darshan-lustre.c 15.5 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
/*
 * Copyright (C) 2015 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
 */

#define _XOPEN_SOURCE 500
#define _GNU_SOURCE

#include "darshan-runtime-config.h"
#include <stdio.h>
#include <unistd.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <time.h>
#include <stdlib.h>
#include <assert.h>
#include <pthread.h>
19 20 21
#include <sys/ioctl.h>

/* XXX stick this into autoconf .h */
22
#include <lustre/lustre_user.h>
23 24 25 26 27

#include "uthash.h"

#include "darshan.h"
#include "darshan-dynamic.h"
28
#include "darshan-lustre.h"
29

30
struct lustre_runtime *lustre_runtime = NULL;
31 32 33 34
static pthread_mutex_t lustre_runtime_mutex = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP;
static int instrumentation_disabled = 0;
static int my_rank = -1;

35 36
static void lustre_runtime_initialize(void);

37 38 39 40
static void lustre_begin_shutdown(void);
static void lustre_get_output_data(MPI_Comm mod_comm, darshan_record_id *shared_recs,
    int shared_rec_count, void **lustre_buf, int *lustre_buf_sz);
static void lustre_shutdown(void);
41 42 43
static int lustre_record_compare(const void* a_p, const void* b_p);
static void lustre_record_reduction_op(void* infile_v, void* inoutfile_v,
    int *len, MPI_Datatype *datatype);
44 45 46 47

#define LUSTRE_LOCK() pthread_mutex_lock(&lustre_runtime_mutex)
#define LUSTRE_UNLOCK() pthread_mutex_unlock(&lustre_runtime_mutex)

48
void darshan_instrument_lustre_file(const char* filepath, int fd)
49
{
50
    struct lustre_record_runtime *rec_rt;
51
    struct darshan_lustre_record *rec;
52
    struct darshan_fs_info fs_info;
53
    darshan_record_id rec_id;
54
    int limit_flag;
55 56 57 58 59
    int i;
    struct lov_user_md *lum;
    size_t lumsize = sizeof(struct lov_user_md) +
        LOV_MAX_STRIPE_COUNT * sizeof(struct lov_user_ost_data);
    size_t rec_size;
60 61

    LUSTRE_LOCK();
62 63
    /* make sure the lustre module is already initialized */
    lustre_runtime_initialize();
64

65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87
    /* if we can't issue ioctl, we have no counter data at all */
    if ( (lum = calloc(1, lumsize)) == NULL )
        return;

    /* find out the OST count of this file so we can allocate memory */
    lum->lmm_magic = LOV_USER_MAGIC;
    lum->lmm_stripe_count = LOV_MAX_STRIPE_COUNT;

    /* -1 means ioctl failed, likely because file isn't on Lustre */
    if ( ioctl( fd, LL_IOC_LOV_GETSTRIPE, (void *)lum ) == -1 )
    {
        free(lum);
        return;
    }

    rec_size = LUSTRE_RECORD_SIZE( lum->lmm_stripe_count );

    {
        /* broken out for clarity */
        void *end_of_new_record = (char*)lustre_runtime->next_free_record + rec_size;
        void *end_of_rec_buffer = (char*)lustre_runtime->record_buffer + lustre_runtime->record_buffer_max;
        limit_flag = ( end_of_new_record > end_of_rec_buffer );
    }
88 89

    /* register a Lustre file record with Darshan */
90
    fs_info.fs_type = -1;
91 92 93 94 95
    darshan_core_register_record(
        (void *)filepath,
        strlen(filepath),
        DARSHAN_LUSTRE_MOD,
        1,
96
        limit_flag,
97
        &rec_id,
98
        &fs_info);
99 100 101

    /* if record id is 0, darshan has no more memory for instrumenting */
    if(rec_id == 0)
102
    {
103
        free(lum);
104
        LUSTRE_UNLOCK();
105
        return;
106 107 108
    }

    /* search the hash table for this file record, and initialize if not found */
109
    HASH_FIND(hlink, lustre_runtime->record_runtime_hash, &rec_id, sizeof(darshan_record_id), rec_rt );
110 111
    if ( !rec_rt ) {
        /* allocate a new lustre record and append it to the array */
112 113 114 115 116
        rec_rt = &(lustre_runtime->record_runtime_array[lustre_runtime->record_count]);
        rec_rt->record = lustre_runtime->next_free_record;
        rec_rt->record_size = rec_size;
        lustre_runtime->next_free_record = (char*)(lustre_runtime->next_free_record) + rec_size;
        lustre_runtime->record_buffer_used += rec_size;
117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133
        rec = rec_rt->record;
        rec->rec_id = rec_id;
        rec->rank = my_rank;

