model-net-mpi-replay.c 50.5 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/*
 * Copyright (C) 2014 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
 */
#include <ross.h>
#include <inttypes.h>

#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"
16
#include "codes/codes-jobmap.h"
17

18 19
/* turning on track lp will generate a lot of output messages */
#define TRACK_LP -1
20
#define TRACE -1
21
#define MAX_WAIT_REQS 512
22 23 24 25
#define CS_LP_DBG 0
#define lprintf(_fmt, ...) \
        do {if (CS_LP_DBG) printf(_fmt, __VA_ARGS__);} while (0)
#define MAX_STATS 65536
26 27 28 29 30 31

char workload_type[128];
char workload_file[8192];
char offset_file[8192];
static int wrkld_id;
static int num_net_traces = 0;
32
static int alloc_spec = 0;
33 34 35 36 37 38 39

/* Doing LP IO*/
static char lp_io_dir[256] = {'\0'};
static lp_io_handle io_handle;
static unsigned int lp_io_use_suffix = 0;
static int do_lp_io = 0;

40 41 42 43 44 45 46 47 48 49 50
/* variables for loading multiple applications */
/* Xu's additions start */
char workloads_conf_file[8192];
char alloc_file[8192];
int num_traces_of_job[5];
char file_name_of_job[5][8192];

struct codes_jobmap_ctx *jobmap_ctx;
struct codes_jobmap_params_list jobmap_p;
/* Xu's additions end */

51 52
typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
53
typedef int32_t dumpi_req_id;
54 55 56 57 58

static int net_id = 0;
static float noise = 5.0;
static int num_net_lps, num_nw_lps;

59 60 61 62 63
FILE * workload_log = NULL;
FILE * workload_agg_log = NULL;
FILE * workload_meta_log = NULL;

static uint64_t sample_bytes_written = 0;
64

65 66 67 68 69 70 71 72 73 74 75 76
long long num_bytes_sent=0;
long long num_bytes_recvd=0;

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;

/* global variables for codes mapping */
static char lp_group_name[MAX_NAME_LENGTH], lp_type_name[MAX_NAME_LENGTH], annotation[MAX_NAME_LENGTH];
static int mapping_grp_id, mapping_type_id, mapping_rep_id, mapping_offset;

/* runtime option for disabling computation time simulation */
static int disable_delay = 0;
77 78 79
static int enable_sampling = 0;
static double sampling_interval = 5000000;
static double sampling_end_time = 3000000000;
80
static int enable_debug = 1;
81 82

/* MPI_OP_GET_NEXT is for getting next MPI operation when the previous operation completes.
83
* 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.
84 85 86 87 88 89 90 91 92
* 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,
};

93 94 95 96
struct mpi_workload_sample
{
    /* Sampling data */
    int nw_id;
97
    int app_id;
98 99 100 101 102
    unsigned long num_sends_sample;
    unsigned long num_bytes_sample;
    unsigned long num_waits_sample;
    double sample_end_time;
};
103 104 105 106 107 108 109
/* 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;
110
    uint64_t num_bytes;
111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126
    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
{
	dumpi_req_id req_id;
    struct qlist_head ql;
};

/* for wait operations, store the pending operation and number of completed waits so far. */
struct pending_waits
{
    int op_type;
127
    int32_t req_ids[MAX_WAIT_REQS];
128
	int num_completed;
129 130
	int count;
    tw_stime start_time;
131 132 133 134 135 136 137 138 139 140 141 142 143
    struct qlist_head ql;
};

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;
144 145
    int app_id;
    int local_rank;
146 147

    struct rc_stack * processed_ops;
148
    struct rc_stack * matched_reqs;
149 150 151 152 153 154 155 156 157 158

    /* 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;

159

160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176
	/* time spent by the LP in executing the app trace*/
	double start_time;
	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;
177

178 179
    tw_stime cur_interval_end;

180 181
    /* Pending wait operation */
    struct pending_waits * wait_op;
182 183 184 185

    unsigned long num_bytes_sent;
    unsigned long num_bytes_recvd;

