codes-darshan-io-wrkld.c 68.5 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 0.2
#define DEF_INTER_CYC_DELAY_PCT 0.4
17

18
#define DARSHAN_NEGLIGIBLE_DELAY 0.00001
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
/* 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;
40
41
42
43

    off_t next_coll_rd_off;
    off_t next_coll_wr_off;

44
45
46
47
    struct qhash_head hash_link;
};

/* Darshan workload generator's implementation of the CODES workload API */
48
49
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);
50
static int darshan_io_workload_get_rank_cnt(const char *params);
51
52
53
54
55
56
57
58
59
60
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);

61
/* Helper functions for implementing the Darshan workload generator */
62
63
64
65
66
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);
67
static double generate_psx_open_event(struct darshan_file *file, int create_flag,
68
                                      double meta_op_time, double cur_time,
69
                                      struct rank_io_context *io_context, int insert_flag);
70
static double generate_psx_close_event(struct darshan_file *file, double meta_op_time,
71
72
                                       double cur_time, struct rank_io_context *io_context,
                                       int insert_flag);
73
74
static double generate_barrier_event(struct darshan_file *file, int64_t root, double cur_time,
                                     struct rank_io_context *io_context);
75
static double generate_psx_ind_io_events(struct darshan_file *file, int64_t io_ops_this_cycle,
76
77
                                         double inter_io_delay, double cur_time,
                                         struct rank_io_context *io_context);
78
79
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,
80
                                          int64_t aggregator_cnt, double inter_io_delay,
81
82
                                          double meta_op_time, double cur_time,
                                          struct rank_io_context *io_context);
83
84
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);
85
86
87
static void determine_coll_io_params(struct darshan_file *file, int write_flag, int64_t coll_op_cnt,
                                     int64_t agg_cnt, int64_t agg_ndx, size_t *io_sz, off_t *io_off,
                                     struct rank_io_context *io_context);
88
static void calc_io_delays(struct darshan_file *file, int64_t num_opens, int64_t num_io_ops,
89
                           double total_delay, double *first_io_delay, double *close_delay,
90
91
92
                           double *inter_open_delay, double *inter_io_delay);
static void file_sanity_check(struct darshan_file *file, struct darshan_job *job);

93
94
95
96
97
98
/* 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,
99
    .codes_workload_get_rank_cnt = darshan_io_workload_get_rank_cnt,
100
101
};

102
103
static int total_rank_cnt = 0;

104
/* hash table to store per-rank workload contexts */
105
static struct qhash_table *rank_tbl = NULL;
106
static int rank_tbl_pop = 0;
107

108
109
110
/* load the workload generator for this rank, given input params */
static int darshan_io_workload_load(const char *params, int rank)
{
111
112
    darshan_params *d_params = (darshan_params *)params;
    darshan_fd logfile_fd;
113
114
    struct darshan_job job;
    struct darshan_file next_file;
115
    struct rank_io_context *my_ctx;
116
    int ret;
117

118
    if (!d_params)
119
120
        return -1;

121
    /* open the darshan log to begin reading in file i/o info */
122
123
    logfile_fd = darshan_log_open(d_params->log_file_path, "r");
    if (logfile_fd < 0)
124
        return -1;
125

126
127
128
129
130
131
132
    /* 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;
    }
133
134
135
136
137
    if (!total_rank_cnt)
    {
        total_rank_cnt = job.nprocs;
    }
    assert(rank < total_rank_cnt);
138

139
140
141
142
143
144
145
146
147
148
    /* 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();
149
150
    my_ctx->next_coll_rd_off = my_ctx->next_coll_wr_off = 0;

151

152
153
154
155
156
157
158
159
160
161
    /* 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 */
162
            generate_psx_ind_file_events(&next_file, my_ctx);
163
164
165
166
167
168
        }
        /* 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);
169
170
171

