codes-workload-mpi-replay.c 18.1 KB
Newer Older
1 2 3 4 5 6 7 8
/*
 * Copyright (C) 2013 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
 */

/* SUMMARY:
 *
9
 *  MPI replay tool for replaying workloads from the codes workload API.
10 11 12 13 14 15 16 17 18 19
 *
 */

#include <stdio.h>
#include <unistd.h>
#include <assert.h>
#include <inttypes.h>
#include <math.h>
#include <getopt.h>
#include <fcntl.h>
20
#include <sys/stat.h>
21 22 23 24
#include <mpi.h>

#include "codes/codes-workload.h"
#include "codes/quickhash.h"
25
#include "codes/configuration.h"
26

27 28 29
#define DEBUG_PROFILING 1

#define WORKLOAD_DELAY_PCT 1
30

31
/* hash table entry for looking up file descriptor of a workload file id */
32
struct file_info
33 34 35 36 37 38 39 40 41 42 43 44
{
    struct qlist_head hash_link;
    uint64_t file_hash;
    int file_descriptor;
};

int replay_workload_op(struct codes_workload_op replay_op, int rank, long long int op_number);
int hash_file_compare(void *key, struct qlist_head *link);

/* command line options */
static int opt_verbose = 0;
static int opt_noop = 0;
45
static int opt_lockless = 0;
46 47 48 49

/* hash table for storing file descriptors of opened files */
static struct qhash_table *fd_table = NULL;

50 51 52
/* file stream to log rank events to, if verbose turned on */
static FILE *log_stream = NULL;

53 54 55 56 57 58 59 60
/* global variables for profiling different portions of the replay, if enabled */
static double total_open_time = 0.0;
static double total_close_time = 0.0;
static double total_read_time = 0.0;
static double total_write_time = 0.0;
static double total_delay_time = 0.0;
static double total_barrier_time = 0.0;

61 62
void usage(char *exename)
{
63 64
    fprintf(stderr, "Usage: %s [OPTIONS] --conf <conf_file_path>\n       "
            "--test-dir <workload_test_dir>\n\n", exename);
65
    fprintf(stderr, "\t<conf_file_path> : (absolute) path to a valid workload configuration file\n");
66 67
    fprintf(stderr, "\t<workload_test_dir> : the directory to replay the workload I/O in\n");
    fprintf(stderr, "\n\t[OPTIONS] includes:\n");
68
    fprintf(stderr, "\t\t--noop : do not perform i/o\n");
69
    fprintf(stderr, "\t\t--lockless : use bgq lockless file i/o optimization\n");
70
    fprintf(stderr, "\t\t    -v : verbose (output i/o details)\n");
71 72 73 74

    exit(1);
}

75
void parse_args(int argc, char **argv, char **conf_path, char **test_dir)
76 77 78 79
{
    int index;
    static struct option long_opts[] =
    {
80
        {"conf", 1, NULL, 'c'},
81
        {"test-dir", 1, NULL, 'd'},
82
        {"noop", 0, NULL, 'n'},
83
        {"lockless", 0, NULL, 'l'},
84 85 86 87
        {"help", 0, NULL, 0},
        {0, 0, 0, 0}
    };

88
    *conf_path = NULL;
89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104
    *test_dir = NULL;
    while (1)
    {
        int c = getopt_long(argc, argv, "v", long_opts, &index);

        if (c == -1)
            break;

        switch (c)
        {
            case 'v':
                opt_verbose = 1;
                break;
            case 'n':
                opt_noop = 1;
                break;
105 106 107
            case 'l':
                opt_lockless = 1;
                break;
108 109
            case 'c':
                *conf_path = optarg;
110 111 112 113 114 115 116 117 118 119 120 121
                break;
            case 'd':
                *test_dir = optarg;
                break;
            case 0:
            case '?':
            default:
                usage(argv[0]);
                break;
        }
    }

