lp-io.c 10.3 KB
Newer Older
1
/*
Philip Carns's avatar
Philip Carns committed
2
 * Copyright (C) 2013 University of Chicago.
3
 * See COPYRIGHT notice in top-level directory.
Philip Carns's avatar
Philip Carns committed
4
 *
5 6 7 8 9 10
 */

#include <assert.h>
#include <sys/stat.h>
#include <sys/types.h>

Jonathan Jenkins's avatar
Jonathan Jenkins committed
11 12
#include <codes/lp-io.h>
#include <codes/codes.h>
13 14 15 16 17 18 19 20 21 22 23

struct io_buffer
{
    tw_lpid gid;
    int size;
    void* buffer;
    struct io_buffer *next;
};

struct identifier
{
24
    char identifier[64];
25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45
    struct io_buffer *buffers;
    int buffers_count;
    int buffers_total_size;
    struct identifier *next;
};

/* local list of identifiers */
static struct identifier* identifiers = NULL;
int identifiers_count = 0;

/* array of all identifiers used across all procs */
struct identifier global_identifiers[64];
int global_identifiers_count = 0;

static int write_id(char* directory, char* identifier, MPI_Comm comm);

int lp_io_write(tw_lpid gid, char* identifier, int size, void* buffer)
{
    struct identifier* id;
    struct io_buffer *buf;

46
    if(strlen(identifier) >= 64)
47 48 49 50 51
    {
        fprintf(stderr, "Error: identifier %s too big.\n", identifier);
        return(-1);
    }

52
    buf = (struct io_buffer*)malloc(sizeof(*buf));
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
    if(!buf)
        return(-1);

    buf->gid = gid;
    buf->size = size;
    /* allocate copy of data being written */
    buf->buffer = malloc(size);
    if(!buf->buffer)
    {
        free(buf);
        return(-1);
    }
    memcpy(buf->buffer, buffer, size);
    buf->next = NULL;

    /* see if we have this identifier already */
    id = identifiers;
    while(id && (strcmp(identifier, id->identifier) != 0))
        id = id->next;

    if(id)
    {
        /* add buffer to front of list */
        buf->next = id->buffers;
        id->buffers  = buf;
        id->buffers_count++;
        id->buffers_total_size += size;
    }
    else
    {
        /* new identifier */
84
        id = (struct identifier*)malloc(sizeof(*id));
85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103
        if(!id)
        {
            free(buf->buffer);
            free(buf);
            return(-1);
        }
        strcpy(id->identifier, identifier);
        id->next = identifiers;
        identifiers = id;
        id->buffers = buf;
        id->buffers_count = 1;
        id->buffers_total_size = size;
        buf->next = NULL;
        identifiers_count++;
    }

    return(0);
}

104
int lp_io_write_rev(tw_lpid gid, char* identifier){
105
    struct identifier* id, *id_prev;
106 107 108 109 110 111 112 113 114 115
    struct io_buffer *buf, *buf_prev;

    /* find given identifier */
    if(strlen(identifier) >= 32)
    {
        fprintf(stderr, "Error: identifier %s too big.\n", identifier);
        return(-1);
    }
    id = identifiers;
    while (id && (strcmp(identifier, id->identifier) != 0)){
116
        id_prev = id;
117 118 119
        id = id->next;
    }
    if (!id){
Jonathan Jenkins's avatar
Jonathan Jenkins committed
120 121
        fprintf(stderr, "Error: identifier %s not found on reverse for LP %llu.",
                identifier,LLU(gid));
122 123 124 125 126 127 128 129 130 131 132 133 134
        return(-1);
    }

