codes-darshan-io-wrkld.c 57.6 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
17
#define DEF_INTER_IO_DELAY_PCT 0.2
#define DEF_INTER_CYC_DELAY_PCT 0.4

18
19
20
21
#define DARSHAN_NEGLIGIBLE_DELAY .001

#define RANK_HASH_TABLE_SIZE 397

22
23
24
25
26
#define IO_IS_IN_SIZE_BIN_RANGE(size, bin_ndx, bin_min_sizes)                       \
        ((bin_ndx == 9) ?                                                           \
        (size >= bin_min_sizes[bin_ndx]) :                                          \
        ((size >= bin_min_sizes[bin_ndx]) && (size < bin_min_sizes[bin_ndx + 1])))

27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
/* 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 */
45
46
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);
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
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);

/* Helper functions for implementing the (complex, nonfactored) Darshan workload generator */
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
65
                                      double meta_op_time, double cur_time,
                                      struct rank_io_context *io_context);
66
static double generate_psx_close_event(struct darshan_file *file, double meta_op_time,
67
68
69
                                       double cur_time, struct rank_io_context *io_context);
static double generate_barrier_event(struct darshan_file *file, int64_t root, double cur_time,
                                     struct rank_io_context *io_context);
70
71
static double generate_psx_ind_io_events(struct darshan_file *file, int64_t io_ops_this_cycle,
                                         int64_t open_ndx, double inter_io_delay, 
72
73
74
75
76
77
78
                                         double meta_op_time, double cur_time,
                                         struct rank_io_context *io_context);
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,
                                          int64_t aggregator_cnt, int64_t open_ndx,
                                          double inter_io_delay, double meta_op_time, double cur_time,
                                          struct rank_io_context *io_context);
79
80
81
82
83
84
85
static void determine_io_params(struct darshan_file *file, int write_flag, int coll_flag,
                                int64_t io_cycles, size_t *io_sz, off_t *io_off);
static void calc_io_delays(struct darshan_file *file, int64_t num_opens, int64_t num_io_ops,
                           double delay_per_cycle, double *first_io_delay, double *close_delay,
                           double *inter_open_delay, double *inter_io_delay);
static void file_sanity_check(struct darshan_file *file, struct darshan_job *job);

86
87
88
89
90
91
92
93
/* 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,
};

94
/* hash table to store per-rank workload contexts */
95
static struct qhash_table *rank_tbl = NULL;
96

97
98
99
/* load the workload generator for this rank, given input params */
static int darshan_io_workload_load(const char *params, int rank)
{
100
101
    darshan_params *d_params = (darshan_params *)params;
    darshan_fd logfile_fd;
102
103
    struct darshan_job job;
    struct darshan_file next_file;
104
    struct rank_io_context *my_ctx;
105
    int ret;
106

107
    if (!d_params)
108
109
        return -1;

110
111
    /* (re)seed the random number generator */
    srand(time(NULL));
112

113
    /* open the darshan log to begin reading in file i/o info */
114
115
    logfile_fd = darshan_log_open(d_params->log_file_path, "r");
    if (logfile_fd < 0)
116
        return -1;
117

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
131
132
133
134
135
136
    /* 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();

137
138
139
140
141
142
143
144
145
146
    /* 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 */
147
            generate_psx_ind_file_events(&next_file, my_ctx);
148
149
150
151
152
153
        }
        /* 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);
154
155
156

            /* 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);
157
158
159
160
161
        }
    }
    if (ret < 0)
        return -1;

162
    darshan_log_close(logfile_fd);
163

164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
    /* 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));

    /* fill out the info required for this workload group */
    if (darshan_workload_info.group_id == -1)
    {
        darshan_workload_info.group_id = 1;
        darshan_workload_info.min_rank = 0;
        darshan_workload_info.max_rank = job.nprocs - 1;
        darshan_workload_info.num_lrank = job.nprocs;
    }

187
188
189
190
191
192
    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)
{
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
    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;

