codes-workload-mpi-replay.c 18.2 KB
Newer Older
1
2
3
4
5
6
7
8
/*
 * Copyright (C) 2013 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
 */

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

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

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

27
28
#define DEBUG_PROFILING 1

29
/* hash table entry for looking up file descriptor of a workload file id */
30
struct file_info
31
32
33
34
35
36
37
38
39
40
41
42
{
    struct qlist_head hash_link;
    uint64_t file_hash;
    int file_descriptor;
};

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

/* command line options */
static int opt_verbose = 0;
static int opt_noop = 0;
43
static int opt_lockless = 0;
44
static double opt_delay_pct = 1.0;
45
46
47
48

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

49
50
51
/* file stream to log rank events to, if verbose turned on */
static FILE *log_stream = NULL;

52
/* global variables for profiling different portions of the replay, if enabled */
53
#if DEBUG_PROFILING
54
55
56
57
58
59
static double total_open_time = 0.0;
static double total_close_time = 0.0;
static double total_read_time = 0.0;
static double total_write_time = 0.0;
static double total_delay_time = 0.0;
static double total_barrier_time = 0.0;
60
#endif
61

62
63
void usage(char *exename)
{
64
65
    fprintf(stderr, "Usage: %s [OPTIONS] --conf <conf_file_path>\n       "
            "--test-dir <workload_test_dir>\n\n", exename);
66
    fprintf(stderr, "\t<conf_file_path> : (absolute) path to a valid workload configuration file\n");
67
68
    fprintf(stderr, "\t<workload_test_dir> : the directory to replay the workload I/O in\n");
    fprintf(stderr, "\n\t[OPTIONS] includes:\n");
69
    fprintf(stderr, "\t\t--noop : do not perform i/o\n");
70
    fprintf(stderr, "\t\t--lockless : use bgq lockless file i/o optimization\n");
71
    fprintf(stderr, "\t\t    -v : verbose (output i/o details)\n");
72
73
74
75

    exit(1);
}

76
void parse_args(int argc, char **argv, char **conf_path, char **test_dir)
77
78
79
80
{
    int index;
    static struct option long_opts[] =
    {
81
        {"conf", 1, NULL, 'c'},
82
        {"test-dir", 1, NULL, 'd'},
83
        {"noop", 0, NULL, 'n'},
84
        {"lockless", 0, NULL, 'l'},
85
        {"delay", 1, NULL, 'p'},
86
87
88
89
        {"help", 0, NULL, 0},
        {0, 0, 0, 0}
    };

90
    *conf_path = NULL;
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
    *test_dir = NULL;
    while (1)
    {
        int c = getopt_long(argc, argv, "v", long_opts, &index);

        if (c == -1)
            break;

        switch (c)
        {
            case 'v':
                opt_verbose = 1;
                break;
            case 'n':
                opt_noop = 1;
                break;
107
108
109
            case 'l':
                opt_lockless = 1;
                break;
110
111
            case 'c':
                *conf_path = optarg;
112
113
114
115
                break;
            case 'd':
                *test_dir = optarg;
                break;
116
117
118
            case 'p':
                opt_delay_pct = atof(optarg);
                break;
119
120
121
122
123
124
125
126
            case 0:
            case '?':
            default:
                usage(argv[0]);
                break;
        }
    }

127
    if (optind < argc || !(*conf_path) || !(*test_dir))
128
129
130
131
132
133
134
    {
        usage(argv[0]);
    }

    return;
}

135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
int load_workload(char *conf_path, int rank)
{
    config_lpgroups_t paramconf;
    char workload_type[MAX_NAME_LENGTH_WKLD];

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

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

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

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

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

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

        /* get the bgp i/o params from the config file */
        configuration_get_value(&config, "PARAMS", "io_kernel_meta_path",
                                b_params.io_kernel_meta_path, MAX_NAME_LENGTH_WKLD);
172
        configuration_get_value(&config, "PARAMS", "bgp_config_file",
173
174
175
176
177
178
179
180
                                b_params.bgp_config_file, MAX_NAME_LENGTH_WKLD);
        configuration_get_value(&config, "PARAMS", "rank_count", rank_count, 10);
        strcpy(b_params.io_kernel_path, "");
        strcpy(b_params.io_kernel_def_path, "");
        b_params.num_cns = atoi(rank_count);

        return codes_workload_load(workload_type, (char *)&b_params, rank);
    }
181
182
183
    else if (strcmp(workload_type, "recorder_io_workload") == 0) {
        struct recorder_params r_params;

