codes-darshan-io-wrkld.c 55.8 KB
Newer Older
1 2 3 4 5
/*
 * Copyright (C) 2013 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
 */
6
#include <assert.h>
7
#include <math.h>
8 9

#include "codes/codes-workload.h"
10
#include "codes/quickhash.h"
11
#include "codes-workload-method.h"
12

13
#include "darshan-logutils.h"
14

15 16
#define DEF_INTER_IO_DELAY_PCT 0.4
#define DEF_INTER_CYC_DELAY_PCT 0.2
17

18
#define DARSHAN_NEGLIGIBLE_DELAY 0.001
19 20 21

#define RANK_HASH_TABLE_SIZE 397

22
#define MIN(a, b) (((a) < (b)) ? (a) : (b))
23

24 25
#define ALIGN_BY_8(x) ((x) + ((x) % 8))

26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43
/* structure for storing a darshan workload operation (a codes op with 2 timestamps) */
struct darshan_io_op
{
    struct codes_workload_op codes_op;
    double start_time;
    double end_time;
};

/* I/O context structure managed by each rank in the darshan workload */
struct rank_io_context
{
    int64_t my_rank;
    double last_op_time;
    void *io_op_dat;
    struct qhash_head hash_link;
};

/* Darshan workload generator's implementation of the CODES workload API */
44 45
static int darshan_io_workload_load(const char *params, int rank);
static void darshan_io_workload_get_next(int rank, struct codes_workload_op *op);
46
static int darshan_io_workload_get_rank_cnt(const char *params);
47 48 49 50 51 52 53 54 55 56
static int darshan_rank_hash_compare(void *key, struct qhash_head *link);

/* Darshan I/O op data structure access (insert, remove) abstraction */
static void *darshan_init_io_op_dat(void);
static void darshan_insert_next_io_op(void *io_op_dat, struct darshan_io_op *io_op);
static void darshan_remove_next_io_op(void *io_op_dat, struct darshan_io_op *io_op,
                                      double last_op_time);
static void darshan_finalize_io_op_dat(void *io_op_dat);
static int darshan_io_op_compare(const void *p1, const void *p2);

57
/* Helper functions for implementing the Darshan workload generator */
58 59 60 61 62
static void generate_psx_ind_file_events(struct darshan_file *file,
                                         struct rank_io_context *io_context);
static void generate_psx_coll_file_events(struct darshan_file *file,
                                          struct rank_io_context *io_context,
                                          int64_t nprocs, int64_t aggregator_cnt);
63
static double generate_psx_open_event(struct darshan_file *file, int create_flag,
64
                                      double meta_op_time, double cur_time,
65
                                      struct rank_io_context *io_context, int insert_flag);
66
static double generate_psx_close_event(struct darshan_file *file, double meta_op_time,
67 68
                                       double cur_time, struct rank_io_context *io_context,
                                       int insert_flag);
69 70
static double generate_barrier_event(struct darshan_file *file, int64_t root, double cur_time,
                                     struct rank_io_context *io_context);
71
static double generate_psx_ind_io_events(struct darshan_file *file, int64_t io_ops_this_cycle,
72 73
                                         double inter_io_delay, double cur_time,
                                         struct rank_io_context *io_context);
74 75
static double generate_psx_coll_io_events(struct darshan_file *file, int64_t ind_io_ops_this_cycle,
                                          int64_t coll_io_ops_this_cycle, int64_t nprocs,
76
                                          int64_t aggregator_cnt, double inter_io_delay,
77
                                          double cur_time, struct rank_io_context *io_context);
78 79
static void determine_io_params(struct darshan_file *file, int write_flag, int64_t io_this_op,
                                int64_t proc_count, size_t *io_sz, off_t *io_off);
80
static void calc_io_delays(struct darshan_file *file, int64_t num_opens, int64_t num_io_ops,
81
                           double total_delay, double *first_io_delay, double *close_delay,
82 83 84
                           double *inter_open_delay, double *inter_io_delay);
static void file_sanity_check(struct darshan_file *file, struct darshan_job *job);

85 86 87 88 89 90
/* workload method name and function pointers for the CODES workload API */
struct codes_workload_method darshan_io_workload_method =
{
    .method_name = "darshan_io_workload",
    .codes_workload_load = darshan_io_workload_load,
    .codes_workload_get_next = darshan_io_workload_get_next,
91
    .codes_workload_get_rank_cnt = darshan_io_workload_get_rank_cnt,
92 93
};

94 95
static int total_rank_cnt = 0;

96
/* hash table to store per-rank workload contexts */
97
static struct qhash_table *rank_tbl = NULL;
98
static int rank_tbl_pop = 0;
99

100 101 102
/* load the workload generator for this rank, given input params */
static int darshan_io_workload_load(const char *params, int rank)
{
103 104
    darshan_params *d_params = (darshan_params *)params;
    darshan_fd logfile_fd;
105 106
    struct darshan_job job;
    struct darshan_file next_file;
107
    struct rank_io_context *my_ctx;
108
    struct darshan_io_op end_op;
109
    int ret;
110

111
    if (!d_params)
112 113
        return -1;

114
    /* open the darshan log to begin reading in file i/o info */
115 116
    logfile_fd = darshan_log_open(d_params->log_file_path, "r");
    if (logfile_fd < 0)
117
        return -1;
118

119 120 121 122 123 124 125
    /* get the per-job stats from the log */
    ret = darshan_log_getjob(logfile_fd, &job);
    if (ret < 0)
    {
        darshan_log_close(logfile_fd);
        return -1;
    }
126 127 128 129 130
    if (!total_rank_cnt)
    {
        total_rank_cnt = job.nprocs;
    }
    assert(rank < total_rank_cnt);
131

132 133 134 135 136 137 138 139 140 141 142
    /* allocate the i/o context needed by this rank */
    my_ctx = malloc(sizeof(struct rank_io_context));
    if (!my_ctx)
    {
        darshan_log_close(logfile_fd);
        return -1;
    }
    my_ctx->my_rank = (int64_t)rank;
    my_ctx->last_op_time = 0.0;
    my_ctx->io_op_dat = darshan_init_io_op_dat();