    /* 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);
    }
    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;
229
230
231
232

    return;
}

233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
/* 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;

        /* free data structures */
        free(array->op_array);
        free(array);
    }
    else
    {
        struct darshan_io_op *tmp = &(array->op_array[array->op_arr_ndx]);

        if ((tmp->start_time - last_op_time) < DARSHAN_NEGLIGIBLE_DELAY)
        {
            /* 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;
        }
    }
}

/* 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 */
/*                                       */
/*****************************************/
375
376
377

/* generate events for an independently opened file, and store these events */
static void generate_psx_ind_file_events(
378
    struct darshan_file *file, struct rank_io_context *io_context)
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
{
    int64_t io_ops_this_cycle;
    double cur_time = file->fcounters[CP_F_OPEN_TIMESTAMP];
    double delay_per_open;
    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 */
    delay_per_open = (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]) / file->counters[CP_POSIX_OPENS];

    /* calculate synthetic delay values */
    calc_io_delays(file, file->counters[CP_POSIX_OPENS],
                   file->counters[CP_POSIX_READS] + file->counters[CP_POSIX_WRITES],
                   delay_per_open, &first_io_delay, &close_delay,
                   &inter_open_delay, &inter_io_delay);

    /* calculate average meta op time (for i/o and opens/closes) */
    /* TODO: this needs to be updated when we add in stat, seek, etc. */
    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]);

411
412
413
414
415
    /* set the create flag if the file was written to */
    if (file->counters[CP_BYTES_WRITTEN])
    {
        create_flag = 1;
    }
416
417
418
419
420
421

    /* 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 */
422
        cur_time = generate_psx_open_event(file, create_flag, meta_op_time, cur_time, io_context);
423
424
425
426
427
428
429
430
431
432
433
        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 */
        cur_time = generate_psx_ind_io_events(file, io_ops_this_cycle, i, inter_io_delay,
434
                                              meta_op_time, cur_time, io_context);
435
436
437
438
439

        /* 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 */
440
        cur_time = generate_psx_close_event(file, meta_op_time, cur_time, io_context);
441
442
443
444
445
446
447

        /* account for potential interopen delay if more than one open */
        if (file->counters[CP_POSIX_OPENS] > 1)
        {
            cur_time += inter_open_delay;
        }
    }
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721

    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)
{
    int64_t total_io_ops = file->counters[CP_POSIX_READS] + file->counters[CP_POSIX_WRITES];
    int64_t extra_opens = 0;
    int64_t extra_open_stride = 0;
    int64_t extra_io_ops = 0;
    int64_t ind_open_stride = 0;
    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;
    int64_t rank_cnt;
    int64_t aggregator_cnt;
    int create_flag = 0;
    double cur_time = file->fcounters[CP_F_OPEN_TIMESTAMP];
    double delay_per_cycle;
    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;
    int64_t i, j;

    /* 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);

    /* determine delay information */
    delay_per_cycle = (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)) /
                      (file->counters[CP_POSIX_OPENS] / nprocs);

    /* calculate average meta op time (for i/o and opens/closes) */
    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]);

    /* TODO TIMING */ 
    calc_io_delays(file, (file->counters[CP_POSIX_OPENS] / nprocs),
                   round((double)total_io_ops / nprocs), delay_per_cycle,
                   &first_io_delay, &close_delay, &inter_cycle_delay, &inter_io_delay);

    if (file->counters[CP_COLL_OPENS] && file->counters[CP_INDEP_OPENS])
    {
        if (((file->counters[CP_COLL_OPENS] / nprocs) % file->counters[CP_INDEP_OPENS]) == 0)
        {        
            ind_open_stride = (file->counters[CP_COLL_OPENS] / nprocs) /
                              file->counters[CP_INDEP_OPENS];
        }
        else
        {
            ind_open_stride = 1;
        }
    }

