model-net-mpi-replay.c 90.7 KB
Newer Older
1 2 3 4 5 6 7
/*
 * Copyright (C) 2014 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
 */
#include <ross.h>
#include <inttypes.h>
8
#include <sys/stat.h>
9 10 11 12 13 14 15 16

#include "codes/codes-workload.h"
#include "codes/codes.h"
#include "codes/configuration.h"
#include "codes/codes_mapping.h"
#include "codes/model-net.h"
#include "codes/rc-stack.h"
#include "codes/quicklist.h"
17
#include "codes/quickhash.h"
18
#include "codes/codes-jobmap.h"
19

20
/* turning on track lp will generate a lot of output messages */
21
#define MN_LP_NM "modelnet_dragonfly_custom"
22
#define CONTROL_MSG_SZ 64
23
#define TRACE -1
24
#define MAX_WAIT_REQS 512
25
#define CS_LP_DBG 1
26 27
#define RANK_HASH_TABLE_SZ 2000
#define NW_LP_NM "nw-lp"
28 29 30
#define lprintf(_fmt, ...) \
        do {if (CS_LP_DBG) printf(_fmt, __VA_ARGS__);} while (0)
#define MAX_STATS 65536
31 32
//#define PAYLOAD_SZ 8
static int PAYLOAD_SZ = 8;
33 34 35
static int msg_size_hash_compare(
            void *key, struct qhash_head *link);

Noah Wolfe's avatar
Noah Wolfe committed
36 37 38
int total_count_irecv = 0, total_count_isend = 0;
int total_irecvs = 0, total_isends = 0;

39
/* NOTE: Message tracking works in sequential mode only! */
40
static int debug_cols = 0;
41 42 43 44 45
/* Turning on this option slows down optimistic mode substantially. Only turn
 * on if you get issues with wait-all completion with traces. */
static int preserve_wait_ordering = 0;
static int enable_msg_tracking = 0;
static int is_synthetic = 0;
46
tw_lpid TRACK_LP = -1;
47

48
static double total_syn_data = 0;
49
static int unmatched = 0;
50 51 52 53 54
char workload_type[128];
char workload_file[8192];
char offset_file[8192];
static int wrkld_id;
static int num_net_traces = 0;
55
static int num_dumpi_traces = 0;
56
static int64_t EAGER_THRESHOLD = 8192;
57

58
static int alloc_spec = 0;
59 60
static tw_stime self_overhead = 10.0;
static tw_stime mean_interval = 100000;
61 62 63

/* Doing LP IO*/
static char lp_io_dir[256] = {'\0'};
64
static char sampling_dir[32] = {'\0'};
65 66 67 68
static lp_io_handle io_handle;
static unsigned int lp_io_use_suffix = 0;
static int do_lp_io = 0;

69 70 71 72
/* variables for loading multiple applications */
char workloads_conf_file[8192];
char alloc_file[8192];
int num_traces_of_job[5];
73 74 75
tw_stime soft_delay_mpi = 2500;
tw_stime nic_delay = 1000;
tw_stime copy_per_byte_eager = 0.55;
76 77 78 79 80
char file_name_of_job[5][8192];

struct codes_jobmap_ctx *jobmap_ctx;
struct codes_jobmap_params_list jobmap_p;

81 82
/* Variables for Cortex Support */
/* Matthieu's additions start */
83
#ifdef ENABLE_CORTEX_PYTHON
84 85 86
static char cortex_file[512] = "\0";
static char cortex_class[512] = "\0";
static char cortex_gen[512] = "\0";
87
#endif
88 89
/* Matthieu's additions end */

90 91
typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
92
typedef int dumpi_req_id;
93 94

static int net_id = 0;
95
static float noise = 2.0;
96 97 98
static int num_nw_lps = 0, num_mpi_lps = 0;

static int num_syn_clients;
99

100
FILE * workload_log = NULL;
101
FILE * msg_size_log = NULL;
102 103 104 105
FILE * workload_agg_log = NULL;
FILE * workload_meta_log = NULL;

static uint64_t sample_bytes_written = 0;
106

107 108 109
long long num_bytes_sent=0;
long long num_bytes_recvd=0;

110 111 112
long long num_syn_bytes_sent = 0;
long long num_syn_bytes_recvd = 0;

113
static unsigned long msgs_per_tick = 0;
Noah Wolfe's avatar
Noah Wolfe committed
114 115
static double tick_interval = 0;

116 117 118 119 120 121
double max_time = 0,  max_comm_time = 0, max_wait_time = 0, max_send_time = 0, max_recv_time = 0;
double avg_time = 0, avg_comm_time = 0, avg_wait_time = 0, avg_send_time = 0, avg_recv_time = 0;