186 187 188 189
    /* For sampling data */
    int sampling_indx;
    int max_arr_size;
    struct mpi_workload_sample * mpi_wkld_samples;
190
    char output_buf[512];
191 192 193 194
};

/* data for handling reverse computation.
* saved_matched_req holds the request ID of matched receives/sends for wait operations.
195
* 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.
196 197 198
* 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
{
199
   // forward message handler
200
   int msg_type;
201
   int op_type;
202

203 204 205 206
   struct
   {
       tw_lpid src_rank;
       tw_lpid dest_rank;
207
       int64_t num_bytes;
208 209 210 211 212
       int num_matched;
       int data_type;
       double sim_start_time;
       // for callbacks - time message was received
       double msg_send_time;
213
       int16_t req_id;
214
       int tag;
215
       int app_id;
216 217 218 219 220 221 222 223 224
       int found_match;
       short wait_completed;
   } fwd;
   struct
   {
       double saved_send_time;
       double saved_recv_time;
       double saved_wait_time;
       double saved_delay;
225
       int64_t saved_num_bytes;
226
       struct codes_workload_op * saved_op;
227
   } rc;
228 229 230 231
};

/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
232
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
233 234
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
235
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
236 237
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
238
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp);
239 240
/* execute the computational delay */
static void codes_exec_comp_delay(
241
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
242 243 244 245 246 247 248 249 250 251 252 253 254 255 256
/* 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 */
257 258 259 260 261 262 263 264
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);
265 266 267 268 269 270 271 272 273 274 275 276 277 278 279
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);

280
/* Debugging functions, may generate unused function warning */
281
static void print_waiting_reqs(int32_t * reqs, int count)
282 283 284 285
{
    printf("\n Waiting reqs: ");
    int i;
    for(i = 0; i < count; i++ )
286
        printf(" %d ", reqs[i]);
287
}
288 289 290 291 292 293 294 295 296 297 298 299 300 301 302
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;
    qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, mpi_msgs_queue, ql);
            printf(" \n Source %d Dest %d bytes %d tag %d ", current->source_rank, current->dest_rank, current->num_bytes, current->tag);
       }
}
303 304 305 306 307 308 309 310
static void print_completed_queue(struct qlist_head * head)
{
    printf("\n Completed queue: ");
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
      qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, completed_requests, ql);
311
            printf(" %d ", current->req_id);
312 313
       }
}
314
static int clear_completed_reqs(nw_state * s,
315
        tw_lp * lp,
316
        int32_t * reqs, int count)
317
{
318
    int i, matched = 0;
319 320 321 322 323
    for( i = 0; i < count; i++)
    {
      struct qlist_head * ent = NULL;
      qlist_for_each(ent, &s->completed_reqs)
       {
324
            struct completed_requests* current =
325 326 327
                qlist_entry(ent, completed_requests, ql);
            if(current->req_id == reqs[i])
            {
328
                ++matched;
329 330 331 332 333
                qlist_del(&current->ql);
                rc_stack_push(lp, current, free, s->matched_reqs);
            }
       }
    }
334
    return matched;
335
}
336
static void add_completed_reqs(nw_state * s,
337 338
        tw_lp * lp,
        int count)
339 340 341 342
{
    int i;
    for( i = 0; i < count; i++)
    {
343 344
       struct completed_requests * req = rc_stack_pop(s->matched_reqs);
       qlist_add(&req->ql, &s->completed_reqs);
345 346
    }
}
347

348 349 350 351 352 353
/* 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);
}

354
static int notify_posted_wait(nw_state* s,
355
        tw_bf * bf, nw_message * m, tw_lp * lp,
356
        dumpi_req_id completed_req)
357
{
358 359
    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;
360