143 144 145 146 147 148 149 150 151 152
    /* loop over all files contained in the log file */
    while ((ret = darshan_log_getfile(logfile_fd, &job, &next_file)) > 0)
    {
        /* generate all i/o events contained in this independent file */
        if (next_file.rank == rank)
        {
            /* make sure the file i/o counters are valid */
            file_sanity_check(&next_file, &job);

            /* generate i/o events and store them in this rank's workload context */
153
            generate_psx_ind_file_events(&next_file, my_ctx);
154 155 156 157 158 159
        }
        /* generate all i/o events involving this rank in this collective file */
        else if (next_file.rank == -1)
        {
            /* make sure the file i/o counters are valid */
            file_sanity_check(&next_file, &job);
160 161 162

            /* generate collective i/o events and store them in the rank context */
            generate_psx_coll_file_events(&next_file, my_ctx, job.nprocs, d_params->aggregator_cnt);
163
        }
164
        else if (next_file.rank < rank)
165
            continue;
166 167
        else
            break;
168 169 170 171

        assert(next_file.counters[CP_POSIX_OPENS] == 0);
        assert(next_file.counters[CP_POSIX_READS] == 0);
        assert(next_file.counters[CP_POSIX_WRITES] == 0);
172 173 174 175
    }
    if (ret < 0)
        return -1;

176
    darshan_log_close(logfile_fd);
177

178 179 180 181 182 183
    /* tack a workload end file on the end of the events list */
    /* NOTE: this helps us generate a delay from the last close to program termination */
    end_op.codes_op.op_type = CODES_WK_END;
    end_op.start_time = end_op.end_time = job.end_time - job.start_time + 1;
    darshan_insert_next_io_op(my_ctx->io_op_dat, &end_op);

184 185 186 187 188 189 190 191 192 193 194 195 196
    /* finalize the rank's i/o context so i/o ops may be retrieved later (in order) */
    darshan_finalize_io_op_dat(my_ctx->io_op_dat);

    /* initialize the hash table of rank contexts, if it has not been initialized */
    if (!rank_tbl)
    {
        rank_tbl = qhash_init(darshan_rank_hash_compare, quickhash_64bit_hash, RANK_HASH_TABLE_SIZE);
        if (!rank_tbl)
            return -1;
    }

    /* add this rank context to the hash table */
    qhash_add(rank_tbl, &(my_ctx->my_rank), &(my_ctx->hash_link));
197
    rank_tbl_pop++;
198

199 200 201 202 203 204
    return 0;
}

/* pull the next event (independent or collective) for this rank from its event context */
static void darshan_io_workload_get_next(int rank, struct codes_workload_op *op)
{
205 206 207 208 209
    int64_t my_rank = (int64_t)rank;
    struct qhash_head *hash_link = NULL;
    struct rank_io_context *tmp = NULL;
    struct darshan_io_op next_io_op;

210 211
    assert(rank < total_rank_cnt);

212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233
    /* find i/o context for this rank in the rank hash table */
    hash_link = qhash_search(rank_tbl, &my_rank);

    /* terminate the workload if there is no valid rank context */
    if (!hash_link)
    {
        op->op_type = CODES_WK_END;
        return;
    }

    /* get access to the rank's io_context data */
    tmp = qhash_entry(hash_link, struct rank_io_context, hash_link);
    assert(tmp->my_rank == my_rank);

    /* get the next darshan i/o op out of this rank's context */
    darshan_remove_next_io_op(tmp->io_op_dat, &next_io_op, tmp->last_op_time);

    /* free the rank's i/o context if this is the last i/o op */
    if (next_io_op.codes_op.op_type == CODES_WK_END)
    {
        qhash_del(hash_link);
        free(tmp);
234 235
 
        rank_tbl_pop--;
236
        if (!rank_tbl_pop){
237
            qhash_finalize(rank_tbl);
238 239
            rank_tbl = NULL;
        }
240 241 242 243 244 245 246 247 248
    }
    else
    {
        /* else, set the last op time to be the end of the returned op */
        tmp->last_op_time = next_io_op.end_time;
    }

    /* return the codes op contained in the darshan i/o op */
    *op = next_io_op.codes_op;
249 250 251 252

    return;
}

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
static int darshan_io_workload_get_rank_cnt(const char *params)
{
    darshan_params *d_params = (darshan_params *)params;
    darshan_fd logfile_fd;
    struct darshan_job job;
    int ret;

    if (!d_params)
        return -1;

    /* open the darshan log to begin reading in file i/o info */
    logfile_fd = darshan_log_open(d_params->log_file_path, "r");
    if (logfile_fd < 0)
        return -1;

    /* get the per-job stats from the log */
    ret = darshan_log_getjob(logfile_fd, &job);
    if (ret < 0)
    {
        darshan_log_close(logfile_fd);
        return -1;
    }

    darshan_log_close(logfile_fd);

    return job.nprocs;
}

281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367
/* comparison function for comparing two hash keys (used for storing multiple io contexts) */
static int darshan_rank_hash_compare(
    void *key, struct qhash_head *link)
{
    int64_t *in_rank = (int64_t *)key;
    struct rank_io_context *tmp;

    tmp = qhash_entry(link, struct rank_io_context, hash_link);
    if (tmp->my_rank == *in_rank)
        return 1;

    return 0;
}

/*****************************************/
/*                                       */
/*   Darshan I/O op storage abstraction  */
/*                                       */
/*****************************************/

#define DARSHAN_IO_OP_INC_CNT 100000

/* dynamically allocated array data structure for storing darshan i/o events */
struct darshan_io_dat_array
{
    struct darshan_io_op *op_array;
    int64_t op_arr_ndx;
    int64_t op_arr_cnt;
};

/* initialize the dynamic array data structure */
static void *darshan_init_io_op_dat()
{
    struct darshan_io_dat_array *tmp;