        /* implicit assumption here that none of these counters will change
         * after the first time a file is opened.  This may not always be
         * true in the future */
        if ( fs_info.fs_type != -1 ) 
        {
            rec->counters[LUSTRE_OSTS] = fs_info.ost_count;
            rec->counters[LUSTRE_MDTS] = fs_info.mdt_count;
        }
        else
        {
            rec->counters[LUSTRE_OSTS] = -1;
            rec->counters[LUSTRE_MDTS] = -1;
        }
134

135 136 137 138 139 140 141 142 143 144
        rec->counters[LUSTRE_STRIPE_SIZE] = lum->lmm_stripe_size;
        rec->counters[LUSTRE_STRIPE_WIDTH] = lum->lmm_stripe_count;
        rec->counters[LUSTRE_STRIPE_OFFSET] = lum->lmm_stripe_offset;
        for ( i = 0; i < lum->lmm_stripe_count; i++ )
            rec->ost_ids[i] = lum->lmm_objects[i].l_ost_idx;
        free(lum);

        HASH_ADD(hlink, lustre_runtime->record_runtime_hash, record->rec_id, sizeof(darshan_record_id), rec_rt);

        lustre_runtime->record_count++;
145
    }
146 147

    LUSTRE_UNLOCK();
148 149 150
    return;
}

151 152 153
static void lustre_runtime_initialize()
{
    int mem_limit;
154
    int max_records;
155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182
    struct darshan_module_funcs lustre_mod_fns =
    {
        .begin_shutdown = &lustre_begin_shutdown,
        .get_output_data = &lustre_get_output_data,
        .shutdown = &lustre_shutdown
    };

    /* don't do anything if already initialized or instrumenation is disabled */
    if(lustre_runtime || instrumentation_disabled)
        return;

    /* register the lustre module with darshan-core */
    darshan_core_register_module(
        DARSHAN_LUSTRE_MOD,
        &lustre_mod_fns,
        &my_rank,
        &mem_limit,
        NULL);

    /* return if no memory assigned by darshan core */
    if(mem_limit == 0)
        return;

    lustre_runtime = malloc(sizeof(*lustre_runtime));
    if(!lustre_runtime)
        return;
    memset(lustre_runtime, 0, sizeof(*lustre_runtime));

183 184 185
    /* allocate the full size of the memory limit we are given */
    lustre_runtime->record_buffer= malloc(mem_limit);
    if(!lustre_runtime->record_buffer)
186
    {
187
        lustre_runtime->record_buffer_max = 0;
188 189
        return;
    }
190 191 192 193 194 195 196 197 198 199 200 201 202
    lustre_runtime->record_buffer_max = mem_limit;
    lustre_runtime->next_free_record = lustre_runtime->record_buffer;
    memset(lustre_runtime->record_buffer, 0, lustre_runtime->record_buffer_max);

    /* Allocate array of Lustre runtime data.  We calculate the maximum possible
     * number of records that will fit into mem_limit by assuming that each
     * record has the minimum possible OST count, then allocate that many 
     * runtime records.  record_buffer will always run out of memory before
     * we overflow record_runtime_array.
     */
    max_records = mem_limit / sizeof(struct darshan_lustre_record);
    lustre_runtime->record_runtime_array =
        malloc( max_records * sizeof(struct lustre_record_runtime));
203 204
    if(!lustre_runtime->record_runtime_array)
    {
205 206
        lustre_runtime->record_buffer_max = 0;
        free( lustre_runtime->record_buffer );
207 208
        return;
    }
209 210
    memset(lustre_runtime->record_runtime_array, 0,
        max_records * sizeof(struct lustre_record_runtime));
211

212 213 214
    return;
}

215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237
/**************************************************************************
 * Functions exported by Lustre module for coordinating with darshan-core *
 **************************************************************************/

static void lustre_begin_shutdown(void)
{
    assert(lustre_runtime);

    LUSTRE_LOCK();
    /* disable further instrumentation while Darshan shuts down */
    instrumentation_disabled = 1;
    LUSTRE_UNLOCK();

    return;
}

static void lustre_get_output_data(
    MPI_Comm mod_comm,
    darshan_record_id *shared_recs,
    int shared_rec_count,
    void **lustre_buf,
    int *lustre_buf_sz)
{
238
    struct lustre_record_runtime *file;
239 240
    int i;

241 242
    assert(lustre_runtime);

