR);
 127         if (fd < 0)
 128                 rprintf(rank, -1, "open() returned %s\n", strerror(errno));
 129
 130         for (n = 0; n < numloops; n++) {
 131                 /* reset the environment */
 132                 if (rank == 0) {
 133                         ret = truncate(filename, 0);
 134                         if (ret != 0)
 135                                 rprintf(rank, n, "truncate() returned %s\n",
 136                                         strerror(errno) );
 137                         random = rand();
 138                 }
 139                 MPI_Bcast(&random, 1, MPI_INT, 0, MPI_COMM_WORLD);
 140                 CHUNK_SIZE(n) = random % CHUNK_MAX_SIZE;
 141
 142                 if (n % 1000 == 0 && rank == 0)
 143                         printf("loop %d: chunk_size %lu\n", n, CHUNK_SIZE(n));
 144
 145                 if (stat(filename, &stat_buf) < 0)
 146                         rprintf(rank, n, "error stating %s: %s\n",
 147                                 filename, strerror(errno));
 148
 149                 if (stat_buf.st_size != 0)
 150                         rprintf(rank, n, "filesize = %lu. "
 151                                 "Should be zero after truncate\n",
 152                                 stat_buf.st_size);
 153
 154                 MPI_Barrier(MPI_COMM_WORLD);
 155
 156                 /* Do the race */
 157                 offset = rank * CHUNK_SIZE(n);
 158                 lseek(fd, offset, SEEK_SET);
 159
 160                 done = 0;
 161                 do {
 162                         ret = write(fd, chunk_buf[rank] + done,
 163                                     CHUNK_SIZE(n) - done);
 164                         if (ret < 0 && errno != EINTR)
 165                                 rprintf(rank, n, "write() returned %s\n",
 166                                         strerror(errno));
 167                         if (ret > 0)
 168                                 done += ret;
 169                 } while (done != CHUNK_SIZE(n));
 170
 171                 MPI_Barrier(MPI_COMM_WORLD);
 172
 173                 /* Check the result */
 174                 if (stat(filename, &stat_buf) < 0)
 175                         rprintf(rank, n, "error stating %s: %s\n",
 176                                 filename, strerror(errno));
 177
 178                 if (stat_buf.st_size != CHUNK_SIZE(n) * noProcessors) {
 179                         if (n > 0)
 180                                 printf("loop %d: chunk_size %lu, "
 181                                        "file size was %lu\n",
 182                                        n - 1, CHUNK_SIZE(n - 1),
 183                                        CHUNK_SIZE(n - 1) *noProcessors);
 184                         rprintf(rank, n, "invalid file size %lu"
 185                                 " instead of %lu = %lu * %u\n",
 186                                 (unsigned long)stat_buf.st_size,
 187                                 CHUNK_SIZE(n) * noProcessors,
 188                                 CHUNK_SIZE(n), noProcessors);
 189                 }
 190
 191                 if (rank == 0) {
 192                         if (lseek(fd, 0, SEEK_SET) < 0)
 193                                 rprintf(rank, n, "error seeking to 0: %s\n",
 194                                         strerror(errno));
 195
 196                         done = 0;
 197                         do {
 198                                 ret = read(fd, read_buf + done,
 199                                            CHUNK_SIZE(n) * noProcessors - done);
 200                                 if (ret < 0)
 201                                         rprintf(rank, n, "read returned %s\n",
 202                                                 strerror(errno));
 203
 204                                 done += ret;
 205                         } while (done != CHUNK_SIZE(n) * noProcessors);
 206
 207                         for (i = 0; i < noProcessors; i++) {
 208                                 char command[4096];
 209                                 int j;
 210
 211                                 if (!memcmp(read_buf + (i * CHUNK_SIZE(n)),
 212                                             chunk_buf[i], CHUNK_SIZE(n)))
 213                                         continue;
 214
 215                                 /* print out previous chunk sizes */
 216                                 if (n > 0)
 217                                         printf("loop %d: chunk_size %lu\n",
 218                                                n - 1, CHUNK_SIZE(n - 1));
 219
 220                                 printf("loop %d: chunk %d corrupted "
 221                                        "with chunk_size %lu, page_size %d\n",
 222                                        n, i, CHUNK_SIZE(n), getpagesize());
 223                                 printf("ranks:\tpage boundry\tchunk boundry\t"
 224                                        "page boundry\n");
 225                                 for (j = 1 ; j < noProcessors; j++) {
 226                                         int b = j * CHUNK_SIZE(n);
 227                                         printf("%c -> %c:\t%d\t%d\t%d\n",
 228                                                'A' + j - 1, 'A' + j,
 229                                                b & ~(getpagesize()-1), b,
 230                                                (b + getpagesize()) &
 231                                                ~(getpagesize()-1));
 232                                 }
 233
 234                                 sprintf(command, "od -Ad -a %s", filename);
 235                                 ret = system(command);
 236                                 rprintf(0, n, "data check error - exiting\n");
 237                         }
 238                 }
 239                 MPI_Barrier(MPI_COMM_WORLD);
 240         }
 241
 242         printf("Finished after %d loops\n", n);
 243         MPI_Finalize();
 244         return 0;
 245 }