361
    m->fwd.wait_completed = 0;
362

363 364
    if(!wait_elem)
        return 0;
365

366
    int op_type = wait_elem->op_type;
367

368 369 370 371 372
    if(op_type == CODES_WK_WAIT &&
            (wait_elem->req_ids[0] == completed_req))
    {
            wait_completed = 1;
    }
373 374
    else if(op_type == CODES_WK_WAITALL
            || op_type == CODES_WK_WAITANY
375 376 377 378 379 380
            || op_type == CODES_WK_WAITSOME)
    {
        int i;
        for(i = 0; i < wait_elem->count; i++)
        {
            if(wait_elem->req_ids[i] == completed_req)
381
            {
382
                wait_elem->num_completed++;
383
                if(wait_elem->num_completed > wait_elem->count)
384
                    printf("\n Num completed %d count %d LP %llu ",
385 386 387
                            wait_elem->num_completed,
                            wait_elem->count,
                            lp->gid);
388 389
//                if(wait_elem->num_completed > wait_elem->count)
//                    tw_lp_suspend(lp, 1, 0);
390

391
                if(wait_elem->num_completed == wait_elem->count)
392
                {
393 394
                    if(enable_debug)
                        fprintf(workload_log, "\n(%lf) APP ID %d MPI WAITALL COMPLETED AT %ld ", s->app_id, tw_now(lp), s->nw_id);
395
                    wait_completed = 1;
396
                }
397

398
                m->fwd.wait_completed = 1;
399
            }
400
        }
401
    }
402
    return wait_completed;
403
}
404

405
/* reverse handler of MPI wait operation */
406
static void codes_exec_mpi_wait_rc(nw_state* s, tw_lp* lp)
407
{
408
    if(s->wait_op)
409
     {
410 411 412
         struct pending_waits * wait_op = s->wait_op;
         free(wait_op);
         s->wait_op = NULL;
413 414 415 416
     }
   else
    {
        codes_issue_next_event_rc(lp);
417
        completed_requests * qi = rc_stack_pop(s->processed_ops);
418
        qlist_add(&qi->ql, &s->completed_reqs);
419
    }
420
    return;
421
}
422

423
/* execute MPI wait operation */
424
static void codes_exec_mpi_wait(nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op)
425
{
426 427
    /* check in the completed receives queue if the request ID has already been completed.*/
    assert(!s->wait_op);
428
    dumpi_req_id req_id = mpi_op->u.wait.req_id;
429
    struct completed_requests* current = NULL;
430

431 432 433 434 435 436 437
    struct qlist_head * ent = NULL;
    qlist_for_each(ent, &s->completed_reqs)
    {
        current = qlist_entry(ent, completed_requests, ql);
        if(current->req_id == req_id)
        {
            qlist_del(&current->ql);
438
            rc_stack_push(lp, current, free, s->processed_ops);
439 440 441 442
            codes_issue_next_event(lp);
            return;
        }
    }
443 444 445 446 447
    /* If not, add the wait operation in the pending 'waits' list. */
    struct pending_waits* wait_op = malloc(sizeof(struct pending_waits));
    wait_op->op_type = mpi_op->op_type;
    wait_op->req_ids[0] = req_id;
    wait_op->count = 1;
448 449
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
450
    s->wait_op = wait_op;
451

452
    return;
453 454
}

455
static void codes_exec_mpi_wait_all_rc(
456
        nw_state* s,
457 458
        tw_bf * bf,
        nw_message * m,
459
        tw_lp* lp)
460
{
461 462 463 464 465 466 467 468 469 470 471
  if(bf->c1)
  {
    int sampling_indx = s->sampling_indx;
    s->mpi_wkld_samples[sampling_indx].num_waits_sample--;

    if(bf->c2)
    {
        s->cur_interval_end -= sampling_interval;
        s->sampling_indx--;
    }
  }
472 473 474 475 476 477 478 479
  if(s->wait_op)
  {
      struct pending_waits * wait_op = s->wait_op;
      free(wait_op);
      s->wait_op = NULL;
  }
  else
  {
480
      add_completed_reqs(s, lp, m->fwd.num_matched);
481 482 483
      codes_issue_next_event_rc(lp);
  }
  return;
484 485
}
static void codes_exec_mpi_wait_all(
486
        nw_state* s,
487 488
        tw_bf * bf,
        nw_message * m,
489
        tw_lp* lp,
490
        struct codes_workload_op * mpi_op)
491
{
492 493
  if(enable_debug)
    fprintf(workload_log, "\n MPI WAITALL POSTED AT %ld ", s->nw_id);
494