184
        /* get the recorder params from the config file */
185
186
187
188
189
190
        configuration_get_value(&config, "PARAMS", "trace_dir_path",
                                r_params.trace_dir_path, MAX_NAME_LENGTH_WKLD);

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

	}
191
192
193
194
195
196
197
    else
    {
        fprintf(stderr, "Error: Invalid workload type specified (%s)\n", workload_type);
        return -1;
    }
}

198
199
int main(int argc, char *argv[])
{
200
    char *conf_path;
201
    char *replay_test_path;
202
203
    char *log_dir = "log";
    char my_log_path[MAX_NAME_LENGTH_WKLD];
204
205
206
207
208
    int nprocs;
    int myrank;
    int workload_id;
    struct codes_workload_op next_op;
    long long int replay_op_number = 1;
209
210
    double load_start, load_end;
    int first_op = 1;
211
212
213
    int ret = 0;

    /* parse command line args */
214
    parse_args(argc, argv, &conf_path, &replay_test_path);
215
216
217
218
219
220

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

221
222
223
    /* start workload load timer */
    load_start = MPI_Wtime();

224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
    /* change the working directory to be the test directory */
    ret = chdir(replay_test_path);
    if (ret < 0)
    {
        fprintf(stderr, "Unable to change to testing directory (%s)\n", strerror(errno));
        goto error_exit;
    }

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

245
246
    /* initialize workload generator from config file */
    workload_id = load_workload(conf_path, myrank);
247
248
249
250
251
252
253
254
255
256
257
258
259
    if (workload_id < 0)
    {
        goto error_exit;
    }

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

260
261
262
    /* synchronize before replay */
    MPI_Barrier(MPI_COMM_WORLD);

263
264
265
    /* loading is finished */
    load_end = MPI_Wtime();

266
267
268
269
270
271
    /* replay loop */
    while (1)
    {
        /* get the next replay operation from the workload generator */
        codes_workload_get_next(workload_id, myrank, &next_op);

272
273
274
        /* subtract the load time from any initial delay to be fair */
        if (first_op && (next_op.op_type == CODES_WK_DELAY))
        {
275
            first_op = 0;
276
            if (next_op.u.delay.seconds > (load_end - load_start))
277
            {
278
                next_op.u.delay.seconds -= (load_end - load_start);
279
            }
280
            else
281
            {
282
                continue;
283
            }
284
285
        }

286
287
        if (next_op.op_type != CODES_WK_END)
        {
288
289

            if (next_op.op_type == CODES_WK_DELAY)
290
                next_op.u.delay.seconds *= opt_delay_pct;
291

292
293
294
295
296
297
298
299
300
301
302
303
304
305
            /* replay the next workload operation */
            ret = replay_workload_op(next_op, myrank, replay_op_number++);
            if (ret < 0)
            {
                break;
            }
        }
        else
        {
            /* workload replay for this rank is complete */
            break;
        }
    }

306
    if (log_stream)
307
308
        fclose(log_stream);

309
    /* destroy and finalize the file descriptor hash table */
310
    qhash_destroy_and_finalize(fd_table, struct file_info, hash_link, free);
311

312
313
314
315
316
317
#if DEBUG_PROFILING
    printf("Rank %d:\td=%.4lf, b=%.4lf, o=%.4lf, c=%.4lf, r=%.4lf, w=%.4lf\n",
           myrank, total_delay_time, total_barrier_time, total_open_time,
           total_close_time, total_read_time, total_write_time);
#endif

318
319
error_exit:
    MPI_Finalize();
320

321
322
323
324
325
326
327
328
    return ret;
}

int replay_workload_op(struct codes_workload_op replay_op, int rank, long long int op_number)
{
    unsigned int secs;
    unsigned int usecs;
    int open_flags = O_RDWR;
329
    char file_name[250];
330
    int fildes;
331
    struct file_info *tmp_list = NULL;
332
333
334
    struct qlist_head *hash_link = NULL;
    char *buf = NULL;
    int ret;
335
336