    if (file->counters[CP_COLL_OPENS] || file->counters[CP_INDEP_OPENS])
    {
        extra_opens = file->counters[CP_POSIX_OPENS] - file->counters[CP_COLL_OPENS] -
                      file->counters[CP_INDEP_OPENS];
        assert(extra_opens <= ((file->counters[CP_COLL_OPENS] / nprocs) +
               file->counters[CP_INDEP_OPENS]));
        if (extra_opens)
        {
            if ((((file->counters[CP_COLL_OPENS] / nprocs) + file->counters[CP_INDEP_OPENS]) %
                extra_opens) == 0)
            {
                extra_open_stride = ((file->counters[CP_COLL_OPENS] / nprocs) +
                                    file->counters[CP_INDEP_OPENS]) / extra_opens;
            }
            else
            {
                extra_open_stride = 1;
            }
            extra_io_ops = 0;
        }

        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;
    }
    else 
    {
        extra_opens = file->counters[CP_POSIX_OPENS] % nprocs;
        if (extra_opens && (((file->counters[CP_POSIX_OPENS] / nprocs) % extra_opens) == 0))
        {
            extra_open_stride = (file->counters[CP_POSIX_OPENS] / nprocs) / extra_opens;
            extra_io_ops = total_io_ops % nprocs;
        }
        else
        {
            extra_opens = 0;
        }

        total_ind_io_ops = total_io_ops - extra_io_ops;
        total_coll_io_ops = 0;
    }

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

    /* generate all events for this collectively opened file */
    for (i = 0; file->counters[CP_POSIX_OPENS]; i++)
    {
        if (file->counters[CP_COLL_OPENS])
            aggregator_cnt = in_agg_cnt;
        else
            aggregator_cnt = nprocs;

        /* assign any determined 'extra' opens to rank 0 at the beginning of the cycle */
        if (extra_opens && !(i % extra_open_stride))
        {
            assert(create_flag);

            /* generate the open/close events for creating the collective file */
            file->rank = 0;
            cur_time = generate_psx_open_event(file, create_flag, meta_op_time, cur_time, io_context);

            ind_io_ops_this_cycle = ceil((double)extra_io_ops / extra_opens);

            cur_time = generate_psx_coll_io_events(file, ind_io_ops_this_cycle, 0, nprocs,
                                                   aggregator_cnt, i, 0.0, meta_op_time,
                                                   cur_time, io_context);
            extra_io_ops -= ind_io_ops_this_cycle;

            cur_time = generate_psx_close_event(file, meta_op_time, cur_time, io_context);
            create_flag = 0;
            file->rank = -1;
            file->counters[CP_POSIX_OPENS]--;
            extra_opens--;
        }

        if (file->counters[CP_POSIX_OPENS] >= nprocs)
        {
            if (file->counters[CP_COLL_OPENS])
            {
                if (file->counters[CP_INDEP_OPENS] && !(i % ind_open_stride))
                {
                    rank_cnt = ceil((double)file->counters[CP_INDEP_OPENS] /
                                    (file->counters[CP_COLL_OPENS] / nprocs));
                    for (j = 0; j < rank_cnt; j++)
                    {
                        file->rank = j;
                        if (j != (rank_cnt - 1))
                            generate_psx_open_event(file, 0, meta_op_time, cur_time, io_context);
                        else
                            cur_time = generate_psx_open_event(file, 0, meta_op_time,
                                                               cur_time, io_context);
                    }

                    ind_io_ops_this_cycle = ceil(((double)total_ind_io_ops /
                                            file->counters[CP_INDEP_OPENS]) * rank_cnt);

                    cur_time = generate_psx_coll_io_events(file, ind_io_ops_this_cycle, 0,
                                                           nprocs, aggregator_cnt, i, 0.0,
                                                           meta_op_time, cur_time, io_context);
                    total_ind_io_ops -= ind_io_ops_this_cycle;

                    for (j = 0; j < rank_cnt; j++)
                    {
                        file->rank = j;
                        if (j != (rank_cnt - 1))
                            generate_psx_close_event(file, meta_op_time, cur_time, io_context);
                        else
                            cur_time = generate_psx_close_event(file, meta_op_time,
                                                                cur_time, io_context);
                    }
                    file->rank = -1;
                    file->counters[CP_INDEP_OPENS] -= rank_cnt;
                    file->counters[CP_POSIX_OPENS] -= rank_cnt;
                }