495 496 497 498 499 500 501 502
  if(enable_sampling)
  {
    bf->c1 = 1;
    if(tw_now(lp) >= s->cur_interval_end)
    {
        bf->c2 = 1;
        int indx = s->sampling_indx;
        s->mpi_wkld_samples[indx].nw_id = s->nw_id;
503
        s->mpi_wkld_samples[indx].app_id = s->app_id;
504 505 506 507 508 509
        s->mpi_wkld_samples[indx].sample_end_time = s->cur_interval_end;
        s->cur_interval_end += sampling_interval;
        s->sampling_indx++;
    }
    if(s->sampling_indx >= MAX_STATS)
    {
510
        struct mpi_workload_sample * tmp = calloc((MAX_STATS + s->max_arr_size), sizeof(struct mpi_workload_sample));
511 512 513 514 515 516 517 518
        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;
    }
    int indx = s->sampling_indx;
    s->mpi_wkld_samples[indx].num_waits_sample++;
  }
519
  int count = mpi_op->u.waits.count;
520 521
  /* If the count is not less than max wait reqs then stop */
  assert(count < MAX_WAIT_REQS);
522

523
  int i = 0, num_matched = 0;
524
  m->fwd.num_matched = 0;
525

526
  /*if(lp->gid == TRACK)
527
  {
528
      printf("\n MPI Wait all posted ");
529 530
      print_waiting_reqs(mpi_op->u.waits.req_ids, count);
      print_completed_queue(&s->completed_reqs);
531
  }*/
532
      /* check number of completed irecvs in the completion queue */
533 534 535 536 537 538 539
  for(i = 0; i < count; i++)
  {
      dumpi_req_id req_id = mpi_op->u.waits.req_ids[i];
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
      qlist_for_each(ent, &s->completed_reqs)
       {
540
            current = qlist_entry(ent, struct completed_requests, ql);
541 542 543 544
            if(current->req_id == req_id)
                num_matched++;
       }
  }
545

546
  m->fwd.found_match = num_matched;
547 548 549 550
  if(num_matched == count)
  {
    /* No need to post a MPI Wait all then, issue next event */
      /* Remove all completed requests from the list */
551 552 553
      m->fwd.num_matched = clear_completed_reqs(s, lp, mpi_op->u.waits.req_ids, count);
      struct pending_waits* wait_op = s->wait_op;
      free(wait_op);
554 555
      s->wait_op = NULL;
      codes_issue_next_event(lp);
556 557
  }
  else
558 559 560 561 562 563 564 565 566 567 568
  {
      /* If not, add the wait operation in the pending 'waits' list. */
	  struct pending_waits* wait_op = malloc(sizeof(struct pending_waits));
	  wait_op->count = count;
      wait_op->op_type = mpi_op->op_type;
      assert(count < MAX_WAIT_REQS);

      for(i = 0; i < count; i++)
          wait_op->req_ids[i] =  mpi_op->u.waits.req_ids[i];

	  wait_op->num_completed = num_matched;
569
	  wait_op->start_time = tw_now(lp);
570
      s->wait_op = wait_op;
571
  }
572 573
  return;
}
574

575 576
/* search for a matching mpi operation and remove it from the list.
 * Record the index in the list from where the element got deleted.
577
 * Index is used for inserting the element once again in the queue for reverse computation. */
578
static int rm_matching_rcv(nw_state * ns,
579
        tw_bf * bf,
580 581
        nw_message * m,
        tw_lp * lp,
582
        mpi_msgs_queue * qitem)
583 584
{
    int matched = 0;
585
    int index = 0;
586 587
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;
588