    /* initialize the array data structure */
    tmp = malloc(sizeof(struct darshan_io_dat_array));
    assert(tmp);
    tmp->op_array = malloc(DARSHAN_IO_OP_INC_CNT * sizeof(struct darshan_io_op));
    assert(tmp->op_array);
    tmp->op_arr_ndx = 0;
    tmp->op_arr_cnt = DARSHAN_IO_OP_INC_CNT;

    /* return the array info for this rank's i/o context */
    return (void *)tmp;
}

/* store the i/o event in this rank's i/o context */
static void darshan_insert_next_io_op(
    void *io_op_dat, struct darshan_io_op *io_op)
{
    struct darshan_io_dat_array *array = (struct darshan_io_dat_array *)io_op_dat;
    struct darshan_io_op *tmp;

    /* realloc array if it is already full */
    if (array->op_arr_ndx == array->op_arr_cnt)
    {
        tmp = malloc((array->op_arr_cnt + DARSHAN_IO_OP_INC_CNT) * sizeof(struct darshan_io_op));
        assert(tmp);
        memcpy(tmp, array->op_array, array->op_arr_cnt * sizeof(struct darshan_io_op));
        free(array->op_array);
        array->op_array = tmp;
        array->op_arr_cnt += DARSHAN_IO_OP_INC_CNT;
    }

    /* add the darshan i/o op to the array */
    array->op_array[array->op_arr_ndx++] = *io_op;

    return;
}

/* pull the next i/o event out of this rank's i/o context */
static void darshan_remove_next_io_op(
    void *io_op_dat, struct darshan_io_op *io_op, double last_op_time)
{
    struct darshan_io_dat_array *array = (struct darshan_io_dat_array *)io_op_dat;

    /* if the array has been scanned completely already */
    if (array->op_arr_ndx == array->op_arr_cnt)
    {
        /* no more events just end the workload */
        io_op->codes_op.op_type = CODES_WK_END;
    }
    else
    {
        struct darshan_io_op *tmp = &(array->op_array[array->op_arr_ndx]);

368
        if ((tmp->start_time - last_op_time) <= DARSHAN_NEGLIGIBLE_DELAY)
369 370 371 372 373 374 375 376 377 378 379 380 381 382
        {
            /* there is no delay, just return the next op in the array */
            *io_op = *tmp;
            array->op_arr_ndx++;
        }
        else
        {
            /* there is a nonnegligible delay, so generate and return a delay event */
            io_op->codes_op.op_type = CODES_WK_DELAY;
            io_op->codes_op.u.delay.seconds = tmp->start_time - last_op_time;
            io_op->start_time = last_op_time;
            io_op->end_time = tmp->start_time;
        }
    }
383 384 385 386 387 388 389 390 391

    /* if this is the end op, free data structures */
    if (io_op->codes_op.op_type == CODES_WK_END)
    {
        free(array->op_array);
        free(array);
    }

    return;
392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427
}

/* sort the dynamic array in order of i/o op start time */
static void darshan_finalize_io_op_dat(
    void *io_op_dat)
{
    struct darshan_io_dat_array *array = (struct darshan_io_dat_array *)io_op_dat;

    /* sort this rank's i/o op list */
    qsort(array->op_array, array->op_arr_ndx, sizeof(struct darshan_io_op), darshan_io_op_compare);
    array->op_arr_cnt = array->op_arr_ndx;
    array->op_arr_ndx = 0;

    return;
}

/* comparison function for sorting darshan_io_ops in order of start timestamps */
static int darshan_io_op_compare(
    const void *p1, const void *p2)
{
    struct darshan_io_op *a = (struct darshan_io_op *)p1;
    struct darshan_io_op *b = (struct darshan_io_op *)p2;

    if (a->start_time < b->start_time)
        return -1;
    else if (a->start_time > b->start_time)
        return 1;
    else
        return 0;
}

/*****************************************/
/*                                       */
/* Darshan workload generation functions */
/*                                       */
/*****************************************/
428 429 430

/* generate events for an independently opened file, and store these events */
static void generate_psx_ind_file_events(
431
    struct darshan_file *file, struct rank_io_context *io_context)
432 433 434
{
    int64_t io_ops_this_cycle;
    double cur_time = file->fcounters[CP_F_OPEN_TIMESTAMP];
435
    double total_delay;
436 437 438 439 440 441 442 443 444 445 446 447 448
    double first_io_delay = 0.0;
    double close_delay = 0.0;
    double inter_open_delay = 0.0;
    double inter_io_delay = 0.0;
    double meta_op_time;
    int create_flag;
    int64_t i;

    /* if the file was never really opened, just return because we have no timing info */
    if (file->counters[CP_POSIX_OPENS] == 0)
        return;

    /* determine delay available per open-io-close cycle */
449 450 451
    total_delay = file->fcounters[CP_F_CLOSE_TIMESTAMP] - file->fcounters[CP_F_OPEN_TIMESTAMP] -
                  file->fcounters[CP_F_POSIX_READ_TIME] - file->fcounters[CP_F_POSIX_WRITE_TIME] -
                  file->fcounters[CP_F_POSIX_META_TIME];
452 453 454

    /* calculate synthetic delay values */
    calc_io_delays(file, file->counters[CP_POSIX_OPENS],
455 456
                   file->counters[CP_POSIX_READS] + file->counters[CP_POSIX_WRITES], total_delay,
                   &first_io_delay, &close_delay, &inter_open_delay, &inter_io_delay);
457 458 459

    /* calculate average meta op time (for i/o and opens/closes) */
    /* TODO: this needs to be updated when we add in stat, seek, etc. */
460
    meta_op_time = file->fcounters[CP_F_POSIX_META_TIME] / (2 * file->counters[CP_POSIX_OPENS]);
461

462 463 464 465 466
    /* set the create flag if the file was written to */
    if (file->counters[CP_BYTES_WRITTEN])
    {
        create_flag = 1;
    }
467 468 469 470 471 472

    /* generate open/io/close events for all cycles */
    /* TODO: add stats */
    for (i = 0; file->counters[CP_POSIX_OPENS]; i++, file->counters[CP_POSIX_OPENS]--)
    {
        /* generate an open event */
473 474
        cur_time = generate_psx_open_event(file, create_flag, meta_op_time, cur_time,
                                           io_context, 1);
475 476 477 478 479 480 481 482 483 484
        create_flag = 0;