/* runtime option for disabling computation time simulation */
static int disable_delay = 0;
122 123 124
static int enable_sampling = 0;
static double sampling_interval = 5000000;
static double sampling_end_time = 3000000000;
125
static int enable_debug = 0;
126

127 128 129
/* set group context */
struct codes_mctx group_ratio;

130
/* MPI_OP_GET_NEXT is for getting next MPI operation when the previous operation completes.
131
* MPI_SEND_ARRIVED is issued when a MPI message arrives at its destination (the message is transported by model-net and an event is invoked when it arrives.
132 133 134 135 136 137 138
* MPI_SEND_POSTED is issued when a MPI message has left the source LP (message is transported via model-net). */
enum MPI_NW_EVENTS
{
	MPI_OP_GET_NEXT=1,
	MPI_SEND_ARRIVED,
    MPI_SEND_ARRIVED_CB, // for tracking message times on sender
	MPI_SEND_POSTED,
139 140 141 142 143 144
    MPI_REND_ARRIVED,
    MPI_REND_ACK_ARRIVED,
    CLI_BCKGND_FIN,
    CLI_BCKGND_ARRIVE,
    CLI_BCKGND_GEN,
    CLI_NBR_FINISH,
145 146
};

147 148 149 150
struct mpi_workload_sample
{
    /* Sampling data */
    int nw_id;
151
    int app_id;
152
    unsigned long num_sends_sample;
153 154 155 156
    unsigned long num_recvs_sample;
    unsigned long num_cols_sample;
    unsigned long num_delays_sample;
    unsigned long num_wait_all_sample;
157
    unsigned long num_waits_sample;
Noah Wolfe's avatar
Noah Wolfe committed
158 159
    unsigned long num_bytes_sent_sample;
    unsigned long num_bytes_recvd_sample;
160 161 162 163 164 165
    unsigned long min_bytes_sample;
    unsigned long max_bytes_sample;
    double send_time_sample;
    double wait_time_sample;
    double comm_time_sample;
    double comp_time_sample;
166 167
    double sample_end_time;
};
168 169 170 171 172 173 174
/* stores pointers of pending MPI operations to be matched with their respective sends/receives. */
struct mpi_msgs_queue
{
    int op_type;
    int tag;
    int source_rank;
    int dest_rank;
175
    int64_t num_bytes;
176 177 178 179 180 181 182 183
    tw_stime req_init_time;
	dumpi_req_id req_id;
    struct qlist_head ql;
};

/* stores request IDs of completed MPI operations (Isends or Irecvs) */
struct completed_requests
{
184
	int req_id;
185
    struct qlist_head ql;
186
    int index;
187 188 189 190 191 192
};

/* for wait operations, store the pending operation and number of completed waits so far. */
struct pending_waits
{
    int op_type;
193
    int req_ids[MAX_WAIT_REQS];
194
	int num_completed;
195 196
	int count;
    tw_stime start_time;
197 198 199
    struct qlist_head ql;
};

200 201 202 203 204 205 206 207 208
struct msg_size_info
{
    int64_t msg_size;
    int num_msgs;
    tw_stime agg_latency;
    tw_stime avg_latency;
    struct qhash_head * hash_link;
    struct qlist_head ql; 
};
209 210 211 212 213 214 215 216 217 218
typedef struct mpi_msgs_queue mpi_msgs_queue;
typedef struct completed_requests completed_requests;
typedef struct pending_waits pending_waits;

/* state of the network LP. It contains the pointers to send/receive lists */
struct nw_state
{
	long num_events_per_lp;
	tw_lpid nw_id;
	short wrkld_end;
219 220
    int app_id;
    int local_rank;
221

222 223 224
    int is_finished;
    int neighbor_completed;

225
    struct rc_stack * processed_ops;
226
    struct rc_stack * processed_wait_op;
227
    struct rc_stack * matched_reqs;
228
//    struct rc_stack * indices;
229 230 231 232 233 234 235 236 237 238 239 240

    /* count of sends, receives, collectives and delays */
	unsigned long num_sends;
	unsigned long num_recvs;
	unsigned long num_cols;
	unsigned long num_delays;
	unsigned long num_wait;
	unsigned long num_waitall;
	unsigned long num_waitsome;

	/* time spent by the LP in executing the app trace*/
	double start_time;
241 242 243 244 245 246 247 248 249

    double col_time;

    double reduce_time;
    int num_reduce;

    double all_reduce_time;
    int num_all_reduce;