            /* 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);
172
        }
173
        else if (next_file.rank < rank)
174
            continue;
175
176
        else
            break;
177
178
179
180

        assert(next_file.counters[CP_POSIX_OPENS] == 0);
        assert(next_file.counters[CP_POSIX_READS] == 0);
        assert(next_file.counters[CP_POSIX_WRITES] == 0);
181
182
183
184
    }
    if (ret < 0)
        return -1;

185
    darshan_log_close(logfile_fd);
186

187
188
189
190
191
192
193
194
195
196
197
198
199
    /* 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));
200
    rank_tbl_pop++;
201

202
203
204
205
206
207
    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)
{
208
209
210
211
212
    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;

213
214
    assert(rank < total_rank_cnt);

215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
    /* 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);
237
238
 
        rank_tbl_pop--;
239
240
        if (!rank_tbl_pop)
        {
241
            qhash_finalize(rank_tbl);
242
243
            rank_tbl = NULL;
        }
244
245
246
247
248
249
250
251
252
    }
    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;
253
254
255
256

    return;
}

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
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;
}

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
368
369
370
371
/* 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]);

372
        if ((tmp->start_time - last_op_time) <= DARSHAN_NEGLIGIBLE_DELAY)
373
374
375
376
377
378
379
380
381
382
383
384
385
386
        {
            /* 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;
        }
    }
387
388
389
390
391
392
393
394
395

    /* 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;
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
428
429
430
431
}

/* 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 */
/*                                       */
/*****************************************/
432
433
434

/* generate events for an independently opened file, and store these events */
static void generate_psx_ind_file_events(
435
    struct darshan_file *file, struct rank_io_context *io_context)
436
437
438
{
    int64_t io_ops_this_cycle;
    double cur_time = file->fcounters[CP_F_OPEN_TIMESTAMP];
439
    double total_delay;
440
441
442
443
444
445
446
447
448
449
450
451
452
    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 */
453
454
455
    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];
456
457
458

    /* calculate synthetic delay values */
    calc_io_delays(file, file->counters[CP_POSIX_OPENS],
459
460
                   file->counters[CP_POSIX_READS] + file->counters[CP_POSIX_WRITES], total_delay,
                   &first_io_delay, &close_delay, &inter_open_delay, &inter_io_delay);
461
462
463

    /* calculate average meta op time (for i/o and opens/closes) */
    /* TODO: this needs to be updated when we add in stat, seek, etc. */
464
    meta_op_time = file->fcounters[CP_F_POSIX_META_TIME] / (2 * file->counters[CP_POSIX_OPENS]);
465

466
467
468
469
470
    /* set the create flag if the file was written to */
    if (file->counters[CP_BYTES_WRITTEN])
    {
        create_flag = 1;
    }
471
472
473
474
475
476

    /* 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 */
477
478
        cur_time = generate_psx_open_event(file, create_flag, meta_op_time, cur_time,
                                           io_context, 1);
479
480
481
482
483
484
485
486
487
488
        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 */
489
        cur_time = generate_psx_ind_io_events(file, io_ops_this_cycle, inter_io_delay,
490
                                              cur_time, io_context);
491
492
493
494
495

        /* 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 */
496
        cur_time = generate_psx_close_event(file, meta_op_time, cur_time, io_context, 1);
497
498
499
500
501
502
503

        /* account for potential interopen delay if more than one open */
        if (file->counters[CP_POSIX_OPENS] > 1)
        {
            cur_time += inter_open_delay;
        }
    }
504
505
506
507
508
509
510
511
512

    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)
{
513
514
515
516
517
    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;
518
519
    int64_t extra_opens = 0;
    int64_t extra_io_ops = 0;
520
    int64_t total_io_ops = file->counters[CP_POSIX_READS] + file->counters[CP_POSIX_WRITES];
521
522
523
524
    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;
525
    int64_t rank_cnt;
526
527
    int create_flag = 0;
    double cur_time = file->fcounters[CP_F_OPEN_TIMESTAMP];
528
    double total_delay;
529
530
531
532
533
    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;
534
    int64_t i;
535
536
537
538
539
540
541
542
543
544