122
    if (optind < argc || !(*conf_path) || !(*test_dir))
123 124 125 126 127 128 129
    {
        usage(argv[0]);
    }

    return;
}

130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166
int load_workload(char *conf_path, int rank)
{
    config_lpgroups_t paramconf;
    char workload_type[MAX_NAME_LENGTH_WKLD];

    /* load the config file across all ranks */
    configuration_load(conf_path, MPI_COMM_WORLD, &config);

    /* get the PARAMS group out of the config file */
    configuration_get_lpgroups(&config, "PARAMS", &paramconf);

    /* get the workload type out of PARAMS */
    configuration_get_value(&config, "PARAMS", "workload_type",
                            workload_type, MAX_NAME_LENGTH_WKLD);

    /* set up the workload parameters and load into the workload API */
    if (strcmp(workload_type, "darshan_io_workload") == 0)
    {
        struct darshan_params d_params;
        char aggregator_count[10];

        /* get the darshan params from the config file */
        configuration_get_value(&config, "PARAMS", "log_file_path",
                                d_params.log_file_path, MAX_NAME_LENGTH_WKLD);
        configuration_get_value(&config, "PARAMS", "aggregator_count", aggregator_count, 10);
        d_params.aggregator_cnt = atoi(aggregator_count);

        return codes_workload_load(workload_type, (char *)&d_params, rank);
    }
    else if (strcmp(workload_type, "bgp_io_workload") == 0)
    {
        struct bgp_params b_params;
        char rank_count[10];

        /* get the bgp i/o params from the config file */
        configuration_get_value(&config, "PARAMS", "io_kernel_meta_path",
                                b_params.io_kernel_meta_path, MAX_NAME_LENGTH_WKLD);
167
        configuration_get_value(&config, "PARAMS", "bgp_config_file",
168 169 170 171 172 173 174 175
                                b_params.bgp_config_file, MAX_NAME_LENGTH_WKLD);
        configuration_get_value(&config, "PARAMS", "rank_count", rank_count, 10);
        strcpy(b_params.io_kernel_path, "");
        strcpy(b_params.io_kernel_def_path, "");
        b_params.num_cns = atoi(rank_count);

        return codes_workload_load(workload_type, (char *)&b_params, rank);
    }
176 177 178
    else if (strcmp(workload_type, "recorder_io_workload") == 0) {
        struct recorder_params r_params;

179
        /* get the recorder params from the config file */
180 181 182 183 184 185
        configuration_get_value(&config, "PARAMS", "trace_dir_path",
                                r_params.trace_dir_path, MAX_NAME_LENGTH_WKLD);

        return codes_workload_load(workload_type, (char *)&r_params, rank);

	}
186 187 188 189 190 191 192
    else
    {
        fprintf(stderr, "Error: Invalid workload type specified (%s)\n", workload_type);
        return -1;
    }
}

193 194
int main(int argc, char *argv[])
{
195
    char *conf_path;
196
    char *replay_test_path;
197 198
    char *log_dir = "log";
    char my_log_path[MAX_NAME_LENGTH_WKLD];
199 200 201 202 203
    int nprocs;
    int myrank;
    int workload_id;
    struct codes_workload_op next_op;
    long long int replay_op_number = 1;
204 205
    double load_start, load_end;
    int first_op = 1;
206 207 208
    int ret = 0;

    /* parse command line args */
209
    parse_args(argc, argv, &conf_path, &replay_test_path);
210 211 212 213 214 215

    /* initialize MPI */
    MPI_Init(&argc, &argv);
    MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
    MPI_Comm_rank(MPI_COMM_WORLD, &myrank);

216 217 218
    /* start workload load timer */
    load_start = MPI_Wtime();

219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239
    /* change the working directory to be the test directory */
    ret = chdir(replay_test_path);
    if (ret < 0)
    {
        fprintf(stderr, "Unable to change to testing directory (%s)\n", strerror(errno));
        goto error_exit;
    }