250 251 252 253 254 255 256 257 258 259 260 261 262 263 264
	double elapsed_time;
	/* time spent in compute operations */
	double compute_time;
	/* time spent in message send/isend */
	double send_time;
	/* time spent in message receive */
	double recv_time;
	/* time spent in wait operation */
	double wait_time;
	/* FIFO for isend messages arrived on destination */
	struct qlist_head arrival_queue;
	/* FIFO for irecv messages posted but not yet matched with send operations */
	struct qlist_head pending_recvs_queue;
	/* List of completed send/receive requests */
	struct qlist_head completed_reqs;
265

266
    tw_stime cur_interval_end;
267
    
268 269
    /* Pending wait operation */
    struct pending_waits * wait_op;
270

271 272 273 274 275 276
    /* Message size latency information */
    struct qhash_table * msg_sz_table;
    struct qlist_head msg_sz_list;

    /* quick hash for maintaining message latencies */

277 278 279
    unsigned long num_bytes_sent;
    unsigned long num_bytes_recvd;

Noah Wolfe's avatar
Noah Wolfe committed
280 281
    unsigned long msgs_per_tick_sent;

282 283 284 285
    /* For sampling data */
    int sampling_indx;
    int max_arr_size;
    struct mpi_workload_sample * mpi_wkld_samples;
286
    char output_buf[512];
287
    char output_buf2[16384];
288
    char col_stats[64];
289 290 291 292
};

/* data for handling reverse computation.
* saved_matched_req holds the request ID of matched receives/sends for wait operations.
293
* ptr_match_op holds the matched MPI operation which are removed from the queues when a send is matched with the receive in forward event handler.
294 295 296
* network event being sent. op is the MPI operation issued by the network workloads API. rv_data holds the data for reverse computation (TODO: Fill this data structure only when the simulation runs in optimistic mode). */
struct nw_message
{
297
   // forward message handler
298
   int msg_type;
299
   int op_type;
300
   model_net_event_return event_rc;
301

302 303 304
   struct
   {
       tw_lpid src_rank;
305
       int dest_rank;
306
       int64_t num_bytes;
307 308 309 310 311
       int num_matched;
       int data_type;
       double sim_start_time;
       // for callbacks - time message was received
       double msg_send_time;
312
       int req_id;
313
       int tag;
314
       int app_id;
315 316 317 318 319 320 321 322 323
       int found_match;
       short wait_completed;
   } fwd;
   struct
   {
       double saved_send_time;
       double saved_recv_time;
       double saved_wait_time;
       double saved_delay;
324
       int64_t saved_num_bytes;
325 326
       int64_t saved_min_bytes;
       int64_t saved_max_bytes;
Noah Wolfe's avatar
Noah Wolfe committed
327
       int64_t saved_msgs_per_tick_sent;
328
   } rc;
329 330
};

331
static void send_ack_back(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, mpi_msgs_queue * mpi_op);
332 333

static void send_ack_back_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
334 335
/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
336
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op, int is_rend);
337 338
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
339
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
340 341
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
342
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp);
343 344
/* execute the computational delay */
static void codes_exec_comp_delay(
345
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op, tw_bf * bf);
346 347 348 349 350 351 352 353 354 355 356 357 358 359 360
/* gets the next MPI operation from the network-workloads API. */
static void get_next_mpi_operation(
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
/* reverse handler of get next mpi operation. */
static void get_next_mpi_operation_rc(
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
/* Makes a call to get_next_mpi_operation. */
static void codes_issue_next_event(tw_lp* lp);
/* reverse handler of next operation */
static void codes_issue_next_event_rc(tw_lp* lp);


///////////////////// HELPER FUNCTIONS FOR MPI MESSAGE QUEUE HANDLING ///////////////
/* upon arrival of local completion message, inserts operation in completed send queue */
/* upon arrival of an isend operation, updates the arrival queue of the network */
361 362 363 364 365 366 367 368
static void update_completed_queue(
        nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp, dumpi_req_id req_id);
/* reverse of the above function */
static void update_completed_queue_rc(
        nw_state*s,
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp);
369 370 371 372 373 374 375 376 377 378 379 380 381 382 383
static void update_arrival_queue(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
/* reverse of the above function */
static void update_arrival_queue_rc(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
/* callback to a message sender for computing message time */
static void update_message_time(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
/* reverse for computing message time */
static void update_message_time_rc(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);

/* conversion from seconds to eanaoseconds */
static tw_stime s_to_ns(tw_stime ns);

384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420
/* Helper functions for Sampling */
void increment_sampling_check_rc(nw_state* s, tw_bf* bf);
void increment_sampling_check(nw_state* s, tw_lp* lp, tw_bf* bf);