589 590
    qlist_for_each(ent, &ns->pending_recvs_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
591
        if((qi->num_bytes == qitem->num_bytes)
592 593
                && ((qi->tag == qitem->tag) || qi->tag == -1)
                && ((qi->source_rank == qitem->source_rank) || qi->source_rank == -1))
594 595 596 597
        {
            matched = 1;
            break;
        }
598
        ++index;
599
    }
600

601 602
    if(matched)
    {
603
        m->rc.saved_recv_time = ns->recv_time;
604
        ns->recv_time += (tw_now(lp) - qi->req_init_time);
605

606 607
        if(qi->op_type == CODES_WK_IRECV)
            update_completed_queue(ns, bf, m, lp, qi->req_id);
608 609
        else if(qi->op_type == CODES_WK_RECV)
            codes_issue_next_event(lp);
610

611
        qlist_del(&qi->ql);
612

613
        rc_stack_push(lp, qi, free, ns->processed_ops);
614
        return index;
615 616 617 618
    }
    return -1;
}

619
static int rm_matching_send(nw_state * ns,
620 621 622
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp, mpi_msgs_queue * qitem)
623 624 625 626 627
{
    int matched = 0;
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;

628
    int index = 0;
629 630
    qlist_for_each(ent, &ns->arrival_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
631
        if((qi->num_bytes == qitem->num_bytes)
632 633 634 635 636 637
                && (qi->tag == qitem->tag || qitem->tag == -1)
                && ((qi->source_rank == qitem->source_rank) || qitem->source_rank == -1))
        {
            matched = 1;
            break;
        }
638
        ++index;
639 640 641 642
    }

    if(matched)
    {
643
        m->rc.saved_recv_time = ns->recv_time;
644 645 646 647 648
        ns->recv_time += (tw_now(lp) - qitem->req_init_time);

        if(qitem->op_type == CODES_WK_IRECV)
            update_completed_queue(ns, bf, m, lp, qitem->req_id);

649
        qlist_del(&qi->ql);
650

651
        return index;
652 653 654 655 656
    }
    return -1;
}
static void codes_issue_next_event_rc(tw_lp * lp)
{
657
	    tw_rand_reverse_unif(lp->rng);
658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677
}

/* Trigger getting next event at LP */
static void codes_issue_next_event(tw_lp* lp)
{
   tw_event *e;
   nw_message* msg;

   tw_stime ts;

   ts = g_tw_lookahead + 0.1 + tw_rand_exponential(lp->rng, noise);
   e = tw_event_new( lp->gid, ts, lp );
   msg = tw_event_data(e);

   msg->msg_type = MPI_OP_GET_NEXT;
   tw_event_send(e);
}

/* Simulate delays between MPI operations */
static void codes_exec_comp_delay(
678
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
679 680 681 682 683
{
	tw_event* e;
	tw_stime ts;
	nw_message* msg;

684
    m->rc.saved_delay = s->compute_time;
685 686
    s->compute_time += s_to_ns(mpi_op->u.delay.seconds);
    ts = s_to_ns(mpi_op->u.delay.seconds);
687 688

	ts += g_tw_lookahead + 0.1 + tw_rand_exponential(lp->rng, noise);
689

690 691 692
	e = tw_event_new( lp->gid, ts , lp );
	msg = tw_event_data(e);
	msg->msg_type = MPI_OP_GET_NEXT;
693 694
	tw_event_send(e);

695 696 697
}

/* reverse computation operation for MPI irecv */
698
static void codes_exec_mpi_recv_rc(
699 700 701
        nw_state* ns,
        tw_bf * bf,
        nw_message* m,
702
        tw_lp* lp)
703
{
704
	num_bytes_recvd -= m->rc.saved_num_bytes;
705 706
	ns->recv_time = m->rc.saved_recv_time;
	if(m->fwd.found_match >= 0)
707
	  {
708
		ns->recv_time = m->rc.saved_recv_time;
709 710 711 712
        int queue_count = qlist_count(&ns->arrival_queue);

        mpi_msgs_queue * qi = rc_stack_pop(ns->processed_ops);