    /* set the path for logging this rank's events, if verbose is turned on */
    if (opt_verbose)
    {
        mkdir(log_dir, 0755);
        snprintf(my_log_path, MAX_NAME_LENGTH_WKLD, "%s/rank-%d.log", log_dir, myrank);
        log_stream = fopen(my_log_path, "w");
        if (log_stream == NULL)
        {
            fprintf(stderr, "Unable to open log file %s\n", my_log_path);
            goto error_exit;
        }
    }

240 241
    /* initialize workload generator from config file */
    workload_id = load_workload(conf_path, myrank);
242 243 244 245 246 247 248 249 250 251 252 253 254
    if (workload_id < 0)
    {
        goto error_exit;
    }

    /* initialize hash table for storing file descriptors */
    fd_table = qhash_init(hash_file_compare, quickhash_64bit_hash, 29);
    if (!fd_table)
    {
        fprintf(stderr, "File descriptor hash table memory error\n");
        goto error_exit;
    }

255 256 257
    /* loading is finished */
    load_end = MPI_Wtime();

258 259 260 261 262 263
    /* replay loop */
    while (1)
    {
        /* get the next replay operation from the workload generator */
        codes_workload_get_next(workload_id, myrank, &next_op);

264 265 266
        /* subtract the load time from any initial delay to be fair */
        if (first_op && (next_op.op_type == CODES_WK_DELAY))
        {
267
            first_op = 0;
268
            if (next_op.u.delay.seconds > (load_end - load_start))
269
            {
270
                next_op.u.delay.seconds -= (load_end - load_start);
271
            }
272
            else
273 274 275
            {
                printf("Replay workload is %.4lf seconds behind original workload\n",
                       (double)(load_end - load_start));
276
                continue;
277
            }
278 279
        }

280 281
        if (next_op.op_type != CODES_WK_END)
        {
282 283 284 285

            if (next_op.op_type == CODES_WK_DELAY)
                next_op.u.delay.seconds *= WORKLOAD_DELAY_PCT;

286 287 288 289 290 291 292 293 294 295 296 297 298 299
            /* replay the next workload operation */
            ret = replay_workload_op(next_op, myrank, replay_op_number++);
            if (ret < 0)
            {
                break;
            }
        }
        else
        {
            /* workload replay for this rank is complete */
            break;
        }
    }

300
    if (log_stream)
301 302
        fclose(log_stream);

303
    /* destroy and finalize the file descriptor hash table */
304
    qhash_destroy_and_finalize(fd_table, struct file_info, hash_link, free);
305

306 307 308 309 310 311
#if DEBUG_PROFILING
    printf("Rank %d:\td=%.4lf, b=%.4lf, o=%.4lf, c=%.4lf, r=%.4lf, w=%.4lf\n",
           myrank, total_delay_time, total_barrier_time, total_open_time,
           total_close_time, total_read_time, total_write_time);
#endif

312 313
error_exit:
    MPI_Finalize();
314

315 316 317 318 319 320 321 322
    return ret;
}

int replay_workload_op(struct codes_workload_op replay_op, int rank, long long int op_number)
{
    unsigned int secs;
    unsigned int usecs;
    int open_flags = O_RDWR;
323
    char file_name[250];
324
    int fildes;
325
    struct file_info *tmp_list = NULL;
326 327 328
    struct qlist_head *hash_link = NULL;
    char *buf = NULL;
    int ret;
329 330 331 332 333
    double start, end;

#if DEBUG_PROFILING
    start = MPI_Wtime();
#endif
334 335 336 337 338

    switch (replay_op.op_type)
    {
        case CODES_WK_DELAY:
            if (opt_verbose)
339
                fprintf(log_stream, "[Rank %d] Operation %lld : DELAY %lf seconds\n",
340 341 342 343 344
                       rank, op_number, replay_op.u.delay.seconds);

