model-net-mpi-replay.c 38.5 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
/*
 * 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"

17
#define TRACK -1
18
#define TRACE -1
19
#define MAX_WAIT_REQS 512
20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40

char workload_type[128];
char workload_file[8192];
char offset_file[8192];
static int wrkld_id;
static int num_net_traces = 0;

/* 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;

typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
typedef int16_t dumpi_req_id;

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

41 42 43 44
#define CS_LP_DBG 0
#define lprintf(_fmt, ...) \
        do {if (CS_LP_DBG) printf(_fmt, __VA_ARGS__);} while (0)

45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92
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;

/* MPI_OP_GET_NEXT is for getting next MPI operation when the previous operation completes.
* 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. 
* 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,
};

/* 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;
    int num_bytes;
    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;
93
    int16_t req_ids[MAX_WAIT_REQS];
94
	int num_completed;
95 96
	int count;
    tw_stime start_time;
97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112
    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;

    struct rc_stack * processed_ops;
	struct rc_stack * matched_qitems;
113
    struct rc_stack * matched_reqs;
114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140

    /* 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;
	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;
141 142 143

    /* Pending wait operation */
    struct pending_waits * wait_op;
144 145 146 147 148 149 150 151 152
};

/* data for handling reverse computation.
* saved_matched_req holds the request ID of matched receives/sends for wait operations.
* 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. 
* 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
{
   int msg_type;
153 154 155 156 157 158 159 160 161 162 163 164 165 166 167
   int op_type;
   tw_lpid src_rank;
   tw_lpid dest_rank;
   int num_bytes;
   int data_type;
   double sim_start_time;
   
   // for callbacks - time message was received
   double msg_send_time;
   int16_t req_id;   
   int tag;
   int found_match;
    short wait_completed;
    dumpi_req_id saved_matched_req;

168
    struct codes_workload_op * op;
169 170 171 172
    double saved_send_time;
    double saved_recv_time;
    double saved_wait_time;
    struct mpi_msgs_queue * saved_item;
173 174 175 176
};

/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
177
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
178 179
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
180
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
181 182
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
183
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp, struct codes_workload_op * mpi_op);
184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201
/* execute the computational delay */
static void codes_exec_comp_delay(
        nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op);
/* 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 */
202 203 204 205 206 207 208 209
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);
210 211 212 213 214 215 216 217 218 219 220 221 222 223 224
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);

225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275
static void print_waiting_reqs(int16_t * reqs, int count)
{
    printf("\n Waiting reqs: ");
    int i;
    for(i = 0; i < count; i++ )
        printf(" %d ", reqs[i]);
}
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);
            printf(" %ld ", current->req_id);
       }
}
static void clear_completed_reqs(nw_state * s,
        tw_lp * lp,
        int16_t * reqs, int count)
{
    int i;
    for( i = 0; i < count; i++)
    {
      printf("\n req %d ", reqs[i]);
      print_completed_queue(&s->completed_reqs);
      struct qlist_head * ent = NULL;
      qlist_for_each(ent, &s->completed_reqs)
       {
            struct completed_requests* current = 
                qlist_entry(ent, completed_requests, ql);
            if(current->req_id == reqs[i])
            {
                qlist_del(&current->ql);
                rc_stack_push(lp, current, free, s->matched_reqs);
            }
       }
    }
}
static void add_completed_reqs(nw_state * s, 
        int16_t * reqs, int count)
{
    int i;
    for( i = 0; i < count; i++)
    {
       struct completed_requests * req = rc_stack_pop(s->matched_reqs); 
       req->req_id = reqs[i];
       qlist_add(&req->ql, &s->completed_reqs); 
    }
}
276 277 278 279 280 281
/* 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);
}

282 283 284
static int notify_posted_wait(nw_state* s,
        tw_bf * bf, nw_message * m, tw_lp * lp, 
        dumpi_req_id completed_req)
285
{
286 287
    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;
288

289 290 291 292 293 294
    if(!wait_elem)
    {
        m->wait_completed = 0;
        return 0;
    }
    int op_type = wait_elem->op_type;
295

296 297 298 299 300 301 302 303 304 305 306 307 308
    if(op_type == CODES_WK_WAIT &&
            (wait_elem->req_ids[0] == completed_req))
    {
            wait_completed = 1;
    }
    else if(op_type == CODES_WK_WAITALL 
            || op_type == CODES_WK_WAITANY 
            || op_type == CODES_WK_WAITSOME)
    {
        int i;
        for(i = 0; i < wait_elem->count; i++)
        {
            if(wait_elem->req_ids[i] == completed_req)
309
            {
310 311 312 313 314 315 316
                wait_elem->num_completed++;
                assert(wait_elem->num_completed <= wait_elem->count);

                if(wait_elem->num_completed == wait_elem->count)
                    wait_completed = 1;
           
                m->wait_completed = 1;
317
            }
318
        }
319
    }
320
    return wait_completed;
321
}
322