// Sampling Function: Always to be called after undoing the increment
void increment_sampling_check_rc(nw_state* s, tw_bf* bf)
{
    if(bf->c13)
    {
        s->sampling_indx--;
        s->cur_interval_end -= sampling_interval;
    }
}

// Sampling Function: Always to be called before incrementing the metric
void increment_sampling_check(nw_state* s, tw_lp* lp, tw_bf* bf)
{
    if(tw_now(lp) >= s->cur_interval_end)
    {
        bf->c13 = 1;
        int indx = s->sampling_indx;
        s->mpi_wkld_samples[indx].nw_id = s->nw_id;
        s->mpi_wkld_samples[indx].app_id = s->app_id;
        s->mpi_wkld_samples[indx].sample_end_time = s->cur_interval_end;
        s->sampling_indx++;
        s->cur_interval_end += sampling_interval;
    }
    if(s->sampling_indx >= MAX_STATS)
    {
        struct mpi_workload_sample * tmp = (struct mpi_workload_sample*)calloc((MAX_STATS + s->max_arr_size), sizeof(struct mpi_workload_sample));
        memcpy(tmp, s->mpi_wkld_samples, s->sampling_indx);
        free(s->mpi_wkld_samples);
        s->mpi_wkld_samples = tmp;
        s->max_arr_size += MAX_STATS;
    }
}

421
/*static void update_message_size_rc(
422 423 424 425
        struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
426
{*/
427
/*TODO: Complete reverse handler */
428
/*    (void)ns;
429 430 431
    (void)lp;
    (void)bf;
    (void)m;
432
}*/
433 434 435 436 437 438 439 440 441 442
/* update the message size */
static void update_message_size(
        struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m,
        mpi_msgs_queue * qitem,
        int is_eager,
        int is_send)
{
443 444 445
            (void)bf;
            (void)is_eager;

446 447 448 449 450 451 452 453 454 455 456 457 458 459
            struct qhash_head * hash_link = NULL;
            tw_stime msg_init_time = qitem->req_init_time;
        
            if(!ns->msg_sz_table)
                ns->msg_sz_table = qhash_init(msg_size_hash_compare, quickhash_64bit_hash, RANK_HASH_TABLE_SZ); 
            
            hash_link = qhash_search(ns->msg_sz_table, &(qitem->num_bytes));

            if(is_send)
                msg_init_time = m->fwd.sim_start_time;
            
            /* update hash table */
            if(!hash_link)
            {
460
                struct msg_size_info * msg_info = (struct msg_size_info*)malloc(sizeof(struct msg_size_info));
461 462
                msg_info->msg_size = qitem->num_bytes;
                msg_info->num_msgs = 1;
463
                msg_info->agg_latency = tw_now(lp) - msg_init_time;
464
                msg_info->avg_latency = msg_info->agg_latency;
465
                qhash_add(ns->msg_sz_table, &(msg_info->msg_size), msg_info->hash_link);
466 467 468 469 470 471 472
                qlist_add(&msg_info->ql, &ns->msg_sz_list);
                //printf("\n Msg size %d aggregate latency %f num messages %d ", m->fwd.num_bytes, msg_info->agg_latency, msg_info->num_msgs);
            }
            else
            {
                struct msg_size_info * tmp = qhash_entry(hash_link, struct msg_size_info, hash_link);
                tmp->num_msgs++;
473
                tmp->agg_latency += tw_now(lp) - msg_init_time;  
474 475 476 477 478 479 480 481 482 483
                tmp->avg_latency = (tmp->agg_latency / tmp->num_msgs);
//                printf("\n Msg size %d aggregate latency %f num messages %d ", qitem->num_bytes, tmp->agg_latency, tmp->num_msgs);
            }
}
static void notify_background_traffic_rc(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
484 485 486
    (void)ns;
    (void)bf;
    (void)m;
487 488 489 490 491
        
    int num_jobs = codes_jobmap_get_num_jobs(jobmap_ctx); 
    
    for(int i = 0; i < num_jobs - 1; i++)
        tw_rand_reverse_unif(lp->rng); 
492 493 494 495 496 497 498 499
}

static void notify_background_traffic(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
500 501 502
        (void)bf;
        (void)m;

503 504 505 506 507 508 509 510 511 512 513 514 515 516 517
        struct codes_jobmap_id jid; 
        jid = codes_jobmap_to_local_id(ns->nw_id, jobmap_ctx);
        
        int num_jobs = codes_jobmap_get_num_jobs(jobmap_ctx); 
        
        for(int other_id = 0; other_id < num_jobs; other_id++)
        {
            if(other_id == jid.job)
                continue;

            struct codes_jobmap_id other_jid;
            other_jid.job = other_id;

            int num_other_ranks = codes_jobmap_get_num_ranks(other_id, jobmap_ctx);