            if (!opt_noop)
            {
                /* satisfy delay using second delay then microsecond delay */
345 346
                secs = (unsigned int)replay_op.u.delay.seconds;
                usecs = (unsigned int)((replay_op.u.delay.seconds - secs) * 1000 * 1000);
347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363
                ret = sleep(secs);
                if (ret)
                {
                    /* error in sleep */
                    errno = EINTR;
                    fprintf(stderr, "Rank %d failure on operation %lld [DELAY: %s]\n",
                            rank, op_number, strerror(errno));
                    return -1;
                }
                ret = usleep(usecs);
                if (ret < 0)
                {
                    /* error in usleep */
                    fprintf(stderr, "Rank %d failure on operation %lld [DELAY: %s]\n",
                            rank, op_number, strerror(errno));
                    return -1;
                }
364 365 366 367
#if DEBUG_PROFILING
                end = MPI_Wtime();
                total_delay_time += (end - start);
#endif
368 369 370 371
            }
            return 0;
        case CODES_WK_BARRIER:
            if (opt_verbose)
372
                fprintf(log_stream, "[Rank %d] Operation %lld : BARRIER\n", rank, op_number);
373 374 375 376 377 378 379 380 381 382 383 384

            if (!opt_noop)
            {
                /* implement barrier using MPI global barrier on all ranks */
                ret = MPI_Barrier(MPI_COMM_WORLD);
                if (ret != MPI_SUCCESS)
                {
                    /* error in MPI_Barrier */
                    fprintf(stderr, "Rank %d failure on operation %lld [BARRIER: %s]\n",
                            rank, op_number, "Invalid communicator");
                    return -1;
                }
385 386 387 388 389

#if DEBUG_PROFILING
                end = MPI_Wtime();
                total_barrier_time += (end - start);
#endif
390 391 392 393
            }
            return 0;
        case CODES_WK_OPEN:
            if (opt_verbose)
394
                fprintf(log_stream, "[Rank %d] Operation %lld: %s file %"PRIu64"\n", rank, op_number,
395 396 397 398 399 400 401 402 403
                       (replay_op.u.open.create_flag) ? "CREATE" : "OPEN", replay_op.u.open.file_id);

            if (!opt_noop)
            {
                /* set the create flag, if necessary */
                if (replay_op.u.open.create_flag)
                    open_flags |= O_CREAT;

                /* write the file hash to string to be used as the actual file name */
404 405 406 407 408
                if (!opt_lockless)
                    snprintf(file_name, sizeof(file_name), "%"PRIu64, replay_op.u.open.file_id);
                else
                    snprintf(file_name, sizeof(file_name), "bglockless:%"PRIu64,
                             replay_op.u.open.file_id);
409 410 411 412 413 414 415 416 417 418 419 420

                /* perform the open operation */
                fildes = open(file_name, open_flags, 0666);
                if (fildes < 0)
                {
                    fprintf(stderr, "Rank %d failure on operation %lld [%s: %s]\n",
                            rank, op_number, (replay_op.u.open.create_flag) ? "CREATE" : "OPEN",
                            strerror(errno));
                    return -1;
                }

                /* save the file descriptor for this file in a hash table to be retrieved later */
421
                tmp_list = malloc(sizeof(struct file_info));
422 423 424 425 426 427 428 429 430
                if (!tmp_list)
                {
                    fprintf(stderr, "No memory available for file hash entry\n");
                    return -1;
                }

                tmp_list->file_hash = replay_op.u.open.file_id;
                tmp_list->file_descriptor = fildes;
                qhash_add(fd_table, &(replay_op.u.open.file_id), &(tmp_list->hash_link));
431 432 433 434 435

#if DEBUG_PROFILING
                end = MPI_Wtime();
                total_open_time += (end - start);
#endif
436 437 438 439
            }
            return 0;
        case CODES_WK_CLOSE:
            if (opt_verbose)
440 441
                fprintf(log_stream, "[Rank %d] Operation %lld : CLOSE file %"PRIu64"\n",
                        rank, op_number, replay_op.u.close.file_id);
442 443 444 445 446 447

            if (!opt_noop)
            {
                /* search for the corresponding file descriptor in the hash table */
                hash_link = qhash_search_and_remove(fd_table, &(replay_op.u.close.file_id));
                assert(hash_link);
448
                tmp_list = qhash_entry(hash_link, struct file_info, hash_link);
449 450
                fildes = tmp_list->file_descriptor;
                free(tmp_list);
451