713
        if(!m->fwd.found_match)
714 715 716
        {
            qlist_add(&qi->ql, &ns->arrival_queue);
        }
717
        else if(m->fwd.found_match >= queue_count)
718 719 720
        {
            qlist_add_tail(&qi->ql, &ns->arrival_queue);
        }
721
        else if(m->fwd.found_match > 0 && m->fwd.found_match < queue_count)
722
        {
723 724 725 726
            int index = 1;
            struct qlist_head * ent = NULL;
            qlist_for_each(ent, &ns->arrival_queue)
            {
727
               if(index == m->fwd.found_match)
728 729 730 731
               {
                 qlist_add(&qi->ql, ent);
                 break;
               }
732
               index++;
733
            }
734
        }
735
        if(qi->op_type == CODES_WK_IRECV)
736
        {
737
            update_completed_queue_rc(ns, bf, m, lp);
738
        }
739 740
        codes_issue_next_event_rc(lp);
      }
741
	else if(m->fwd.found_match < 0)
742
	    {
743
            struct qlist_head * ent = qlist_pop_back(&ns->pending_recvs_queue);
744 745
            mpi_msgs_queue * qi = qlist_entry(ent, mpi_msgs_queue, ql);
            free(qi);
746

747
            if(m->op_type == CODES_WK_IRECV)
748
                codes_issue_next_event_rc(lp);
749 750 751
	    }
}

752
/* Execute MPI Irecv operation (non-blocking receive) */
753
static void codes_exec_mpi_recv(
754
        nw_state* s,
755
        tw_bf * bf,
756 757
        nw_message * m,
        tw_lp* lp,
758
        struct codes_workload_op * mpi_op)
759 760 761 762 763
{
/* Once an irecv is posted, list of completed sends is checked to find a matching isend.
   If no matching isend is found, the receive operation is queued in the pending queue of
   receive operations. */

764
	m->rc.saved_recv_time = s->recv_time;
765 766
    m->rc.saved_num_bytes = mpi_op->u.recv.num_bytes;

767 768 769 770 771 772 773 774 775 776 777
	num_bytes_recvd += mpi_op->u.recv.num_bytes;

    mpi_msgs_queue * recv_op = (mpi_msgs_queue*) malloc(sizeof(mpi_msgs_queue));
    recv_op->req_init_time = tw_now(lp);
    recv_op->op_type = mpi_op->op_type;
    recv_op->source_rank = mpi_op->u.recv.source_rank;
    recv_op->dest_rank = mpi_op->u.recv.dest_rank;
    recv_op->num_bytes = mpi_op->u.recv.num_bytes;
    recv_op->tag = mpi_op->u.recv.tag;
    recv_op->req_id = mpi_op->u.recv.req_id;

778 779 780 781
    if(s->nw_id == TRACK_LP)
        printf("\n Receive op posted num bytes %d source %d ", recv_op->num_bytes,
                recv_op->source_rank);

782
	int found_matching_sends = rm_matching_send(s, bf, m, lp, recv_op);
783 784 785 786

	/* save the req id inserted in the completed queue for reverse computation. */
	if(found_matching_sends < 0)
	  {
787
	   	  m->fwd.found_match = -1;
788
          qlist_add_tail(&recv_op->ql, &s->pending_recvs_queue);
789

790 791 792
	       /* for mpi irecvs, this is a non-blocking receive so just post it and move on with the trace read. */
		if(mpi_op->op_type == CODES_WK_IRECV)
		   {
793
			codes_issue_next_event(lp);
794 795
			return;
		   }
796
      }
797 798
	else
	  {
799
        m->fwd.found_match = found_matching_sends;
800
        codes_issue_next_event(lp);
801 802
	    rc_stack_push(lp, recv_op, free, s->processed_ops);
      }
803 804
}