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

                coll_io_ops_this_cycle = ceil((double)total_coll_io_ops /
                                         (file->counters[CP_COLL_OPENS] / nprocs));
                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_POSIX_OPENS] / nprocs));
            }
            else
            {
                coll_io_ops_this_cycle = 0;
                ind_io_ops_this_cycle = ceil((double)total_ind_io_ops /
                                        (file->counters[CP_POSIX_OPENS] / nprocs));
            }

            /* perform an open across all ranks (rank == -1) */
            cur_time = generate_psx_open_event(file, create_flag, meta_op_time,
                                               cur_time, io_context);
            create_flag = 0;

            /* account for potential delay between the open and first i/o */
            cur_time += first_io_delay;

            /* generate the io events */
            cur_time = generate_psx_coll_io_events(file, ind_io_ops_this_cycle,
                                                   coll_io_ops_this_cycle, nprocs, aggregator_cnt,
                                                   i, inter_io_delay, meta_op_time,
                                                   cur_time, io_context);
            total_ind_io_ops -= ind_io_ops_this_cycle;
            total_coll_io_ops -= coll_io_ops_this_cycle;

            /* account for potential delay between last i/o operation and file close */
            cur_time += close_delay;
            
            /* generate the corresponding close event for the open at the start of the loop */
            cur_time = generate_psx_close_event(file, meta_op_time, cur_time, io_context);

            file->counters[CP_POSIX_OPENS] -= nprocs;
            if (file->counters[CP_COLL_OPENS])
                file->counters[CP_COLL_OPENS] -= nprocs;

            /* account for any delay between open-close cycles */
            if (file->counters[CP_POSIX_OPENS])
                cur_time += inter_cycle_delay;
        }
        else
        {
            /* open the file across participating ranks */
            rank_cnt = file->counters[CP_POSIX_OPENS];
            for (j = 0; j < rank_cnt; j++)
            {
                file->rank = j;
                if (j != (rank_cnt - 1))
                    generate_psx_open_event(file, 0, meta_op_time, cur_time, io_context);
                else
                    cur_time = generate_psx_open_event(file, 0, meta_op_time, cur_time, io_context);
            }

            /* account for potential delay between the open and first i/o */
            cur_time += first_io_delay;

            ind_io_ops_this_cycle = ceil((double)total_ind_io_ops /
                                         (file->counters[CP_POSIX_OPENS] / nprocs));

            cur_time = generate_psx_coll_io_events(file, ind_io_ops_this_cycle, 0, nprocs,
                                                   rank_cnt, i, inter_io_delay, meta_op_time,
                                                   cur_time, io_context);
            total_ind_io_ops -= ind_io_ops_this_cycle;

            /* account for potential delay between last i/o operation and file close */
            cur_time += close_delay;

            /* close the file across participating ranks */
            for (j = 0; j < rank_cnt; j++)
            {
                file->rank = j;
                if (j != (rank_cnt - 1))
                    generate_psx_close_event(file, meta_op_time, cur_time, io_context);
                else
                    cur_time = generate_psx_close_event(file, meta_op_time, cur_time, io_context);
            }
            file->rank = -1;
            file->counters[CP_POSIX_OPENS] -= rank_cnt;
        }
    }
722
723
724
725
726
727

    return;
}

/* fill in an open event structure and store it with the rank context */
static double generate_psx_open_event(
728
729
    struct darshan_file *file, int create_flag, double meta_op_time,
    double cur_time, struct rank_io_context *io_context)
730
{
731
732
733
734
735
736
737
    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
    };
738
739
740

    /* set the end time of the event based on time spent in POSIX meta operations */
    cur_time += meta_op_time;
741
    next_io_op.end_time = cur_time;
742