518
            lprintf("\n Other ranks %d ", num_other_ranks);
519 520 521 522 523 524 525 526 527 528 529 530
            tw_stime ts = (1.1 * g_tw_lookahead) + tw_rand_exponential(lp->rng, mean_interval/10000);
            tw_lpid global_dest_id;
     
            for(int k = 0; k < num_other_ranks; k++)    
            {
                other_jid.rank = k;
                int intm_dest_id = codes_jobmap_to_global_id(other_jid, jobmap_ctx); 
                global_dest_id = codes_mapping_get_lpid_from_relative(intm_dest_id, NULL, NW_LP_NM, NULL, 0);

                tw_event * e;
                struct nw_message * m_new;  
                e = tw_event_new(global_dest_id, ts, lp);
531
                m_new = (struct nw_message*)tw_event_data(e);
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
                m_new->msg_type = CLI_BCKGND_FIN;
                tw_event_send(e);   
            }
        }
        return;
}
static void notify_neighbor_rc(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
       if(bf->c0)
       {
            notify_background_traffic_rc(ns, lp, bf, m);
            return;
       }
   
       if(bf->c1)
       {
          tw_rand_reverse_unif(lp->rng); 
       }
} 
static void notify_neighbor(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
    if(ns->local_rank == num_dumpi_traces - 1 
            && ns->is_finished == 1
            && ns->neighbor_completed == 1)
    {
565
//        printf("\n All workloads completed, notifying background traffic ");
566 567 568 569 570 571 572 573 574 575 576 577 578
        bf->c0 = 1;
        notify_background_traffic(ns, lp, bf, m);
        return;
    }
    
    struct codes_jobmap_id nbr_jid;
    nbr_jid.job = ns->app_id;
    tw_lpid global_dest_id;

    if(ns->is_finished == 1 && (ns->neighbor_completed == 1 || ns->local_rank == 0))
    {
        bf->c1 = 1;

579
//        printf("\n Local rank %d notifying neighbor %d ", ns->local_rank, ns->local_rank+1);
580 581 582 583 584 585 586 587 588 589
        tw_stime ts = (1.1 * g_tw_lookahead) + tw_rand_exponential(lp->rng, mean_interval/10000);
        nbr_jid.rank = ns->local_rank + 1;
        
        /* Send a notification to the neighbor about completion */
        int intm_dest_id = codes_jobmap_to_global_id(nbr_jid, jobmap_ctx); 
        global_dest_id = codes_mapping_get_lpid_from_relative(intm_dest_id, NULL, NW_LP_NM, NULL, 0);
       
        tw_event * e;
        struct nw_message * m_new;  
        e = tw_event_new(global_dest_id, ts, lp);
590
        m_new = (struct nw_message*)tw_event_data(e); 
591 592 593 594 595 596 597 598 599 600
        m_new->msg_type = CLI_NBR_FINISH;
        tw_event_send(e);   
    }
}
void finish_bckgnd_traffic_rc(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
601 602 603 604
        (void)b;
        (void)msg;
        (void)lp;

605 606 607 608 609 610 611 612 613
        ns->is_finished = 0;
        return;
}
void finish_bckgnd_traffic(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
614 615
        (void)b;
        (void)msg;
616
        ns->is_finished = 1;
Noah Wolfe's avatar
Noah Wolfe committed
617
        lprintf("\n LP %llu completed sending data %lu completed at time %lf ", LLU(lp->gid), ns->num_bytes_sent, tw_now(lp));
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
        return;
}

void finish_nbr_wkld_rc(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
    ns->neighbor_completed = 0;
    
    notify_neighbor_rc(ns, lp, b, msg);
}

void finish_nbr_wkld(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
    ns->neighbor_completed = 1;

    notify_neighbor(ns, lp, b, msg);
}
static void gen_synthetic_tr_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
    if(bf->c0)
        return;

Noah Wolfe's avatar
Noah Wolfe committed
647 648 649 650 651
    s->msgs_per_tick_sent--;
    if(bf->c17){
        s->msgs_per_tick_sent = m->rc.saved_msgs_per_tick_sent;
    }

652
    model_net_event_rc2(lp, &m->event_rc);
Noah Wolfe's avatar
Noah Wolfe committed
653 654 655
    s->num_bytes_sent -= PAYLOAD_SZ;
    s->num_sends--;
    s->compute_time = m->rc.saved_delay;
656

Noah Wolfe's avatar
Noah Wolfe committed
657 658 659
    if(enable_sampling)
    {
       int indx = s->sampling_indx;
660