    /* 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);

545
546
547
548
    total_delay = file->fcounters[CP_F_CLOSE_TIMESTAMP] -
                  file->fcounters[CP_F_OPEN_TIMESTAMP] -
                  file->fcounters[CP_F_SLOWEST_RANK_TIME];

549
550
    if (file->counters[CP_COLL_OPENS] || file->counters[CP_INDEP_OPENS])
    {
551
552
553
554
        /* Subtract a portion of MPI r/w overhead from the total delay.
         * This accounts for time spent synchronizing, since we generate barrier events.
         * Note: 60% of mpi read/write time seems to be an appropriate approx of sync duration.
         */
555
556
557
//        total_delay -= ((((file->fcounters[CP_F_MPI_READ_TIME] +
//                       file->fcounters[CP_F_MPI_WRITE_TIME]) / nprocs) -
//                       file->fcounters[CP_F_SLOWEST_RANK_TIME]) * .4);
558

559
560
        extra_opens = file->counters[CP_POSIX_OPENS] - file->counters[CP_COLL_OPENS] -
                      file->counters[CP_INDEP_OPENS];
561

562
563
        total_coll_opens = file->counters[CP_COLL_OPENS];
        total_ind_opens = file->counters[CP_POSIX_OPENS] - total_coll_opens - extra_opens;
564
565
566
567

        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;

568
569
        if (file->counters[CP_COLL_OPENS])
        {
570
            open_cycles = total_coll_opens / nprocs;
571
572
573
574
            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);
        }
575
        else
576
577
578
579
580
        {
            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);
        }
581
582
583
584
585
586
587
588
589
590
591
    }
    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;
592
593
594
595
596
            if (extra_io_ops != extra_opens)
            {
                extra_opens = 0;
                extra_io_ops = 0;
            }
597
598
599
        }

        total_coll_opens = 0;
600
        total_ind_opens = file->counters[CP_POSIX_OPENS] - extra_opens;
601
602
603

        total_ind_io_ops = total_io_ops - extra_io_ops;
        total_coll_io_ops = 0;
604
605
606
607

        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);
608
609
610
    }
    assert(extra_opens <= open_cycles);

611
    /* calculate average meta op time (for i/o and opens/closes) */
612
613
    meta_op_time = file->fcounters[CP_F_POSIX_META_TIME] / ((2 * file->counters[CP_POSIX_OPENS]) +
                   file->counters[CP_POSIX_READS] + file->counters[CP_POSIX_WRITES]);
614

615
616
617
618
619
    /* it is rare to overwrite existing files, so set the create flag */
    if (file->counters[CP_BYTES_WRITTEN])
    {
        create_flag = 1;
    }
620

621
622
    /* generate all events for this collectively opened file */
    for (i = 0; i < open_cycles; i++)
623
    {
624
625
        ind_opens_this_cycle = ceil((double)total_ind_opens / (open_cycles - i));
        coll_opens_this_cycle = total_coll_opens / (open_cycles - i);
626

627
628
629
        /* assign any extra opens to rank 0 (these may correspond to file creations or
         * header reads/writes)
         */
630
        if (extra_opens && !(i % (open_cycles / extra_opens)))
631
        {
632
633
            cur_time = generate_psx_open_event(file, create_flag, meta_op_time, cur_time,
                                               io_context, (io_context->my_rank == 0));
634
635
            create_flag = 0;

636
637
            if (!file->counters[CP_COLL_OPENS] && !file->counters[CP_INDEP_OPENS])
            {
638
                /* TODO: we probably want to use ind_io here */
639
                cur_time = generate_psx_coll_io_events(file, 1, 0, nprocs, nprocs, 0.0,
640
                                                       meta_op_time, 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
        /* TODO: look at this, use ind_io */
650
        while (ind_opens_this_cycle)
651
        {
652
653
654
655
656
            if (ind_opens_this_cycle >= nprocs)
                rank_cnt = nprocs;
            else
                rank_cnt = ind_opens_this_cycle;