323
/* reverse handler of MPI wait operation */
324
static void codes_exec_mpi_wait_rc(nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op)
325
{
326
    if(s->wait_op)
327
     {
328 329 330
         struct pending_waits * wait_op = s->wait_op;
         free(wait_op);
         s->wait_op = NULL;
331 332 333 334
     }
   else
    {
        codes_issue_next_event_rc(lp);
335 336
        completed_requests * qi = rc_stack_pop(s->matched_reqs);
        qlist_add(&qi->ql, &s->completed_reqs);
337
    }
338
    return;
339
}
340

341
/* execute MPI wait operation */
342
static void codes_exec_mpi_wait(nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op)
343
{
344 345
    /* check in the completed receives queue if the request ID has already been completed.*/
    assert(!s->wait_op);
346
    dumpi_req_id req_id = mpi_op->u.wait.req_id;
347
    struct completed_requests* current = NULL;
348

349 350 351 352 353 354 355 356
    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);
            rc_stack_push(lp, current, free, s->matched_reqs);
357 358 359 360
            codes_issue_next_event(lp);
            return;
        }
    }
361 362 363 364 365
    /* 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;
366 367
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
368
    s->wait_op = wait_op;
369

370
    return;
371 372
}

373 374 375 376 377 378
static void codes_exec_mpi_wait_all_rc(
        nw_state* s, 
        tw_bf * bf,
        nw_message * m,
        tw_lp* lp, 
        struct codes_workload_op * mpi_op)
379
{
380 381 382 383 384 385 386 387 388 389 390 391
  if(s->wait_op)
  {
      struct pending_waits * wait_op = s->wait_op;
      free(wait_op);
      s->wait_op = NULL;
  }
  else
  {
      add_completed_reqs(s, mpi_op->u.waits.req_ids, mpi_op->u.waits.count);
      codes_issue_next_event_rc(lp);
  }
  return;
392 393
}
static void codes_exec_mpi_wait_all(
394 395 396 397 398
        nw_state* s, 
        tw_bf * bf,
        nw_message * m,
        tw_lp* lp, 
        struct codes_workload_op * mpi_op)
399 400 401
{
  //assert(!s->pending_waits);
  int count = mpi_op->u.waits.count;
402 403
  /* If the count is not less than max wait reqs then stop */
  assert(count < MAX_WAIT_REQS);
404

405
  int i = 0, num_matched = 0;
406

407
  if(s->nw_id == TRACK)
408
  {
409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424
      print_waiting_reqs(mpi_op->u.waits.req_ids, count);
      print_completed_queue(&s->completed_reqs);
  }
      /* check number of completed irecvs in the completion queue */ 
  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)
       {
            current = qlist_entry(ent, completed_requests, ql);
            if(current->req_id == req_id)
                num_matched++;
       }
  }
425

426 427 428 429 430 431 432 433 434 435
  if(s->nw_id == TRACK)
    printf("\n num matched %ld count %ld ", num_matched, count);
  m->found_match = num_matched;
  if(num_matched == count)
  {
    /* No need to post a MPI Wait all then, issue next event */
      /* Remove all completed requests from the list */
      s->wait_op = NULL;
      clear_completed_reqs(s, lp, mpi_op->u.waits.req_ids, count);
      codes_issue_next_event(lp);
436 437
  }
  else
438 439 440 441 442 443 444 445 446 447 448
  {
      /* 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;
449
	  wait_op->start_time = tw_now(lp);
450 451
      s->wait_op = wait_op;
    
452
  }
453 454
  return;
}
455 456 457 458

/* search for a matching mpi operation and remove it from the list. 
 * Record the index in the list from where the element got deleted. 
 * Index is used for inserting the element once again in the queue for reverse computation. */
459 460 461 462 463
static int rm_matching_rcv(nw_state * ns, 
        tw_bf * bf,
        nw_message * m, 
        tw_lp * lp, 
        mpi_msgs_queue * qitem)
464 465
{
    int matched = 0;
466
    int index = 0;
467 468
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;
469