    /* attempt to find previous LP's write. This is made easier by the fact 
     * that the list is stored newest-to-oldest */
    buf = id->buffers; 
    buf_prev = NULL;
    while (buf){
        if (buf->gid == gid){ break; }
        buf_prev = buf;
        buf = buf->next;
    }
    if (!buf){
Jonathan Jenkins's avatar
Jonathan Jenkins committed
135
        fprintf(stderr, "Error: no lp-io write buffer found for LP %llu (reverse write)\n", LLU(gid));
136 137 138 139
        return(-1);
    }

    id->buffers_count--;
140
    id->buffers_total_size -= buf->size;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
141 142 143
    if (id->buffers_count == 0){
        /* seg faults caused with empty identifiers for some reason - remove
         * this ID */
144 145 146 147 148 149
        if (id == identifiers){
            identifiers = id->next;
        }
        else{
            id_prev->next = id->next;
        }
Jonathan Jenkins's avatar
Jonathan Jenkins committed
150 151 152
        free(id);
        identifiers_count--;
    }
153 154
    /* remove the buffer from the list 
     * (NULLs for end-of-list are preserved) */
155
    else if (buf == id->buffers) { /* buf is head of list */
156 157 158 159 160 161 162 163 164
        id->buffers = buf->next;
    }
    else { /* buf is in list, has a previous element */
        buf_prev->next = buf->next; 
    }
    free(buf->buffer);
    free(buf);
    return(0);
}
165 166 167 168 169 170 171 172 173 174 175

int lp_io_prepare(char *directory, int flags, lp_io_handle* handle, MPI_Comm comm)
{
    int ret;
    int rank;
    int save_err;

    MPI_Comm_rank(comm, &rank);

    if(flags & LP_IO_UNIQ_SUFFIX)
    {
176
        *handle = (lp_io_handle)malloc(256);
177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355
        if(!(*handle))
        {
            return(-1);
        }
        if(rank == 0)
            sprintf(*handle, "%s-%ld-%ld", directory, (long)getpid(), (long)time(NULL));
        MPI_Bcast(*handle, 256, MPI_CHAR, 0, comm);
    }
    else
    {
        *handle = strdup(directory);
        if(!(*handle))
        {
            return(-1);
        }
    }

    if(rank == 0)
    {
        /* create output directory */
        ret = mkdir(*handle, 0775);
        if(ret != 0)
        {
            save_err = errno;

            fprintf(stderr, "mkdir(\"%s\") failed: %s\n", *handle, strerror(save_err));
            return(-1);
        }
    }

    MPI_Barrier(comm);

    return(0);
}

int lp_io_flush(lp_io_handle handle, MPI_Comm comm)
{
    int comm_size;
    int rank;
    MPI_Status status;
    int ret;
    int i;
    struct identifier *id;
    int match = 0;

    char* directory  = handle;

    memset(global_identifiers, 0, 64*sizeof(global_identifiers[0]));

    MPI_Comm_size(comm, &comm_size);
    MPI_Comm_rank(comm, &rank);

    /* The first thing we need to do is come up with a global list of
     * identifiers.  We can't really guarantee that every MPI proc had data
     * written to every identifier, but we want to collectively write each
     * ID.
     */

    /* To do this, each rank will receive an initial set of ideas from
     * rank+1 and add its own identifiers to the set before transmitting to
     * rank-1.  Rank 0 will then broadcast the results to everyone.
     */

    /* get array so far from rank +1 */
    if(rank != comm_size-1)
    {
        ret = MPI_Recv(&global_identifiers_count, 1, MPI_INT, rank+1, 0, comm, &status);
        assert(ret == 0);
        ret = MPI_Recv(global_identifiers, 64*sizeof(global_identifiers[0]), MPI_BYTE,
            rank+1, 0, comm, &status);
        assert(ret == 0);
    }

    /* add our own entries to the array */
    id = identifiers;
    while(id)
    {
        match = 0;
        for(i = 0; i<global_identifiers_count; i++)
        {
            if(strcmp(id->identifier, global_identifiers[i].identifier) == 0)
            {
                match = 1;
                break;
            }
        }
        if(!match)
        {
            assert(global_identifiers_count < 64);
            global_identifiers[global_identifiers_count] = *id;
            global_identifiers_count++;
        }

        id = id->next;
    }

    /* send results to rank-1 */
    if(rank != 0)
    {
        ret = MPI_Send(&global_identifiers_count, 1, MPI_INT, rank-1, 0, comm);
        assert(ret == 0);
        ret = MPI_Send(global_identifiers, 64*sizeof(global_identifiers[0]), MPI_BYTE,
            rank-1, 0, comm);
        assert(ret == 0);