243 244 245
    /* if there are globally shared files, do a shared file reduction */
    /* NOTE: the shared file reduction is also skipped if the 
     * DARSHAN_DISABLE_SHARED_REDUCTION environment variable is set.
246
     */
247
    if (shared_rec_count && !getenv("DARSHAN_DISABLE_SHARED_REDUCTION"))
248 249 250 251
    {
        /* necessary initialization of shared records */
        for(i = 0; i < shared_rec_count; i++)
        {
252
            HASH_FIND(hlink, lustre_runtime->record_runtime_hash, &shared_recs[i],
253 254 255
                sizeof(darshan_record_id), file);
            assert(file);

256
            file->record->rank = -1;
257 258 259 260 261 262
        }

        /* sort the array of files descending by rank so that we get all of the 
         * shared files (marked by rank -1) in a contiguous portion at end 
         * of the array
         */
263
        sort_lustre_records();
264

265 266 267
        /* simply drop all shared records from non-root ranks by truncating
         * the record array and recalculating the size of the used buffer
         */
268
        if (my_rank != 0)
269
        {
270
            lustre_runtime->record_count -= shared_rec_count;
271 272 273 274 275
            lustre_runtime->record_buffer_used = 0;
            for ( i = 0; i < lustre_runtime->record_count; i++ )
                lustre_runtime->record_buffer_used += 
                    LUSTRE_RECORD_SIZE( (lustre_runtime->record_runtime_array[i]).record->counters[LUSTRE_STRIPE_WIDTH] );
        }
276
    }
277

278 279
    *lustre_buf = (void *)(lustre_runtime->record_buffer);
    *lustre_buf_sz = lustre_runtime->record_buffer_used;
280

281 282 283 284 285 286 287
    return;
}

static void lustre_shutdown(void)
{
    assert(lustre_runtime);

288
    HASH_CLEAR(hlink, lustre_runtime->record_runtime_hash);
289
    free(lustre_runtime->record_runtime_array);
290
    free(lustre_runtime->record_buffer);
291
    free(lustre_runtime);
292 293 294 295 296
    lustre_runtime = NULL;

    return;
}

297 298 299
/* compare function for sorting file records by descending rank */
static int lustre_record_compare(const void* a_p, const void* b_p)
{
300 301
    const struct lustre_record_runtime* a = a_p;
    const struct lustre_record_runtime* b = b_p;
302

303
    if (a->record->rank < b->record->rank)
304
        return 1;
305
    if (a->record->rank > b->record->rank)
306 307
        return -1;

308 309 310 311 312 313 314
    /* if ( a->record->rank == b->record->rank ) we MUST do a secondary
     * sort so that the order of qsort is fully deterministic and consistent
     * across all MPI ranks.  Without a secondary sort, the sort order can
     * be affected by rank-specific variations (e.g., the order in which
     * files are first opened).
     */

315 316 317
    return 0;
}

318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380
/*
 * Sort the record_runtimes and records by MPI rank to facilitate shared redux.
 * This requires craftiness and additional heap utilization because the records
 * (but not record_runtimes) have variable size.  Currently has to temporarily
 * duplicate the entire record_buffer; there is room for more memory-efficient
 * optimization if this becomes a scalability issue.
 */
int sort_lustre_records()
{
    int i;
    struct darshan_lustre_record *rec;
    struct lustre_record_runtime *rec_rt, *tmp_rec_rt;
    char  *new_buf, *p;

    /* Create a new buffer to store an entire replica of record_buffer.  Since
     * we know the exact size of record_buffer's useful data at this point, we
     * can allocate the exact amount we need instead of record_buffer_max */
    new_buf = malloc(lustre_runtime->record_buffer_used);
    p = new_buf;
    if ( !new_buf )
        return 1;

    /* qsort breaks the hash table, so delete it now to free its memory buffers
     * and prevent later confusion */
    HASH_ITER( hlink, lustre_runtime->record_runtime_hash, rec_rt, tmp_rec_rt )
        HASH_DELETE( hlink, lustre_runtime->record_runtime_hash, rec_rt );

    /* sort the runtime records, which is has fixed-length elements */
    qsort(
        lustre_runtime->record_runtime_array,
        lustre_runtime->record_count,
        sizeof(struct lustre_record_runtime),
        lustre_record_compare
    );

    /* rebuild the hash and array with the qsorted runtime records */
    for ( i = 0; i < lustre_runtime->record_count; i++ )
    {
        rec_rt = &(lustre_runtime->record_runtime_array[i]);
        HASH_ADD(hlink, lustre_runtime->record_runtime_hash, record->rec_id, sizeof(darshan_record_id), rec_rt );
    }