#if DEBUG_PROFILING
337
    double start, end;
338
339
    start = MPI_Wtime();
#endif
340
341
342
343
344

    switch (replay_op.op_type)
    {
        case CODES_WK_DELAY:
            if (opt_verbose)
345
                fprintf(log_stream, "[Rank %d] Operation %lld : DELAY %lf seconds\n",
346
347
348
349
350
                       rank, op_number, replay_op.u.delay.seconds);

            if (!opt_noop)
            {
                /* satisfy delay using second delay then microsecond delay */
351
352
                secs = (unsigned int)replay_op.u.delay.seconds;
                usecs = (unsigned int)((replay_op.u.delay.seconds - secs) * 1000 * 1000);
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
                ret = sleep(secs);
                if (ret)
                {
                    /* error in sleep */
                    errno = EINTR;
                    fprintf(stderr, "Rank %d failure on operation %lld [DELAY: %s]\n",
                            rank, op_number, strerror(errno));
                    return -1;
                }
                ret = usleep(usecs);
                if (ret < 0)
                {
                    /* error in usleep */
                    fprintf(stderr, "Rank %d failure on operation %lld [DELAY: %s]\n",
                            rank, op_number, strerror(errno));
                    return -1;
                }
370
371
372
373
#if DEBUG_PROFILING
                end = MPI_Wtime();
                total_delay_time += (end - start);
#endif
374
375
376
377
            }
            return 0;
        case CODES_WK_BARRIER:
            if (opt_verbose)
378
                fprintf(log_stream, "[Rank %d] Operation %lld : BARRIER\n", rank, op_number);
379
380
381
382
383
384
385
386
387
388
389
390

            if (!opt_noop)
            {
                /* implement barrier using MPI global barrier on all ranks */
                ret = MPI_Barrier(MPI_COMM_WORLD);
                if (ret != MPI_SUCCESS)
                {
                    /* error in MPI_Barrier */
                    fprintf(stderr, "Rank %d failure on operation %lld [BARRIER: %s]\n",
                            rank, op_number, "Invalid communicator");
                    return -1;
                }
391
392
393
394
395

#if DEBUG_PROFILING
                end = MPI_Wtime();
                total_barrier_time += (end - start);
#endif
396
397
398
399
            }
            return 0;
        case CODES_WK_OPEN:
            if (opt_verbose)
400
                fprintf(log_stream, "[Rank %d] Operation %lld: %s file %"PRIu64"\n", rank, op_number,
401
402
403
404
405
406
407
408
409
                       (replay_op.u.open.create_flag) ? "CREATE" : "OPEN", replay_op.u.open.file_id);

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

                /* write the file hash to string to be used as the actual file name */
410
411
412
413
414
                if (!opt_lockless)
                    snprintf(file_name, sizeof(file_name), "%"PRIu64, replay_op.u.open.file_id);
                else
                    snprintf(file_name, sizeof(file_name), "bglockless:%"PRIu64,
                             replay_op.u.open.file_id);
415
416
417
418
419
420
421
422
423
424
425
426

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

                /* save the file descriptor for this file in a hash table to be retrieved later */
427
                tmp_list = malloc(sizeof(struct file_info));
428
429
430
431
432
433
434
435
436
                if (!tmp_list)
                {
                    fprintf(stderr, "No memory available for file hash entry\n");
                    return -1;
                }

                tmp_list->file_hash = replay_op.u.open.file_id;
                tmp_list->file_descriptor = fildes;
                qhash_add(fd_table, &(replay_op.u.open.file_id), &(tmp_list->hash_link));
437
438
439
440
441

#if DEBUG_PROFILING
                end = MPI_Wtime();
                total_open_time += (end - start);
#endif
442
443
444
445
            }
            return 0;
        case CODES_WK_CLOSE:
            if (opt_verbose)
446
447
                fprintf(log_stream, "[Rank %d] Operation %lld : CLOSE file %"PRIu64"\n",
                        rank, op_number, replay_op.u.close.file_id);
448
449
450
451
452
453

            if (!opt_noop)
            {
                /* search for the corresponding file descriptor in the hash table */
                hash_link = qhash_search_and_remove(fd_table, &(replay_op.u.close.file_id));
                assert(hash_link);
454
                tmp_list = qhash_entry(hash_link, struct file_info, hash_link);
455
456
                fildes = tmp_list->file_descriptor;
                free(tmp_list);
457