    }

    /* broadcast results to everyone */
    ret = MPI_Bcast(&global_identifiers_count, 1, MPI_INT, 0, comm);
    assert(ret == 0);
    ret = MPI_Bcast(global_identifiers, 64*sizeof(global_identifiers[0]), MPI_BYTE, 0, comm);
    assert(ret == 0);

    if(rank == 0)
    {
        printf("LP-IO: writing output to %s/\n", directory);
        printf("LP-IO: data files:\n");
    }

    for(i=0; i<global_identifiers_count; i++)
    {
        if(rank == 0)
        {
            printf("   %s/%s\n", directory, global_identifiers[i].identifier);
        }

        ret = write_id(directory, global_identifiers[i].identifier, comm);
        if(ret < 0)
        {
            return(ret);
        }
    }

    free(handle);

    return(0);
}

static int write_id(char* directory, char* identifier, MPI_Comm comm)
{
    char file[256];
    MPI_File fh;
    int ret;
    long my_size = 0;
    long my_offset = 0;
    struct identifier* id;
    struct io_buffer *buf;
    char err_string[MPI_MAX_ERROR_STRING];
    int err_len;
    MPI_Datatype mtype;
    int *lengths;
    MPI_Aint *displacements;
    MPI_Aint base;
    int i;
    MPI_Status status;

    sprintf(file, "%s/%s", directory, identifier);
    ret = MPI_File_open(comm, file, MPI_MODE_CREATE|MPI_MODE_WRONLY|MPI_MODE_EXCL, MPI_INFO_NULL, &fh);
    if(ret != 0)
    {
        MPI_Error_string(ret, err_string, &err_len);
        fprintf(stderr, "Error: MPI_File_open(%s) failure: %s\n", file, err_string);
        return(-1);
    }

    /* see if we have any data for this id */
    id = identifiers;
    while(id && (strcmp(identifier, id->identifier) != 0))
        id = id->next;

    /* find my offset */
    if(id)
        my_size = id->buffers_total_size;
    MPI_Scan(&my_size, &my_offset, 1, MPI_LONG, MPI_SUM, comm);
    my_offset -= my_size;

    /* build datatype for our buffers */
    if(id)
    {
356
        lengths = (int*)malloc(id->buffers_count*sizeof(int));
357
        assert(lengths);
358
        displacements = (MPI_Aint*)malloc(id->buffers_count*sizeof(MPI_Aint));
359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377
        assert(displacements);

        /* NOTE: some versions of MPI-IO have a bug related to using
         * MPI_BOTTOM with hindexed types. We therefore use first pointer as
         * base and adjust the others accordingly.
         */
        buf = id->buffers;
        base = (MPI_Aint)buf->buffer;
        /* NOTE: filling in datatype arrays backwards to get buffers in the
         * order they were added on each process
         */
        i = id->buffers_count-1;
        while(buf)
        {
            displacements[i] = (MPI_Aint)buf->buffer - (MPI_Aint)base;
            lengths[i] = buf->size;
            i--;
            buf = buf->next;
        }
378
        MPI_Type_create_hindexed(id->buffers_count, lengths, displacements,
379 380
            MPI_BYTE, &mtype);
        MPI_Type_commit(&mtype);
381 382
        free(lengths);
        free(displacements);
383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416

        ret = MPI_File_write_at_all(fh, my_offset, id->buffers->buffer, 1, mtype, &status);

        MPI_Type_free(&mtype);
    }
    else
    {
        /* nothing to write, but participate in collective anyway */
        ret = MPI_File_write_at_all(fh, my_offset, NULL, 0, MPI_BYTE, &status);
    }
    if(ret != 0)
    {
        fprintf(stderr, "Error: MPI_File_write_at(%s) failure.\n", file);
        return(-1);
    }
    
    ret = MPI_File_close(&fh);
    if(ret != 0)
    {
        fprintf(stderr, "Error: MPI_File_close(%s) failure.\n", file);
        return(-1);
    }

    return(0);
}

/*
 * Local variables:
 *  c-indent-level: 4
 *  c-basic-offset: 4
 * End:
 *
 * vim: ft=c ts=8 sts=4 sw=4 expandtab
 */