657
658
            cur_time = generate_psx_open_event(file, create_flag, meta_op_time, cur_time,
                                               io_context, (io_context->my_rank < rank_cnt));
659
660
            create_flag = 0;

661
662
663
664
            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,
665
666
                                                   nprocs, inter_io_delay, meta_op_time,
                                                   cur_time, io_context);
667
668
669
            total_ind_io_ops -= ind_io_ops_this_cycle;

            cur_time += close_delay;
670

671
672
            cur_time = generate_psx_close_event(file, meta_op_time, cur_time, io_context,
                                                (io_context->my_rank < rank_cnt));
673

674
675
676
677
678
679
            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;
680
681
        }

682
        while (coll_opens_this_cycle)
683
        {
684
            assert(!create_flag);
685

686
            cur_time = generate_barrier_event(file, 0, cur_time, io_context);
687
688

            cur_time = generate_psx_open_event(file, create_flag, meta_op_time,
689
                                               cur_time, io_context, 1);
690

691
692
693
694
695
696
697
698
699
700
701
702
            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,
703
704
                                                   inter_io_delay, meta_op_time, cur_time,
                                                   io_context);
705
706
707
708
            total_ind_io_ops -= ind_io_ops_this_cycle;
            total_coll_io_ops -= coll_io_ops_this_cycle;

            cur_time += close_delay;
709

710
            cur_time = generate_psx_close_event(file, meta_op_time, cur_time, io_context, 1);
711
712

            file->counters[CP_POSIX_OPENS] -= nprocs;
713
714
            file->counters[CP_COLL_OPENS] -= nprocs;
            coll_opens_this_cycle -= nprocs;
715
716
717
718
            total_coll_opens -= nprocs;

            if (file->counters[CP_POSIX_OPENS])
                cur_time += inter_cycle_delay;
719
720
        }
    }
721
722
723
724
725
726

    return;
}

/* fill in an open event structure and store it with the rank context */
static double generate_psx_open_event(
727
    struct darshan_file *file, int create_flag, double meta_op_time,
728
    double cur_time, struct rank_io_context *io_context, int insert_flag)
729
{
730
731
732
733
734
735
736
    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
    };
737
738
739

    /* set the end time of the event based on time spent in POSIX meta operations */
    cur_time += meta_op_time;
740
    next_io_op.end_time = cur_time;
741

742
    /* store the open event (if this rank performed it) */
743
    if (insert_flag)
744
        darshan_insert_next_io_op(io_context->io_op_dat, &next_io_op);
745
746
747
748
749
750

    return cur_time;
}

/* fill in a close event structure and store it with the rank context */
static double generate_psx_close_event(
751
    struct darshan_file *file, double meta_op_time, double cur_time,
752
    struct rank_io_context *io_context, int insert_flag)
753
{
754
755
756
757
758
759
    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
    };
760
761
762

    /* set the end time of the event based on time spent in POSIX meta operations */
    cur_time += meta_op_time;
763
    next_io_op.end_time = cur_time;
764

765
    /* store the close event (if this rank performed it) */
766
    if (insert_flag)
767
        darshan_insert_next_io_op(io_context->io_op_dat, &next_io_op);
768
769
770
771
772
773

    return cur_time;
}

/* fill in a barrier event structure and store it with the rank context */
static double generate_barrier_event(
774
    struct darshan_file *file, int64_t root, double cur_time, struct rank_io_context *io_context)
775
{
776
777
778
779
780
781
782
    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
    };
783
784

    cur_time += .000001; /* small synthetic delay representing time to barrier */
785
    next_io_op.end_time = cur_time;
786

787
    /* store the barrier event */
788
    if (file->rank == -1)
789
        darshan_insert_next_io_op(io_context->io_op_dat, &next_io_op);
790
791
792
793
794
795

    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(
796
    struct darshan_file *file, int64_t io_ops_this_cycle, double inter_io_delay,
797
    double cur_time, struct rank_io_context *io_context)
798
799
800
801
802
803
804
805
806
{
    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;
807
    struct darshan_io_op next_io_op;
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
841
842
843