        /* account for potential delay from first open to first io */
        cur_time += first_io_delay;

        io_ops_this_cycle = ceil((double)(file->counters[CP_POSIX_READS] +
                                 file->counters[CP_POSIX_WRITES]) /
                                 file->counters[CP_POSIX_OPENS]);

        /* perform the calculated number of i/o operations for this file open */
485
        cur_time = generate_psx_ind_io_events(file, io_ops_this_cycle, inter_io_delay,
486
                                              cur_time, io_context);
487 488 489 490 491

        /* account for potential delay from last io to close */
        cur_time += close_delay;

        /* generate a close for the open event at the start of the loop */
492
        cur_time = generate_psx_close_event(file, meta_op_time, cur_time, io_context, 1);
493 494 495 496 497 498 499

        /* account for potential interopen delay if more than one open */
        if (file->counters[CP_POSIX_OPENS] > 1)
        {
            cur_time += inter_open_delay;
        }
    }
500 501 502 503 504 505 506 507 508

    return;
}

/* generate events for the i/o ops stored in a collectively opened file for this rank */
void generate_psx_coll_file_events(
    struct darshan_file *file, struct rank_io_context *io_context,
    int64_t nprocs, int64_t in_agg_cnt)
{
509 510 511 512 513
    int64_t open_cycles;
    int64_t total_ind_opens;
    int64_t total_coll_opens;
    int64_t ind_opens_this_cycle;
    int64_t coll_opens_this_cycle;
514 515
    int64_t extra_opens = 0;
    int64_t extra_io_ops = 0;
516
    int64_t total_io_ops = file->counters[CP_POSIX_READS] + file->counters[CP_POSIX_WRITES];
517 518 519 520
    int64_t total_ind_io_ops;
    int64_t total_coll_io_ops;
    int64_t ind_io_ops_this_cycle;
    int64_t coll_io_ops_this_cycle;
521
    int64_t rank_cnt;
522 523
    int create_flag = 0;
    double cur_time = file->fcounters[CP_F_OPEN_TIMESTAMP];
524
    double total_delay;
525 526 527 528 529
    double first_io_delay = 0.0;
    double close_delay = 0.0;
    double inter_cycle_delay = 0.0;
    double inter_io_delay = 0.0;
    double meta_op_time;
530
    int64_t i;
531 532 533 534 535 536 537 538 539 540

    /* the collective file was never opened (i.e., just stat-ed), so return */
    if (!(file->counters[CP_POSIX_OPENS]))
        return;

    /*  in this case, posix opens are less than mpi opens...
     *  this is probably a mpi deferred open -- assume app will not use this, currently.
     */
    assert(file->counters[CP_POSIX_OPENS] >= nprocs);

541 542
    if (file->counters[CP_COLL_OPENS] || file->counters[CP_INDEP_OPENS])
    {
543 544
        extra_opens = file->counters[CP_POSIX_OPENS] - file->counters[CP_COLL_OPENS] -
                      file->counters[CP_INDEP_OPENS];
545

546 547
        total_coll_opens = file->counters[CP_COLL_OPENS];
        total_ind_opens = file->counters[CP_POSIX_OPENS] - total_coll_opens - extra_opens;
548 549 550 551

        total_ind_io_ops = file->counters[CP_INDEP_READS] + file->counters[CP_INDEP_WRITES];
        total_coll_io_ops = (file->counters[CP_COLL_READS] + file->counters[CP_COLL_WRITES]) / nprocs;

552 553 554 555 556 557 558 559
        if (file->counters[CP_COLL_OPENS])
        {
            total_delay = (file->fcounters[CP_F_CLOSE_TIMESTAMP] -
                           file->fcounters[CP_F_OPEN_TIMESTAMP] -
                           (file->fcounters[CP_F_POSIX_READ_TIME] / in_agg_cnt) -
                           (file->fcounters[CP_F_POSIX_WRITE_TIME] / in_agg_cnt) -
                           (file->fcounters[CP_F_POSIX_META_TIME] / in_agg_cnt));

560
            open_cycles = total_coll_opens / nprocs;
561 562 563 564
            calc_io_delays(file, ceil(((double)(total_coll_opens + total_ind_opens)) / nprocs),
                           total_coll_io_ops + ceil((double)total_ind_io_ops / nprocs), total_delay,
                           &first_io_delay, &close_delay, &inter_cycle_delay, &inter_io_delay);
        }
565
        else
566 567 568 569 570 571
        {
            total_delay = (file->fcounters[CP_F_CLOSE_TIMESTAMP] -
                           file->fcounters[CP_F_OPEN_TIMESTAMP] -
                           (file->fcounters[CP_F_POSIX_READ_TIME] / nprocs) -
                           (file->fcounters[CP_F_POSIX_WRITE_TIME] / nprocs) -
                           (file->fcounters[CP_F_POSIX_META_TIME] / nprocs));
572

573 574 575 576
            open_cycles = ceil((double)total_ind_opens / nprocs);
            calc_io_delays(file, open_cycles, ceil((double)total_ind_io_ops / nprocs), total_delay,
                           &first_io_delay, &close_delay, &inter_cycle_delay, &inter_io_delay);
        }
577 578 579 580 581 582 583 584 585 586 587
    }
    else
    {
        extra_opens = file->counters[CP_POSIX_OPENS] % nprocs;
        if (extra_opens && ((file->counters[CP_POSIX_OPENS] / nprocs) % extra_opens))
        {
            extra_opens = 0;
        }
        else
        {
            extra_io_ops = total_io_ops % nprocs;
588 589 590 591 592
            if (extra_io_ops != extra_opens)
            {
                extra_opens = 0;
                extra_io_ops = 0;
            }
593 594 595
        }

        total_coll_opens = 0;
596
        total_ind_opens = file->counters[CP_POSIX_OPENS] - extra_opens;
597 598 599

        total_ind_io_ops = total_io_ops - extra_io_ops;
        total_coll_io_ops = 0;
600