Noah Wolfe's avatar
Noah Wolfe committed
661 662 663 664
       s->mpi_wkld_samples[indx].num_sends_sample--;
       s->mpi_wkld_samples[indx].num_bytes_sent_sample -= PAYLOAD_SZ;
       increment_sampling_check_rc(s, bf);
    }
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
}

/* generate synthetic traffic */
static void gen_synthetic_tr(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
    if(s->is_finished == 1)
    {
        bf->c0 = 1;
        return;
    }

    /* Get job information */
    tw_lpid global_dest_id;

    struct codes_jobmap_id jid;
    jid = codes_jobmap_to_local_id(s->nw_id, jobmap_ctx); 

    int num_clients = codes_jobmap_get_num_ranks(jid.job, jobmap_ctx);
    int dest_svr = tw_rand_integer(lp->rng, 0, num_clients - 1);

    if(dest_svr == s->local_rank)
    {
       dest_svr = (s->local_rank + 1) % num_clients;
    }
   
    jid.rank = dest_svr;

    int intm_dest_id = codes_jobmap_to_global_id(jid, jobmap_ctx); 
    global_dest_id = codes_mapping_get_lpid_from_relative(intm_dest_id, NULL, NW_LP_NM, NULL, 0);

    nw_message remote_m;
    remote_m.fwd.sim_start_time = tw_now(lp);
    remote_m.fwd.dest_rank = dest_svr;
    remote_m.msg_type = CLI_BCKGND_ARRIVE;
    remote_m.fwd.num_bytes = PAYLOAD_SZ;
    remote_m.fwd.app_id = s->app_id;
    remote_m.fwd.src_rank = s->local_rank;

    m->event_rc = model_net_event(net_id, "synthetic-tr", global_dest_id, PAYLOAD_SZ, 0.0, 
            sizeof(nw_message), (const void*)&remote_m, 
            0, NULL, lp);
    
Noah Wolfe's avatar
Noah Wolfe committed
707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736
    s->num_bytes_sent += PAYLOAD_SZ;
    num_bytes_sent += PAYLOAD_SZ; 
    s->num_sends++;
    if(enable_sampling)
    {
        increment_sampling_check(s, lp, bf);
        int indx = s->sampling_indx;
        s->mpi_wkld_samples[indx].num_sends_sample++;
        s->mpi_wkld_samples[indx].num_bytes_sent_sample += PAYLOAD_SZ;
        s->mpi_wkld_samples[indx].send_time_sample = s->send_time;
        m->rc.saved_max_bytes = s->mpi_wkld_samples[indx].max_bytes_sample;
        m->rc.saved_min_bytes = s->mpi_wkld_samples[indx].min_bytes_sample;
        s->mpi_wkld_samples[indx].comp_time_sample = s->compute_time;
    }

    int tick_delay = 0;
    /* Check if num messages sent has reached total for the current tick window */
    /* If so, issue next message to start at the beggining of the next tick */
    //printf("LP:%llu, msgs_per_tick_sent:%llu, msgs_per_tick:%llu, tick_delay:%d, tick_interval:%llu, tw_now:%llu\n", LLU(lp->gid),LLU(s->msgs_per_tick_sent), LLU(msgs_per_tick), tick_delay, LLU(tick_interval), LLU(tw_now(lp)));
    if(s->msgs_per_tick_sent >= msgs_per_tick){
        tick_delay = tick_interval - (LLU(tw_now(lp)) % LLU(tick_interval));
        m->rc.saved_delay = s->compute_time;
        s->compute_time += tick_delay;
        m->rc.saved_msgs_per_tick_sent = s->msgs_per_tick_sent;
        s->msgs_per_tick_sent = 0;
        bf->c17 = 1;
        if(LLU(lp->gid) == 0)
            printf("All messages transfered before tick. LP:%llu, msgs_per_tick_sent:%llu, msgs_per_tick:%llu, tick_delay:%d, tick_interval:%llu, tw_now:%llu\n", LLU(lp->gid),LLU(s->msgs_per_tick_sent), LLU(msgs_per_tick), tick_delay, LLU(tick_interval), LLU(tw_now(lp)));
    }
    s->msgs_per_tick_sent++;
737 738