    /* 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 */
844
        determine_io_params(file, rw, 1, 1, &io_sz, &io_off);
845
846
847
        if (!rw)
        {
            /* generate a read event */
848
849
850
851
852
            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;
853
854
855
856
857

            /* set the end time based on observed bandwidth and io size */
            if (rd_bw == 0.0)
                io_op_time = 0.0;
            else
858
                io_op_time = (io_sz / rd_bw);
859
860

            /* update time, accounting for metadata time */
861
            cur_time += io_op_time;
862
            next_io_op.end_time = cur_time;
863
864
865
866
867
            file->counters[CP_POSIX_READS]--;
        }
        else
        {
            /* generate a write event */
868
869
870
871
872
            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;
873
874
875
876
877

            /* set the end time based on observed bandwidth and io size */
            if (wr_bw == 0.0)
                io_op_time = 0.0;
            else
878
                io_op_time = (io_sz / wr_bw);
879
880

            /* update time, accounting for metadata time */
881
            cur_time += io_op_time;
882
            next_io_op.end_time = cur_time;
883
884
885
886
887
888
889
            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);

890
891
        /* store the i/o event */
        darshan_insert_next_io_op(io_context->io_op_dat, &next_io_op);
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911

        /* 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 */
912
            cur_time += inter_io_delay;
913
914
915
916
917
918
919
920
921
922
923
924
        }
    }

    /* 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;
}

925
926
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,
927
928
    int64_t nprocs, int64_t aggregator_cnt, double inter_io_delay, double meta_op_time,
    double cur_time, struct rank_io_context *io_context)
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 tmp_rank;
    int64_t next_ind_io_rank = 0;
    int64_t io_cnt;
938
    double ranks_per_aggregator = (double)(nprocs - 1) / (aggregator_cnt - 1);
939
    int64_t ind_ops_remaining = 0;
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
    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;

963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
        /* 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);
        }

985
986
987
988
989
990
991
        /* 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];
    }

992
993
994
995
996
    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;

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

1036
1037
1038
            for (j = 0; j < io_cnt; j += aggregator_cnt)
            {
                int64_t tmp_coll_cnt = MIN(io_cnt - j, aggregator_cnt);
1039
                int64_t tmp_agg_ndx;
1040
1041
1042

                cur_time = generate_barrier_event(file, 0, cur_time, io_context);

1043
1044
1045
                tmp_agg_ndx = (int64_t)round(io_context->my_rank / ranks_per_aggregator);
                if ((round(tmp_agg_ndx * ranks_per_aggregator) == io_context->my_rank) &&
                    (tmp_agg_ndx < tmp_coll_cnt))
1046
                {
1047
                    determine_coll_io_params(file, rw, io_cnt, tmp_coll_cnt, tmp_agg_ndx + 1, 
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
                                             &io_sz, &io_off, io_context);
                    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);
                        
1064
                        next_io_op.end_time = cur_time + io_op_time + meta_op_time;
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
                        file->counters[CP_POSIX_READS] -= tmp_coll_cnt;
                    }
                    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);

1082
                        next_io_op.end_time = cur_time + io_op_time + meta_op_time;
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
                        file->counters[CP_POSIX_WRITES] -= tmp_coll_cnt;
                    }

                    darshan_insert_next_io_op(io_context->io_op_dat, &next_io_op);

                    cur_time = next_io_op.end_time;
                }
                else
                {
                    if (!rw)
                    {
                        file->counters[CP_POSIX_READS] -= tmp_coll_cnt;
                    }
                    else
                    {
                        file->counters[CP_POSIX_WRITES] -= tmp_coll_cnt;
                    }
                }
                psx_rw_ops_remaining -= tmp_coll_cnt;
                assert(file->counters[CP_POSIX_READS] >= 0);
                assert(file->counters[CP_POSIX_WRITES] >= 0);
            }
1105
        }
1106
1107
        io_ops_this_rw--;
        max_cur_time = cur_time;
1108