601 602 603 604 605 606
        total_delay = (file->fcounters[CP_F_CLOSE_TIMESTAMP] -
                       file->fcounters[CP_F_OPEN_TIMESTAMP] -
                       (file->fcounters[CP_F_POSIX_READ_TIME] / nprocs) -
                       (file->fcounters[CP_F_POSIX_WRITE_TIME] / nprocs) -
                       (file->fcounters[CP_F_POSIX_META_TIME] / nprocs));

607 608 609
        open_cycles = ceil((double)total_ind_opens / nprocs);
        calc_io_delays(file, open_cycles, ceil((double)total_ind_io_ops / nprocs), total_delay,
                       &first_io_delay, &close_delay, &inter_cycle_delay, &inter_io_delay);
610 611 612
    }
    assert(extra_opens <= open_cycles);

613
    /* calculate average meta op time (for i/o and opens/closes) */
614
    meta_op_time = file->fcounters[CP_F_POSIX_META_TIME] / (2 * file->counters[CP_POSIX_OPENS]);
615

616 617 618 619 620
    /* it is rare to overwrite existing files, so set the create flag */
    if (file->counters[CP_BYTES_WRITTEN])
    {
        create_flag = 1;
    }
621

622 623
    /* generate all events for this collectively opened file */
    for (i = 0; i < open_cycles; i++)
624
    {
625 626
        ind_opens_this_cycle = ceil((double)total_ind_opens / (open_cycles - i));
        coll_opens_this_cycle = total_coll_opens / (open_cycles - i);
627

628 629 630
        /* assign any extra opens to rank 0 (these may correspond to file creations or
         * header reads/writes)
         */
631
        if (extra_opens && !(i % (open_cycles / extra_opens)))
632
        {
633 634
            cur_time = generate_psx_open_event(file, create_flag, meta_op_time, cur_time,
                                               io_context, (io_context->my_rank == 0));
635 636
            create_flag = 0;

637 638 639
            if (!file->counters[CP_COLL_OPENS] && !file->counters[CP_INDEP_OPENS])
            {
                cur_time = generate_psx_coll_io_events(file, 1, 0, nprocs, nprocs, 0.0,
640
                                                       cur_time, io_context);
641 642
                extra_io_ops--;
            }
643

644 645
            cur_time = generate_psx_close_event(file, meta_op_time, cur_time, io_context,
                                                (io_context->my_rank == 0));
646 647 648
            file->counters[CP_POSIX_OPENS]--;
        }

649
        while (ind_opens_this_cycle)
650
        {
651 652 653 654 655
            if (ind_opens_this_cycle >= nprocs)
                rank_cnt = nprocs;
            else
                rank_cnt = ind_opens_this_cycle;

656 657
            cur_time = generate_psx_open_event(file, create_flag, meta_op_time, cur_time,
                                               io_context, (io_context->my_rank < rank_cnt));
658 659
            create_flag = 0;

660 661 662 663
            cur_time += first_io_delay;

            ind_io_ops_this_cycle = ceil(((double)total_ind_io_ops / total_ind_opens) * rank_cnt);
            cur_time = generate_psx_coll_io_events(file, ind_io_ops_this_cycle, 0, nprocs,
664
                                                   nprocs, inter_io_delay, cur_time, io_context);
665 666 667
            total_ind_io_ops -= ind_io_ops_this_cycle;

            cur_time += close_delay;
668

669 670
            cur_time = generate_psx_close_event(file, meta_op_time, cur_time, io_context,
                                                (io_context->my_rank < rank_cnt));
671

672 673 674 675 676 677
            file->counters[CP_POSIX_OPENS] -= rank_cnt;
            ind_opens_this_cycle -= rank_cnt;
            total_ind_opens -= rank_cnt;

            if (file->counters[CP_POSIX_OPENS])
                cur_time += inter_cycle_delay;
678 679
        }

680
        while (coll_opens_this_cycle)
681
        {
682
            assert(!create_flag);
683

684
            cur_time = generate_barrier_event(file, 0, cur_time, io_context);
685 686

            cur_time = generate_psx_open_event(file, create_flag, meta_op_time,
687
                                               cur_time, io_context, 1);
688

689 690 691 692 693 694 695 696 697 698 699 700
            cur_time += first_io_delay;

            if (file->counters[CP_INDEP_OPENS])
                ind_io_ops_this_cycle = 0;
            else
                ind_io_ops_this_cycle = ceil((double)total_ind_io_ops / 
                                             (file->counters[CP_COLL_OPENS] / nprocs));

            coll_io_ops_this_cycle = ceil((double)total_coll_io_ops / 
                                          (file->counters[CP_COLL_OPENS] / nprocs));
            cur_time = generate_psx_coll_io_events(file, ind_io_ops_this_cycle,
                                                   coll_io_ops_this_cycle, nprocs, in_agg_cnt,
701
                                                   inter_io_delay, cur_time, io_context);
702 703 704 705
            total_ind_io_ops -= ind_io_ops_this_cycle;
            total_coll_io_ops -= coll_io_ops_this_cycle;

            cur_time += close_delay;
706

707
            cur_time = generate_psx_close_event(file, meta_op_time, cur_time, io_context, 1);
708 709

            file->counters[CP_POSIX_OPENS] -= nprocs;
710 711
            file->counters[CP_COLL_OPENS] -= nprocs;
            coll_opens_this_cycle -= nprocs;
712 713 714 715
            total_coll_opens -= nprocs;

            if (file->counters[CP_POSIX_OPENS])
                cur_time += inter_cycle_delay;
716 717
        }
    }
718 719 720 721 722 723

    return;
}

/* fill in an open event structure and store it with the rank context */
static double generate_psx_open_event(
724
    struct darshan_file *file, int create_flag, double meta_op_time,
725
    double cur_time, struct rank_io_context *io_context, int insert_flag)
726
{
727 728 729 730 731 732 733
    struct darshan_io_op next_io_op = 
    {
        .codes_op.op_type = CODES_WK_OPEN,
        .codes_op.u.open.file_id = file->hash,
        .codes_op.u.open.create_flag = create_flag,
        .start_time = cur_time
    };
734 735 736