    /* create reordered record buffer, then copy it back in place */
    for ( i = 0; i < lustre_runtime->record_count; i++ )
    {
        rec_rt = &(lustre_runtime->record_runtime_array[i]);
        memcpy( p, rec_rt->record, rec_rt->record_size );
        /* fix record pointers within each runtime record too - pre-emptively
         * point them at where they will live in record_buffer after we memcpy
         * below */
        rec_rt->record = (struct darshan_lustre_record *)((char*)(lustre_runtime->record_buffer) + (p - new_buf));

        p += rec_rt->record_size;
    }
    memcpy( 
        lustre_runtime->record_buffer, 
        new_buf, 
        lustre_runtime->record_buffer_used );

    free(new_buf);
    return 0;
}

381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419
/* this is just boilerplate reduction code that isn't currently used */
static void lustre_record_reduction_op(void* infile_v, void* inoutfile_v,
    int *len, MPI_Datatype *datatype)
{
    struct darshan_lustre_record tmp_record;
    struct darshan_lustre_record *infile = infile_v;
    struct darshan_lustre_record *inoutfile = inoutfile_v;
    int i, j;

    assert(lustre_runtime);

    for( i=0; i<*len; i++ )
    {
        memset(&tmp_record, 0, sizeof(struct darshan_lustre_record));
        tmp_record.rec_id = infile->rec_id;
        tmp_record.rank = -1;

        /* preserve only rank 0's value */
        for( j = LUSTRE_OSTS; j < LUSTRE_NUM_INDICES; j++)
        {
            if ( my_rank == 0 ) 
            {
                tmp_record.counters[j] = infile->counters[j];
            }
            else
            {
                tmp_record.counters[j] = inoutfile->counters[j];
            }
        }

        /* update pointers */
        *inoutfile = tmp_record;
        inoutfile++;
        infile++;
    }

    return;
}

420 421 422 423 424 425 426 427 428 429 430 431 432 433 434
/*
 *  Dump the memory structure of our records and runtime records
 */
void print_lustre_runtime( void )
{
    int i, j;
    struct darshan_lustre_record *rec;

    /* print what we just loaded */
    for ( i = 0; i < lustre_runtime->record_count; i++ )
    {
        rec = (lustre_runtime->record_runtime_array[i]).record;
        printf( "File %2d\n", i );
        for ( j = 0; j < LUSTRE_NUM_INDICES; j++ )
        {
435
            printf( "  Counter %-2d: %10ld, addr %ld\n", 
436 437 438 439 440 441
                j, 
                rec->counters[j],
                (char*)(&(rec->counters[j])) - (char*)(lustre_runtime->record_buffer) );
        }
        for ( j = 0; j < rec->counters[LUSTRE_STRIPE_WIDTH]; j++ )
        {
442 443
            if ( j > 0 && j % 2 == 0 ) printf("\n");
            printf( "  Stripe  %-2d: %10ld, addr %-9d", 
444 445 446 447
                j, 
                rec->ost_ids[j],
                (char*)(&(rec->ost_ids[j])) - (char*)(lustre_runtime->record_buffer) );
        }
448
        printf( "\n" );
449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486
    }
    return;
}

/*
 *  Dump the order in which records appear in memory
 */
void print_array( void )
{
    int i;
    struct lustre_record_runtime *rec_rt;
    printf("*** DUMPING RECORD LIST BY ARRAY SEQUENCE\n");
    for ( i = 0; i < lustre_runtime->record_count; i++ )
    {
        rec_rt = &(lustre_runtime->record_runtime_array[i]);
        printf( "*** record %d rank %d osts %d\n", 
            rec_rt->record->rec_id, 
            rec_rt->record->rank,
            rec_rt->record->counters[LUSTRE_STRIPE_WIDTH]);
    }
}
void print_hash( void )
{
    struct lustre_record_runtime *rec_rt, *tmp_rec_rt;
    printf("*** DUMPING RECORD LIST BY HASH SEQUENCE\n");
    HASH_ITER( hlink, lustre_runtime->record_runtime_hash, rec_rt, tmp_rec_rt )
    {
        printf( "*** record %d rank %d osts %d\n", 
            rec_rt->record->rec_id, 
            rec_rt->record->rank,
            rec_rt->record->counters[LUSTRE_STRIPE_WIDTH]);
    }
    return;
}




487 488 489 490 491 492 493 494
/*
 * Local variables:
 *  c-indent-level: 4
 *  c-basic-offset: 4
 * End:
 *
 * vim: ts=8 sts=4 sw=4 expandtab
 */