1109
#if 0
1110
1111
        for (j = 0; j < io_cnt; j++)
        {
1112
            determine_coll_io_params(file, rw, io_cnt - j, &io_sz, &io_off);
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
            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);
                
1128
                next_io_op.end_time = cur_time + io_op_time;
1129
1130
1131
1132
1133
                file->counters[CP_POSIX_READS]--;
            }
            else
            {
                /* generate a write event */
1134
                ext_io_op.codes_op.op_type = CODES_WK_WRITE;
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
                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);

1146
                next_io_op.end_time = cur_time + io_op_time;
1147
1148
1149
1150
1151
1152
                file->counters[CP_POSIX_WRITES]--;
            }
            psx_rw_ops_remaining--;
            assert(file->counters[CP_POSIX_READS] >= 0);
            assert(file->counters[CP_POSIX_WRITES] >= 0);

1153
            /* TODO store the i/o event */
1154
1155
1156
1157
1158
1159
1160
            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;
1161
            if (ind_coll && (tmp_rank >= (ranks_per_aggregator * aggregator_cnt)))
1162
1163
1164
            {
                tmp_rank = 0;
                cur_time = max_cur_time;
1165
1166
                if (j != io_cnt -1)
                    cur_time = generate_barrier_event(file, 0, cur_time, io_context);
1167
1168
1169
            }
        }
        io_ops_this_rw--;
1170
#endif
1171

1172
1173
1174
1175
1176
1177
1178
1179
        if (ind_coll)
        {
            cur_time = max_cur_time;
            if (i != (total_io_ops_this_cycle - 1))
                cur_time += inter_io_delay;
        }
        else
        {
1180
1181
            if (tmp_rank == (nprocs - 1) || (i == (total_io_ops_this_cycle - 1)))
            {
1182
1183
                cur_time = max_cur_time;

1184
1185
1186
                if (i != (total_io_ops_this_cycle - 1))
                    cur_time += inter_io_delay;
            }
1187
1188
        }

1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
        /* 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;
    }

1228
    return cur_time;
1229
1230
}

1231
1232
1233
1234
1235
1236
1237
1238
static void determine_coll_io_params(
    struct darshan_file *file, int write_flag, int64_t coll_op_cnt, int64_t agg_cnt,
    int64_t agg_ndx, size_t *io_sz, off_t *io_off, struct rank_io_context *io_context)
{
    static int64_t size_bins_left = 0;
    static int64_t agg_size_bins[10] = { 0 };
    static off_t agg_off = 0;
    int i, j;
1239
    int start_ndx, end_ndx;
1240
1241
1242
1243
1244
1245
    off_t *next_coll_off;
    int64_t *all_size_bins = NULL;
    int64_t *common_accesses = &(file->counters[CP_ACCESS1_ACCESS]); /* 4 common accesses */
    int64_t *common_access_counts = &(file->counters[CP_ACCESS1_COUNT]); /* common access counts */
    int64_t *total_io_size = NULL;
    int64_t tmp_cnt;
1246
1247
1248
    int64_t switch_cnt;
    int64_t leftover;
    int64_t bins_to_use;
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
    const int64_t size_bin_min_vals[10] = { 0, 100, 1024, 10 * 1024, 100 * 1024, 1024 * 1024,
                                            4 * 1024 * 1024, 10 * 1024 * 1024, 100 * 1024 * 1024,
                                            1024 * 1024 * 1024 };
    const int64_t size_bin_max_vals[10] = { 100, 1024, 10 * 1024, 100 * 1024, 1024 * 1024,
                                            4 * 1024 * 1024, 10 * 1024 * 1024, 100 * 1024 * 1024,
                                            1024 * 1024 * 1024, INT64_MAX };
    
    if (write_flag)
    {
        all_size_bins = &(file->counters[CP_SIZE_WRITE_0_100]);
        total_io_size = &(file->counters[CP_BYTES_WRITTEN]);
        next_coll_off = &(io_context->next_coll_wr_off);
    }
    else
    {
        all_size_bins = &(file->counters[CP_SIZE_READ_0_100]);
        total_io_size = &(file->counters[CP_BYTES_READ]);
        next_coll_off = &(io_context->next_coll_rd_off);
    }