    /* New event after MEAN_INTERVAL */  
Noah Wolfe's avatar
Noah Wolfe committed
739
    tw_stime ts = 0.1 + tick_delay + mean_interval + tw_rand_exponential(lp->rng, noise); 
740 741 742
    tw_event * e;
    nw_message * m_new;
    e = tw_event_new(lp->gid, ts, lp);
743
    m_new = (struct nw_message*)tw_event_data(e);
744 745 746 747 748 749
    m_new->msg_type = CLI_BCKGND_GEN;
    tw_event_send(e);
}

void arrive_syn_tr_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
750 751 752
    (void)bf;
    (void)m;
    (void)lp;
753 754
//    printf("\n Data arrived %d total data %ld ", m->fwd.num_bytes, s->syn_data);
    int data = m->fwd.num_bytes;
Noah Wolfe's avatar
Noah Wolfe committed
755 756 757 758 759 760 761 762 763 764 765
    s->num_bytes_recvd -= data;
    s->num_recvs--;
    s->send_time = m->rc.saved_send_time;
    num_bytes_recvd -= data;
    if(enable_sampling)
    {
        int indx = s->sampling_indx;
        s->mpi_wkld_samples[indx].num_recvs_sample--;
        s->mpi_wkld_samples[indx].num_bytes_recvd_sample -= PAYLOAD_SZ;
        increment_sampling_check_rc(s, bf);
    }
766 767 768
}
void arrive_syn_tr(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
769 770 771
    (void)bf;
    (void)lp;

772 773
//    printf("\n Data arrived %d total data %ld ", m->fwd.num_bytes, s->syn_data);
    int data = m->fwd.num_bytes;
Noah Wolfe's avatar
Noah Wolfe committed
774 775 776 777 778 779 780 781 782 783 784 785
    s->num_bytes_recvd += data;
    s->num_recvs++;
    m->rc.saved_send_time = s->send_time;
    s->send_time += tw_now(lp) - m->fwd.sim_start_time;
    num_bytes_recvd += data;
    if(enable_sampling)
    {
        increment_sampling_check(s, lp, bf);
        int indx = s->sampling_indx;
        s->mpi_wkld_samples[indx].num_recvs_sample++;
        s->mpi_wkld_samples[indx].num_bytes_recvd_sample += PAYLOAD_SZ;
    }
786
}
787
/* Debugging functions, may generate unused function warning */
788
/*static void print_waiting_reqs(uint32_t * reqs, int count)
789
{
790
    lprintf("\n Waiting reqs: %d count", count);
791 792
    int i;
    for(i = 0; i < count; i++ )
793
        lprintf(" %d ", reqs[i]);
794
}*/
795 796 797 798 799 800 801 802 803
static void print_msgs_queue(struct qlist_head * head, int is_send)
{
    if(is_send)
        printf("\n Send msgs queue: ");
    else
        printf("\n Recv msgs queue: ");

    struct qlist_head * ent = NULL;
    mpi_msgs_queue * current = NULL;
804 805
    int count = 0;
    int total_bytes = 0;
806 807 808
    qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, mpi_msgs_queue, ql);
809 810 811 812
            count++;
            total_bytes += current->num_bytes;
            if(count == 0)
                printf(" \n irecv:%d isend:%d op_type:%d Source %d Dest %d bytes %"PRId64" tag %d count:%d total_bytes:%d req_time:%llu", CODES_WK_IRECV, CODES_WK_ISEND, current->op_type, current->source_rank, current->dest_rank, current->num_bytes, current->tag, count, total_bytes, current->req_init_time);
813 814
       }
}
815
/*static void print_completed_queue(tw_lp * lp, struct qlist_head * head)
816
{
817
//    printf("\n Completed queue: ");
818 819
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
820
      tw_output(lp, "\n");
821 822 823
      qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, completed_requests, ql);
824
            tw_output(lp, " %llu ", current->req_id);
825
       }
826
}*/
827
static int clear_completed_reqs(nw_state * s,
828
        tw_lp * lp,
829
        int * reqs, int count)
830
{
831 832 833
    (void)s;
    (void)lp;

834
    int i, matched = 0;
835

836 837 838
    for( i = 0; i < count; i++)
    {
      struct qlist_head * ent = NULL;
839 840 841
      struct completed_requests * current = NULL;
      struct completed_requests * prev = NULL;

842
      int index = 0;
843 844
      qlist_for_each(ent, &s->completed_reqs)
       {
845 846
           if(prev)
           {
847
              rc_stack_push(lp, prev, free, s->matched_reqs);
848 849
              prev = NULL;
           }
850
            
851 852
           current = qlist_entry(ent, completed_requests, ql);
           current->index = index; 
853 854
            if(current->req_id == reqs[i])
            {
855
                ++matched;
856
                qlist_del(&current->ql);
857
                prev = current;
858
            }
859
            ++index;
860
       }
861 862

      if(prev)
863 864 865 866
      {
         rc_stack_push(lp, prev, free, s->matched_reqs);
         prev = NULL;
      }
867
    }
868
    return matched;
869
}
870
static void add_completed_reqs(nw_state * s,
871 872
        tw_lp * lp,
        int count)
873
{
874
    (void)lp;
875
    for(int i = 0; i < count; i++)
876
    {
877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903
       struct completed_requests * req = (struct completed_requests*)rc_stack_pop(s->matched_reqs);