452 453 454 455 456 457 458 459
                /* perform the close operation */
                ret = close(fildes);
                if (ret < 0)
                {
                    fprintf(stderr, "Rank %d failure on operation %lld [CLOSE: %s]\n",
                            rank, op_number, strerror(errno));
                    return -1;
                }
460 461 462 463 464

#if DEBUG_PROFILING
                end = MPI_Wtime();
                total_close_time += (end - start);
#endif
465 466 467 468
            }
            return 0;
        case CODES_WK_WRITE:
            if (opt_verbose)
469
                fprintf(log_stream, "[Rank %d] Operation %lld : WRITE file %"PRIu64" (sz = %"PRId64
470
                       ", off = %"PRId64")\n",
471
                       rank, op_number, replay_op.u.write.file_id, replay_op.u.write.size,
472 473 474 475 476 477 478
                       replay_op.u.write.offset);

            if (!opt_noop)
            {
                /* search for the corresponding file descriptor in the hash table */
                hash_link = qhash_search(fd_table, &(replay_op.u.write.file_id));
                assert(hash_link);
479
                tmp_list = qhash_entry(hash_link, struct file_info, hash_link);
480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496
                fildes = tmp_list->file_descriptor;

                /* perform the write operation */
                buf = malloc(replay_op.u.write.size);
                if (!buf)
                {
                    fprintf(stderr, "No memory available for write buffer\n");
                    return -1;
                }
                ret = pwrite(fildes, buf, replay_op.u.write.size, replay_op.u.write.offset);
                free(buf);
                if (ret < 0)
                {
                    fprintf(stderr, "Rank %d failure on operation %lld [WRITE: %s]\n",
                            rank, op_number, strerror(errno));
                    return -1;
                }
497 498 499 500 501

#if DEBUG_PROFILING
                end = MPI_Wtime();
                total_write_time += (end - start);
#endif
502 503 504 505
            }
            return 0;
        case CODES_WK_READ:
            if (opt_verbose)
506 507
                fprintf(log_stream, "[Rank %d] Operation %lld : READ file %"PRIu64" (sz = %"PRId64
                        ", off = %"PRId64")\n", rank, op_number, replay_op.u.read.file_id,
508
                       replay_op.u.read.size, replay_op.u.read.offset);
509

510 511 512 513 514
            if (!opt_noop)
            {
                /* search for the corresponding file descriptor in the hash table */
                hash_link = qhash_search(fd_table, &(replay_op.u.read.file_id));
                assert(hash_link);
515
                tmp_list = qhash_entry(hash_link, struct file_info, hash_link);
516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532
                fildes = tmp_list->file_descriptor;

                /* perform the write operation */
                buf = malloc(replay_op.u.read.size);
                if (!buf)
                {
                    fprintf(stderr, "No memory available for write buffer\n");
                    return -1;
                }
                ret = pread(fildes, buf, replay_op.u.read.size, replay_op.u.read.offset);
                free(buf);
                if (ret < 0)
                {
                    fprintf(stderr, "Rank %d failure on operation %lld [READ: %s]\n",
                            rank, op_number, strerror(errno));
                    return -1;
                }
533 534 535 536 537

#if DEBUG_PROFILING
                end = MPI_Wtime();
                total_read_time += (end - start);
#endif
538 539 540 541
            }
            return 0;
        default:
            fprintf(stderr, "** Rank %d: INVALID OPERATION (op count = %lld) **\n", rank, op_number);
542
            return 0;
543 544 545 546 547 548 549
    }

}

int hash_file_compare(void *key, struct qlist_head *link)
{
    uint64_t *in_file_hash = (uint64_t *)key;
550
    struct file_info *tmp_file;
551

552
    tmp_file = qlist_entry(link, struct file_info, hash_link);
553 554 555 556 557
    if (tmp_file->file_hash == *in_file_hash)
        return 1;

    return 0;
}
558 559 560 561 562 563 564 565 566 567

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