458
459
460
461
462
463
464
465
                /* perform the close operation */
                ret = close(fildes);
                if (ret < 0)
                {
                    fprintf(stderr, "Rank %d failure on operation %lld [CLOSE: %s]\n",
                            rank, op_number, strerror(errno));
                    return -1;
                }
466
467
468
469
470

#if DEBUG_PROFILING
                end = MPI_Wtime();
                total_close_time += (end - start);
#endif
471
472
473
474
            }
            return 0;
        case CODES_WK_WRITE:
            if (opt_verbose)
475
                fprintf(log_stream, "[Rank %d] Operation %lld : WRITE file %"PRIu64" (sz = %"PRId64
476
                       ", off = %"PRId64")\n",
477
                       rank, op_number, replay_op.u.write.file_id, replay_op.u.write.size,
478
479
480
481
482
483
484
                       replay_op.u.write.offset);

            if (!opt_noop)
            {
                /* search for the corresponding file descriptor in the hash table */
                hash_link = qhash_search(fd_table, &(replay_op.u.write.file_id));
                assert(hash_link);
485
                tmp_list = qhash_entry(hash_link, struct file_info, hash_link);
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
                fildes = tmp_list->file_descriptor;

                /* perform the write operation */
                buf = malloc(replay_op.u.write.size);
                if (!buf)
                {
                    fprintf(stderr, "No memory available for write buffer\n");
                    return -1;
                }
                ret = pwrite(fildes, buf, replay_op.u.write.size, replay_op.u.write.offset);
                free(buf);
                if (ret < 0)
                {
                    fprintf(stderr, "Rank %d failure on operation %lld [WRITE: %s]\n",
                            rank, op_number, strerror(errno));
                    return -1;
                }
503
504
505
506
507

#if DEBUG_PROFILING
                end = MPI_Wtime();
                total_write_time += (end - start);
#endif
508
509
510
511
            }
            return 0;
        case CODES_WK_READ:
            if (opt_verbose)
512
513
                fprintf(log_stream, "[Rank %d] Operation %lld : READ file %"PRIu64" (sz = %"PRId64
                        ", off = %"PRId64")\n", rank, op_number, replay_op.u.read.file_id,
514
                       replay_op.u.read.size, replay_op.u.read.offset);
515

516
517
518
519
520
            if (!opt_noop)
            {
                /* search for the corresponding file descriptor in the hash table */
                hash_link = qhash_search(fd_table, &(replay_op.u.read.file_id));
                assert(hash_link);
521
                tmp_list = qhash_entry(hash_link, struct file_info, hash_link);
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
                fildes = tmp_list->file_descriptor;

                /* perform the write operation */
                buf = malloc(replay_op.u.read.size);
                if (!buf)
                {
                    fprintf(stderr, "No memory available for write buffer\n");
                    return -1;
                }
                ret = pread(fildes, buf, replay_op.u.read.size, replay_op.u.read.offset);
                free(buf);
                if (ret < 0)
                {
                    fprintf(stderr, "Rank %d failure on operation %lld [READ: %s]\n",
                            rank, op_number, strerror(errno));
                    return -1;
                }
539
540
541
542
543

#if DEBUG_PROFILING
                end = MPI_Wtime();
                total_read_time += (end - start);
#endif
544
545
546
547
            }
            return 0;
        default:
            fprintf(stderr, "** Rank %d: INVALID OPERATION (op count = %lld) **\n", rank, op_number);
548
            return 0;
549
550
551
552
553
554
555
    }

}

int hash_file_compare(void *key, struct qlist_head *link)
{
    uint64_t *in_file_hash = (uint64_t *)key;
556
    struct file_info *tmp_file;
557

558
    tmp_file = qlist_entry(link, struct file_info, hash_link);
559
560
561
562
563
    if (tmp_file->file_hash == *in_file_hash)
        return 1;

    return 0;
}
564
565
566
567
568
569
570
571
572
573

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