    /* set the end time of the event based on time spent in POSIX meta operations */
    cur_time += meta_op_time;
737
    next_io_op.end_time = cur_time;
738

739
    /* store the open event (if this rank performed it) */
740
    if (insert_flag)
741
        darshan_insert_next_io_op(io_context->io_op_dat, &next_io_op);
742 743 744 745 746 747

    return cur_time;
}

/* fill in a close event structure and store it with the rank context */
static double generate_psx_close_event(
748
    struct darshan_file *file, double meta_op_time, double cur_time,
749
    struct rank_io_context *io_context, int insert_flag)
750
{
751 752 753 754 755 756
    struct darshan_io_op next_io_op =
    {
        .codes_op.op_type = CODES_WK_CLOSE,
        .codes_op.u.close.file_id = file->hash,
        .start_time = cur_time
    };
757 758 759

    /* set the end time of the event based on time spent in POSIX meta operations */
    cur_time += meta_op_time;
760
    next_io_op.end_time = cur_time;
761

762
    /* store the close event (if this rank performed it) */
763
    if (insert_flag)
764
        darshan_insert_next_io_op(io_context->io_op_dat, &next_io_op);
765 766 767 768 769 770

    return cur_time;
}

/* fill in a barrier event structure and store it with the rank context */
static double generate_barrier_event(
771
    struct darshan_file *file, int64_t root, double cur_time, struct rank_io_context *io_context)
772
{
773 774 775 776 777 778 779
    struct darshan_io_op next_io_op =
    {
        .codes_op.op_type = CODES_WK_BARRIER, 
        .codes_op.u.barrier.count = -1, /* all processes */
        .codes_op.u.barrier.root = root,
        .start_time = cur_time
    };
780 781

    cur_time += .000001; /* small synthetic delay representing time to barrier */
782
    next_io_op.end_time = cur_time;
783

784
    /* store the barrier event */
785
    if (file->rank == -1)
786
        darshan_insert_next_io_op(io_context->io_op_dat, &next_io_op);
787 788 789 790 791 792

    return cur_time;
}

/* generate all i/o events for one independent file open and store them with the rank context */
static double generate_psx_ind_io_events(
793
    struct darshan_file *file, int64_t io_ops_this_cycle, double inter_io_delay,
794
    double cur_time, struct rank_io_context *io_context)
795 796 797 798 799 800 801 802 803
{
    static int rw = -1; /* rw = 1 for write, 0 for read, -1 for uninitialized */
    static int64_t io_ops_this_rw;
    static double rd_bw = 0.0, wr_bw = 0.0;
    int64_t psx_rw_ops_remaining = file->counters[CP_POSIX_READS] + file->counters[CP_POSIX_WRITES];
    double io_op_time;
    size_t io_sz;
    off_t io_off;
    int64_t i;
804
    struct darshan_io_op next_io_op;
805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840

    /* if there are no i/o ops, just return immediately */
    if (!io_ops_this_cycle)
        return cur_time;

    /* initialze static variables when a new file is opened */
    if (rw == -1)
    {
        /* initialize rw to be the first i/o operation found in the log */
        if (file->fcounters[CP_F_WRITE_START_TIMESTAMP] == 0.0)
            rw = 0;
        else if (file->fcounters[CP_F_READ_START_TIMESTAMP] == 0.0)
            rw = 1;
        else
            rw = (file->fcounters[CP_F_READ_START_TIMESTAMP] <
                  file->fcounters[CP_F_WRITE_START_TIMESTAMP]) ? 0 : 1;

        /* determine how many io ops to do before next rw switch */
        if (!rw)
            io_ops_this_rw = file->counters[CP_POSIX_READS] /
                             ((file->counters[CP_RW_SWITCHES] / 2) + 1);
        else
            io_ops_this_rw = file->counters[CP_POSIX_WRITES] /
                             ((file->counters[CP_RW_SWITCHES] / 2) + 1);

        /* initialize the rd and wr bandwidth values using total io size and time */
        if (file->fcounters[CP_F_POSIX_READ_TIME])
            rd_bw = file->counters[CP_BYTES_READ] / file->fcounters[CP_F_POSIX_READ_TIME];
        if (file->fcounters[CP_F_POSIX_WRITE_TIME])
            wr_bw = file->counters[CP_BYTES_WRITTEN] / file->fcounters[CP_F_POSIX_WRITE_TIME];
    }

    /* loop to generate all reads/writes for this open/close sequence */
    for (i = 0; i < io_ops_this_cycle; i++)
    {
        /* calculate what value to use for i/o size and offset */
841
        determine_io_params(file, rw, 1, 1, &io_sz, &io_off);
842 843 844
        if (!rw)
        {
            /* generate a read event */
845 846 847 848 849
            next_io_op.codes_op.op_type = CODES_WK_READ;
            next_io_op.codes_op.u.read.file_id = file->hash;
            next_io_op.codes_op.u.read.size = io_sz;
            next_io_op.codes_op.u.read.offset = io_off;
            next_io_op.start_time = cur_time;
850 851 852 853 854

            /* set the end time based on observed bandwidth and io size */
            if (rd_bw == 0.0)
                io_op_time = 0.0;
            else
855
                io_op_time = (io_sz / rd_bw);
856 857

            /* update time, accounting for metadata time */
858
            cur_time += io_op_time;
859
            next_io_op.end_time = cur_time;
860 861 862 863 864
            file->counters[CP_POSIX_READS]--;
        }
        else
        {
            /* generate a write event */
865 866 867 868 869
            next_io_op.codes_op.op_type = CODES_WK_WRITE;
            next_io_op.codes_op.u.write.file_id = file->hash;
            next_io_op.codes_op.u.write.size = io_sz;
            next_io_op.codes_op.u.write.offset = io_off;
            next_io_op.start_time = cur_time;
870 871 872 873 874

            /* set the end time based on observed bandwidth and io size */
            if (wr_bw == 0.0)
                io_op_time = 0.0;
            else
875
                io_op_time = (io_sz / wr_bw);
876 877