       // turn on only if wait-all unmatched error arises in optimistic mode.
       if(preserve_wait_ordering)
       {
            if(req->index == 0)
            {
                qlist_add(&req->ql, &s->completed_reqs);
            }
            else
            {
                int index = 1;
                struct qlist_head * ent = NULL;
                qlist_for_each(ent, &s->completed_reqs) 
                {
                    if(index == req->index)
                    {
                        qlist_add(&req->ql, ent);
                    }
                }//end qlist
            }// end else*/
       }
       else
       {
               qlist_add(&req->ql, &s->completed_reqs);
       }
    }//end for
904
}
905

906 907 908 909 910 911
/* helper function - maps an MPI rank to an LP id */
static tw_lpid rank_to_lpid(int rank)
{
    return codes_mapping_get_lpid_from_relative(rank, NULL, "nw-lp", NULL, 0);
}

912
static int notify_posted_wait(nw_state* s,
913
        tw_bf * bf, nw_message * m, tw_lp * lp,
914
        int completed_req)
915
{
916 917
    (void)bf;

918 919
    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;
920

921
    m->fwd.wait_completed = 0;
922

923 924
    if(!wait_elem)
        return 0;
925

926
    int op_type = wait_elem->op_type;
927

928 929 930
    if(op_type == CODES_WK_WAIT &&
            (wait_elem->req_ids[0] == completed_req))
    {
931
            m->fwd.wait_completed = 1;
932 933
            wait_completed = 1;
    }
934 935
    else if(op_type == CODES_WK_WAITALL
            || op_type == CODES_WK_WAITANY
936 937 938 939 940 941
            || op_type == CODES_WK_WAITSOME)
    {
        int i;
        for(i = 0; i < wait_elem->count; i++)
        {
            if(wait_elem->req_ids[i] == completed_req)
942
            {
943
                wait_elem->num_completed++;
944
                if(wait_elem->num_completed > wait_elem->count)
945
                    printf("\n Num completed %d count %d LP %llu ",
946 947
                            wait_elem->num_completed,
                            wait_elem->count,
948
                            LLU(lp->gid));
949 950
//                if(wait_elem->num_completed > wait_elem->count)
//                    tw_lp_suspend(lp, 1, 0);
951

952
                if(wait_elem->num_completed >= wait_elem->count)
953
                {
954
                    if(enable_debug)
955
                        fprintf(workload_log, "\n(%lf) APP ID %d MPI WAITALL COMPLETED AT %llu ", tw_now(lp), s->app_id, LLU(s->nw_id));
956
                    wait_completed = 1;
957
                }
958

959
                m->fwd.wait_completed = 1;
960
            }
961
        }
962
    }
963
    return wait_completed;
964
}
965

966
/* reverse handler of MPI wait operation */
967
static void codes_exec_mpi_wait_rc(nw_state* s, tw_bf * bf, tw_lp* lp, nw_message * m)
968
{
969
   if(bf->c1)
970
    {
971
        completed_requests * qi = (completed_requests*)rc_stack_pop(s->processed_ops);
972 973
        if(m->fwd.found_match == 0)
        {
974
            qlist_add(&qi->ql, &s->completed_reqs);
975
        }
976 977 978 979 980 981 982 983 984 985 986 987 988
        else
        {
           int index = 1;
           struct qlist_head * ent = NULL;
           qlist_for_each(ent, &s->completed_reqs)
           {
                if(index == m->fwd.found_match)
                {
                    qlist_add(&qi->ql, ent);
                    break;
                }
                index++;
           }
989
        }
990
        codes_issue_next_event_rc(lp);
991
        return;
992
    }
993 994 995
         struct pending_waits * wait_op = s->wait_op;
         free(wait_op);
         s->wait_op = NULL;
996
}
997

998
/* execute MPI wait operation */
999
static void codes_exec_mpi_wait(nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
1000
{
1001 1002
    /* check in the completed receives queue if the request ID has already been completed.*/
    assert(!s->wait_op);
1003
    int req_id = mpi_op->u.wait.req_id;
1004

1005
    struct completed_requests* current = NULL;
1006

1007
    struct qlist_head * ent = NULL;
1008
    int index = 0;
1009 1010 1011 1012 1013
    qlist_for_each(ent, &s->completed_reqs)
    {
        current = qlist_entry(ent, completed_requests, ql);
        if(current->req_id == req_id)
        {
1014
            bf->c1=1;
1015
            qlist_del(&current->ql);