    /* we enter this if statement if we have not yet calculated which size bins to use for the
     * current collective I/O operation.
     */
    if (!size_bins_left)
    {
        int64_t total_agg_size_bin_cnt = 0;
        int tmp_ndx = 9;

        /* find some size bins that we can assign accesses out of.
         * Note: we require the bins be reasonably close to each other in size, and be less
         * than 100 MiB.
         */
        for (i = 7; i >= 0; i--)
        {
            if (all_size_bins[i])
            {
                if (total_agg_size_bin_cnt)
                {
                    if ((tmp_ndx - 3) <= i)
                    {
                        tmp_ndx = i;
                        total_agg_size_bin_cnt += all_size_bins[i];
                    }
                    break;
                }
                else
                {
                    tmp_ndx = i;
                    total_agg_size_bin_cnt += all_size_bins[i];
                }
            }
        }

        /* assign accesses from found bins proportional to bin size */
        for (i = 7; i >= tmp_ndx; i--)
        {
            if (all_size_bins[i])
            {
                agg_size_bins[i] = ((double)all_size_bins[i] / total_agg_size_bin_cnt) * coll_op_cnt;
                size_bins_left += agg_size_bins[i];
                all_size_bins[i] -= agg_size_bins[i];
            }

            if (size_bins_left == coll_op_cnt) break;
        }

        /* if we still haven't assigned enough accesses, just assign them stupidly */
        if (size_bins_left < coll_op_cnt)
        {
            for (i = 9; i >= 0; i--)
            {
                if (all_size_bins[i])
                {
                    tmp_cnt = MIN(all_size_bins[i], coll_op_cnt - size_bins_left);
                    agg_size_bins[i] += tmp_cnt;
                    size_bins_left += tmp_cnt;
                    all_size_bins[i] -= tmp_cnt;
                }

                if (size_bins_left == coll_op_cnt) break;
            }
        }
        assert(size_bins_left == coll_op_cnt);

        ssize_t tmp_size;
1334
1335
        int64_t tmp_agg_ndx = 1; /* start at aggregator 1 */
        int64_t tmp_agg_cnt;
1336
        int64_t tmp_common_cnts[4];
1337
        int64_t tmp_size_bins[10];
1338
        memcpy(tmp_common_cnts, common_access_counts, 4 * sizeof(int64_t));
1339
1340
        memcpy(tmp_size_bins, agg_size_bins, 10 * sizeof(int64_t));
        tmp_cnt = coll_op_cnt;
1341
1342
        agg_off = *next_coll_off;

1343
        while (tmp_cnt)
1344
        {
1345
1346
1347
1348
1349
1350
            start_ndx = -1;
            end_ndx = -1;
            leftover = 0;
            tmp_agg_cnt = MIN(tmp_cnt, agg_cnt);
        
            for (i = 9; i >= 0; i--)
1351
            {
1352
                if (tmp_size_bins[i])
1353
                {
1354
1355
1356
1357
                    if (start_ndx == -1) start_ndx = i;

                    tmp_agg_cnt -= tmp_size_bins[i];
                    if (tmp_agg_cnt <= 0)
1358
                    {
1359
                        end_ndx = i;
1360
1361
1362
                        break;
                    }
                }
1363
            }
1364

1365
1366
1367
1368
1369
            i = start_ndx;
            tmp_agg_cnt = MIN(tmp_cnt, agg_cnt);
            while (tmp_agg_cnt)
            {
                if ((tmp_size_bins[i] >= tmp_agg_cnt) && !leftover)
1370
                {
1371
                    switch_cnt = tmp_agg_cnt;
1372
                }