            /* update time, accounting for metadata time */
878
            cur_time += io_op_time;
879
            next_io_op.end_time = cur_time;
880 881 882 883 884 885 886
            file->counters[CP_POSIX_WRITES]--;
        }
        psx_rw_ops_remaining--;
        io_ops_this_rw--;
        assert(file->counters[CP_POSIX_READS] >= 0);
        assert(file->counters[CP_POSIX_WRITES] >= 0);

887 888
        /* store the i/o event */
        darshan_insert_next_io_op(io_context->io_op_dat, &next_io_op);
889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908

        /* determine whether to toggle between reads and writes */
        if (!io_ops_this_rw && psx_rw_ops_remaining)
        {
            /* toggle the read/write flag */
            rw ^= 1;
            file->counters[CP_RW_SWITCHES]--;

            /* determine how many io ops to do before next rw switch */
            if (!rw)
                io_ops_this_rw = file->counters[CP_POSIX_READS] /
                                 ((file->counters[CP_RW_SWITCHES] / 2) + 1);
            else
                io_ops_this_rw = file->counters[CP_POSIX_WRITES] /
                                 ((file->counters[CP_RW_SWITCHES] / 2) + 1);
        }

        if (i != (io_ops_this_cycle - 1))
        {
            /* update current time to account for possible delay between i/o operations */
909
            cur_time += inter_io_delay;
910 911 912 913 914 915 916 917 918 919 920 921
        }
    }

    /* reset the static rw flag if this is the last open-close cycle for this file */
    if (file->counters[CP_POSIX_OPENS] == 1)
    {
        rw = -1;
    }

    return cur_time;
}

922 923
static double generate_psx_coll_io_events(
    struct darshan_file *file, int64_t ind_io_ops_this_cycle, int64_t coll_io_ops_this_cycle,
924 925
    int64_t nprocs, int64_t aggregator_cnt, double inter_io_delay, double cur_time,
    struct rank_io_context *io_context)
926 927 928 929 930 931 932 933 934 935 936 937
{
    static int rw = -1; /* rw = 1 for write, 0 for read, -1 for uninitialized */
    static int64_t io_ops_this_rw;
    static double rd_bw = 0.0, wr_bw = 0.0;
    int64_t psx_rw_ops_remaining = file->counters[CP_POSIX_READS] + file->counters[CP_POSIX_WRITES];
    int64_t total_io_ops_this_cycle = ind_io_ops_this_cycle + coll_io_ops_this_cycle;
    int64_t total_coll_io_ops =
            (file->counters[CP_COLL_READS] + file->counters[CP_COLL_WRITES]) / nprocs;
    int64_t tmp_rank;
    int64_t next_ind_io_rank = 0;
    int64_t io_cnt;
    int64_t ranks_per_aggregator = nprocs / aggregator_cnt;
938
    int64_t ind_ops_remaining = 0;
939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961
    double io_op_time;
    double max_cur_time = 0.0;
    int ind_coll;
    size_t io_sz;
    off_t io_off;
    int64_t i, j;
    struct darshan_io_op next_io_op;

    if (!total_io_ops_this_cycle)
        return cur_time;

    /* initialze static variables when a new file is opened */
    if (rw == -1)
    {
        /* initialize rw to be the first i/o operation found in the log */
        if (file->fcounters[CP_F_WRITE_START_TIMESTAMP] == 0.0)
            rw = 0;
        else if (file->fcounters[CP_F_READ_START_TIMESTAMP] == 0.0)
            rw = 1;
        else
            rw = (file->fcounters[CP_F_READ_START_TIMESTAMP] <
                  file->fcounters[CP_F_WRITE_START_TIMESTAMP]) ? 0 : 1;

962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983
        /* determine how many io ops to do before next rw switch */
        if (!rw)
        {
            if (file->counters[CP_COLL_OPENS])
                io_ops_this_rw =
                    ((file->counters[CP_COLL_READS] / nprocs) + file->counters[CP_INDEP_READS]) /
                    ((file->counters[CP_RW_SWITCHES] / (2 * aggregator_cnt)) + 1);
            else
                io_ops_this_rw = file->counters[CP_POSIX_READS] /
                                 ((file->counters[CP_RW_SWITCHES] / (2 * aggregator_cnt)) + 1);
        }
        else
        {
            if (file->counters[CP_COLL_OPENS])
                io_ops_this_rw =
                    ((file->counters[CP_COLL_WRITES] / nprocs) + file->counters[CP_INDEP_WRITES]) /
                    ((file->counters[CP_RW_SWITCHES] / (2 * aggregator_cnt)) + 1);
            else
                io_ops_this_rw = file->counters[CP_POSIX_WRITES] /
                                 ((file->counters[CP_RW_SWITCHES] / (2 * aggregator_cnt)) + 1);
        }

984 985 986 987 988 989 990
        /* initialize the rd and wr bandwidth values using total io size and time */
        if (file->fcounters[CP_F_POSIX_READ_TIME])
            rd_bw = file->counters[CP_BYTES_READ] / file->fcounters[CP_F_POSIX_READ_TIME];
        if (file->fcounters[CP_F_POSIX_WRITE_TIME])
            wr_bw = file->counters[CP_BYTES_WRITTEN] / file->fcounters[CP_F_POSIX_WRITE_TIME];
    }

991 992 993 994 995
    if (coll_io_ops_this_cycle)
        ind_ops_remaining = ceil((double)ind_io_ops_this_cycle / coll_io_ops_this_cycle);
    else
        ind_ops_remaining = ind_io_ops_this_cycle;