743
744
745
    /* store the open event (if this rank performed it) */
    if ((file->rank == io_context->my_rank) || (file->rank == -1))
        darshan_insert_next_io_op(io_context->io_op_dat, &next_io_op);
746
747
748
749
750
751

    return cur_time;
}

/* fill in a close event structure and store it with the rank context */
static double generate_psx_close_event(
752
753
    struct darshan_file *file, double meta_op_time, double cur_time,
    struct rank_io_context *io_context)
754
{
755
756
757
758
759
760
    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
    };
761
762
763

    /* set the end time of the event based on time spent in POSIX meta operations */
    cur_time += meta_op_time;
764
    next_io_op.end_time = cur_time;
765

766
767
768
    /* store the close event (if this rank performed it) */
    if ((file->rank == io_context->my_rank) || (file->rank == -1))
        darshan_insert_next_io_op(io_context->io_op_dat, &next_io_op);
769
770
771
772
773
774

    return cur_time;
}

/* fill in a barrier event structure and store it with the rank context */
static double generate_barrier_event(
775
    struct darshan_file *file, int64_t root, double cur_time, struct rank_io_context *io_context)
776
{
777
778
779
780
781
782
783
    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
    };
784
785

    cur_time += .000001; /* small synthetic delay representing time to barrier */
786
    next_io_op.end_time = cur_time;
787

788
789
790
    /* store the barrier event */
    if ((file->rank == -1) || (file->rank == io_context->my_rank))
        darshan_insert_next_io_op(io_context->io_op_dat, &next_io_op);
791
792
793
794
795
796
797

    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(
    struct darshan_file *file, int64_t io_ops_this_cycle, int64_t open_ndx,
798
    double inter_io_delay, double meta_op_time, double cur_time, struct rank_io_context *io_context)
799
800
801
802
803
804
805
806
807
{
    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;
808
    struct darshan_io_op next_io_op;
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
844
845
846
847
848

    /* 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 */
        determine_io_params(file, rw, 0, file->counters[CP_POSIX_OPENS], &io_sz, &io_off);
        if (!rw)
        {
            /* generate a read event */
849
850
851
852
853
            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;
854
855
856
857
858

            /* set the end time based on observed bandwidth and io size */
            if (rd_bw == 0.0)
                io_op_time = 0.0;
            else
859
                io_op_time = (io_sz / rd_bw);
860
861
862

            /* update time, accounting for metadata time */
            cur_time += (io_op_time + meta_op_time);
863
            next_io_op.end_time = cur_time;
864
865
866
867
868
            file->counters[CP_POSIX_READS]--;
        }
        else
        {
            /* generate a write event */
869
870
871
872
873
            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;
874
875
876
877
878

            /* set the end time based on observed bandwidth and io size */
            if (wr_bw == 0.0)
                io_op_time = 0.0;
            else
879
                io_op_time = (io_sz / wr_bw);
880
881
882

            /* update time, accounting for metadata time */
            cur_time += (io_op_time + meta_op_time);
883
            next_io_op.end_time = cur_time;
884
885
886
887
888
889
890
            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);

891
892
        /* store the i/o event */
        darshan_insert_next_io_op(io_context->io_op_dat, &next_io_op);
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912

        /* 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 */
913
            cur_time += inter_io_delay;
914
915
916
917
918
919
920
921
922
923
924
925
        }
    }

    /* 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;
}

926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
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, int64_t aggregator_cnt, int64_t open_ndx, double inter_io_delay,
    double meta_op_time, double cur_time, struct rank_io_context *io_context)
{
    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;
    int64_t ind_ops_so_far = 0;
    double ind_coll_switch;
    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;

        /* 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);
        }

        /* 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];
    }

    for (i = 0; i < total_io_ops_this_cycle; i++)
    {
        ind_coll_switch = (double)ind_io_ops_this_cycle / (total_io_ops_this_cycle - i);
        if (((double)rand() / (double)(RAND_MAX + 1.0)) < ind_coll_switch)
        {
For faster browsing, not all history is shown. View entire blame