805 806 807 808 809 810 811 812
int get_global_id_of_job_rank(tw_lpid job_rank, int app_id)
{
    struct codes_jobmap_id lid;
    lid.job = app_id;
    lid.rank = job_rank;
    int global_rank = codes_jobmap_to_global_id(lid, jobmap_ctx);
    return global_rank;
}
813
/* executes MPI send and isend operations */
814
static void codes_exec_mpi_send(nw_state* s,
815 816
        tw_bf * bf,
        nw_message * m,
817
        tw_lp* lp,
818
        struct codes_workload_op * mpi_op)
819
{
820 821
	/* model-net event */
    int global_dest_rank = mpi_op->u.send.dest_rank;
822

823 824 825 826 827
    if(alloc_spec)
    {
        global_dest_rank = get_global_id_of_job_rank(mpi_op->u.send.dest_rank, s->app_id);
    }

828
    m->rc.saved_num_bytes = mpi_op->u.send.num_bytes;
829 830
	/* model-net event */
	tw_lpid dest_rank;
831 832 833
    dest_rank = codes_mapping_get_lpid_from_relative(global_dest_rank, NULL, "nw-lp", NULL, 0);
//	if(net_id == DRAGONFLY) /* special handling for the dragonfly case */
/*	{
834
		int num_routers, lps_per_rep, factor;
835 836
		num_routers = codes_mapping_get_lp_count(lp_group_name, 1,
                  "modelnet_dragonfly_router", NULL, 1);
837
	 	lps_per_rep = (2 * num_nw_lps) + num_routers;
838
		factor = global_dest_rank / num_nw_lps;
839
		dest_rank = (lps_per_rep * factor) + (global_dest_rank % num_nw_lps);
840 841 842
	}
	else
	{
843 844
*/		/* other cases like torus/simplenet/loggp etc. */
/*		codes_mapping_get_lp_id(lp_group_name, lp_type_name, NULL, 1,
845
	    	  global_dest_rank, mapping_offset, &dest_rank);
846
	}
847
*/
848
	num_bytes_sent += mpi_op->u.send.num_bytes;
849
    s->num_bytes_sent += mpi_op->u.send.num_bytes;
850

851 852 853 854 855 856 857
    if(enable_sampling)
    {
        if(tw_now(lp) >= s->cur_interval_end)
        {
            bf->c1 = 1;
            int indx = s->sampling_indx;
            s->mpi_wkld_samples[indx].nw_id = s->nw_id;
858
            s->mpi_wkld_samples[indx].app_id = s->app_id;
859 860
            s->mpi_wkld_samples[indx].sample_end_time = s->cur_interval_end;
            s->sampling_indx++;
861
            s->cur_interval_end += sampling_interval;
862 863 864
        }
        if(s->sampling_indx >= MAX_STATS)
        {
865
            struct mpi_workload_sample * tmp = calloc((MAX_STATS + s->max_arr_size), sizeof(struct mpi_workload_sample));
866 867 868 869 870 871 872 873 874
            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;
        }
        int indx = s->sampling_indx;
        s->mpi_wkld_samples[indx].num_sends_sample++;
        s->mpi_wkld_samples[indx].num_bytes_sample += mpi_op->u.send.num_bytes;
    }
875 876 877
	nw_message local_m;
	nw_message remote_m;

878 879 880
    local_m.fwd.sim_start_time = tw_now(lp);
    local_m.fwd.dest_rank = mpi_op->u.send.dest_rank;
    local_m.fwd.src_rank = mpi_op->u.send.source_rank;
881
    local_m.op_type = mpi_op->op_type;
882
    local_m.msg_type = MPI_SEND_POSTED;
883 884 885
    local_m.fwd.tag = mpi_op->u.send.tag;
    local_m.fwd.num_bytes = mpi_op->u.send.num_bytes;
    local_m.fwd.req_id = mpi_op->u.send.req_id;
886
    local_m.fwd.app_id = s->app_id;
887 888 889 890

    remote_m = local_m;
	remote_m.msg_type = MPI_SEND_ARRIVED;

891
	model_net_event(net_id, "test", dest_rank, mpi_op->u.send.num_bytes, 0.0,
892 893
	    sizeof(nw_message