996 997
    for (i = 0; i < total_io_ops_this_cycle; i++)
    {
998
        if (ind_ops_remaining)
999 1000 1001 1002
        {
            ind_coll = 0;
            tmp_rank = (next_ind_io_rank++) % nprocs;
            io_cnt = 1;
1003 1004
            ind_io_ops_this_cycle--;
            ind_ops_remaining--;
1005 1006 1007 1008 1009 1010 1011 1012 1013
            if (!rw)
                file->counters[CP_INDEP_READS]--;
            else
                file->counters[CP_INDEP_WRITES]--;
        }
        else
        {
            ind_coll = 1;
            tmp_rank = 0;
1014
            coll_io_ops_this_cycle--;
1015 1016 1017
            if (!rw)
            {
                io_cnt = ceil((double)(file->counters[CP_POSIX_READS] -
1018
                              file->counters[CP_INDEP_READS]) / 
1019 1020 1021 1022 1023 1024
                              (file->counters[CP_COLL_READS] / nprocs));
                file->counters[CP_COLL_READS] -= nprocs;
            }
            else
            {
                io_cnt = ceil((double)(file->counters[CP_POSIX_WRITES] -
1025
                              file->counters[CP_INDEP_WRITES]) / 
1026 1027 1028
                              (file->counters[CP_COLL_WRITES] / nprocs));
                file->counters[CP_COLL_WRITES] -= nprocs;
            }
1029 1030 1031 1032 1033 1034 1035

            if (coll_io_ops_this_cycle)
                ind_ops_remaining = ceil((double)ind_io_ops_this_cycle / coll_io_ops_this_cycle);
            else
                ind_ops_remaining = ind_io_ops_this_cycle;

            cur_time = generate_barrier_event(file, 0, cur_time, io_context);
1036 1037 1038 1039
        }

        for (j = 0; j < io_cnt; j++)
        {
1040 1041
            determine_io_params(file, rw, (ind_coll) ? io_cnt - j : ind_io_ops_this_cycle + 1,
                                aggregator_cnt, &io_sz, &io_off);
1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056
            if (!rw)
            {
                /* generate a read event */
                next_io_op.codes_op.op_type = CODES_WK_READ;
                next_io_op.codes_op.u.read.file_id = file->hash;
                next_io_op.codes_op.u.read.size = io_sz;
                next_io_op.codes_op.u.read.offset = io_off;
                next_io_op.start_time = cur_time;

                /* set the end time based on observed bandwidth and io size */
                if (rd_bw == 0.0)
                    io_op_time = 0.0;
                else
                    io_op_time = (io_sz / rd_bw);
                
1057
                next_io_op.end_time = cur_time + io_op_time;
1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074
                file->counters[CP_POSIX_READS]--;
            }
            else
            {
                /* generate a write event */
                next_io_op.codes_op.op_type = CODES_WK_WRITE;
                next_io_op.codes_op.u.write.file_id = file->hash;
                next_io_op.codes_op.u.write.size = io_sz;
                next_io_op.codes_op.u.write.offset = io_off;
                next_io_op.start_time = cur_time;

                /* set the end time based on observed bandwidth and io size */
                if (wr_bw == 0.0)
                    io_op_time = 0.0;
                else
                    io_op_time = (io_sz / wr_bw);

1075
                next_io_op.end_time = cur_time + io_op_time;
1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089
                file->counters[CP_POSIX_WRITES]--;
            }
            psx_rw_ops_remaining--;
            assert(file->counters[CP_POSIX_READS] >= 0);
            assert(file->counters[CP_POSIX_WRITES] >= 0);

            /*  store the i/o event */
            if (tmp_rank == io_context->my_rank)
                darshan_insert_next_io_op(io_context->io_op_dat, &next_io_op);

            if (next_io_op.end_time > max_cur_time)
                max_cur_time = next_io_op.end_time;

            tmp_rank += ranks_per_aggregator;
1090
            if (ind_coll && (tmp_rank >= (ranks_per_aggregator * aggregator_cnt)))
1091 1092 1093
            {
                tmp_rank = 0;
                cur_time = max_cur_time;
1094
                cur_time = generate_barrier_event(file, 0, cur_time, io_context);
1095 1096 1097 1098
            }
        }
        io_ops_this_rw--;

1099 1100 1101 1102 1103 1104 1105 1106 1107 1108
        if (ind_coll)
        {
            total_coll_io_ops--;

            cur_time = max_cur_time;
            if (i != (total_io_ops_this_cycle - 1))
                cur_time += inter_io_delay;
        }
        else
        {
1109 1110
            if (tmp_rank == (nprocs - 1) || (i == (total_io_ops_this_cycle - 1)))
            {
1111 1112
                cur_time = max_cur_time;

1113 1114 1115
                if (i != (total_io_ops_this_cycle - 1))
                    cur_time += inter_io_delay;
            }
1116 1117
        }

1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156
        /* determine whether to toggle between reads and writes */
        if (!io_ops_this_rw && psx_rw_ops_remaining)
        {
            /* toggle the read/write flag */
            rw ^= 1;
            file->counters[CP_RW_SWITCHES] -= aggregator_cnt;

            /* determine how many io ops to do before next rw switch */
            if (!rw)
            {
                if (file->counters[CP_COLL_OPENS])
                    io_ops_this_rw =
                        ((file->counters[CP_COLL_READS] / nprocs) +
                        file->counters[CP_INDEP_READS]) / ((file->counters[CP_RW_SWITCHES] /
                        (2 * aggregator_cnt)) + 1);
                else
                    io_ops_this_rw = file->counters[CP_POSIX_READS] /
                                     ((file->counters[CP_RW_SWITCHES] / (2 * aggregator_cnt)) + 1);
            }
            else
            {
                if (file->counters[CP_COLL_OPENS])
                    io_ops_this_rw =
                        ((file->counters[CP_COLL_WRITES] / nprocs) +
                        file->counters[CP_INDEP_WRITES]) / ((file->counters[CP_RW_SWITCHES] /
                        (2 * aggregator_cnt)) + 1);
                else
                    io_ops_this_rw = file->counters[CP_POSIX_WRITES] /
                                     ((file->counters[CP_RW_SWITCHES] / (2 * aggregator_cnt)) + 1);
            }
        }
    }

    /* reset the static rw flag if this is the last open-close cycle for this file */
    if (file->counters[CP_POSIX_OPENS] <= nprocs)
    {
        rw = -1;
    }

1157
    return cur_time;
1158 1159
}

1160
/* WARNING: BRUTE FORCE */