470 471 472
    qlist_for_each(ent, &ns->pending_recvs_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
        if((qi->num_bytes >= qitem->num_bytes)
473 474
                && ((qi->tag == qitem->tag) || qi->tag == -1)
                && ((qi->source_rank == qitem->source_rank) || qi->source_rank == -1))
475 476 477 478
        {
            matched = 1;
            break;
        }
479
        ++index;
480 481 482 483
    }
    
    if(matched)
    {
484 485
        m->saved_item = qi;
        m->saved_recv_time = ns->recv_time;
486
        ns->recv_time += (tw_now(lp) - qi->req_init_time);
487 488 489 490 491
        
        if(qi->op_type == CODES_WK_IRECV)
        {
            update_completed_queue(ns, bf, m, lp, qi->req_id);
        }
492 493
        qlist_del(&qi->ql);
        rc_stack_push(lp, qi, free, ns->matched_qitems);
494
        return index;
495 496 497 498
    }
    return -1;
}

499 500 501 502
static int rm_matching_send(nw_state * ns, 
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp, mpi_msgs_queue * qitem)
503 504 505 506 507
{
    int matched = 0;
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;

508
    int index = 0;
509 510 511 512 513 514 515 516 517
    qlist_for_each(ent, &ns->arrival_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
        if((qi->num_bytes <= qitem->num_bytes) 
                && (qi->tag == qitem->tag || qitem->tag == -1)
                && ((qi->source_rank == qitem->source_rank) || qitem->source_rank == -1))
        {
            matched = 1;
            break;
        }
518
        ++index;
519 520 521 522
    }

    if(matched)
    {
523 524 525 526 527 528 529
        m->saved_recv_time = ns->recv_time;
        m->saved_item = qi;
        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);

530 531
        qlist_del(&qi->ql);
        rc_stack_push(lp, qi, free, ns->matched_qitems);
532
        return index;
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
    }
    return -1;
}
static void codes_issue_next_event_rc(tw_lp * lp)
{
	    tw_rand_reverse_unif(lp->rng);	
}

/* 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(
        nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op)
{
	tw_event* e;
	tw_stime ts;
	nw_message* msg;

        if (disable_delay) {
            ts = 0.0; // no compute time sim
        }
        else {
            s->compute_time += s_to_ns(mpi_op->u.delay.seconds);
            ts = s_to_ns(mpi_op->u.delay.seconds);
        }

	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); 
                
}

/* reverse computation operation for MPI irecv */
584 585 586 587 588 589
static void codes_exec_mpi_recv_rc(
        nw_state* ns, 
        tw_bf * bf, 
        nw_message* m, 
        tw_lp* lp, 
        struct codes_workload_op * mpi_op)
590 591
{
	num_bytes_recvd -= mpi_op->u.recv.num_bytes;
592 593
	ns->recv_time = m->saved_recv_time;
	if(m->found_match >= 0)
594
	  {
595
		ns->recv_time = m->saved_recv_time;
596
        int queue_count = qlist_count(&ns->arrival_queue); 
597 598
        
        mpi_msgs_queue * qi = rc_stack_pop(ns->matched_qitems);	
599
       
600
        if(!m->found_match)
601 602 603
        {
            qlist_add(&qi->ql, &ns->arrival_queue);
        }
604
        else if(m->found_match >= queue_count)
605 606 607
        {
            qlist_add_tail(&qi->ql, &ns->arrival_queue);
        }
608
        else if(m->found_match > 0 && m->found_match < queue_count) 
609
        {
610 611 612 613
            int index = 1;
            struct qlist_head * ent = NULL;
            qlist_for_each(ent, &ns->arrival_queue)
            {
614
               if(index == m->found_match)
615 616 617 618 619 620
               {
                 qlist_add(&qi->ql, ent);
                 break;
               }
               index++; 
            }
621
        }
622 623 624
        if(qi->op_type == CODES_WK_IRECV)
            update_completed_queue_rc(ns, bf, m, lp);
        
625 626
        codes_issue_next_event_rc(lp);
      }
627
	else if(m->found_match < 0)
628
	    {
629 630 631 632 633 634
            struct qlist_head * ent = qlist_pop_back(&ns->pending_recvs_queue); 
            mpi_msgs_queue * qi = qlist_entry(ent, mpi_msgs_queue, ql);
            free(qi);
            
            if(mpi_op->op_type == CODES_WK_IRECV)
                codes_issue_next_event_rc(lp);
635 636 637 638
	    }
}

/* Execute MPI Irecv operation (non-blocking receive) */ 
639 640 641 642 643 644
static void codes_exec_mpi_recv(
        nw_state* s, 
        tw_bf * bf,
        nw_message * m, 
        tw_lp* lp, 
        struct codes_workload_op * mpi_op)
645 646 647 648 649
{
/* 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. */

650
	m->saved_recv_time = s->recv_time;
651 652 653 654 655 656 657 658 659 660 661
	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;

662
	int found_matching_sends = rm_matching_send(s, bf, m, lp, recv_op);
663 664 665 666

	/* save the req id inserted in the completed queue for reverse computation. */
	if(found_matching_sends < 0)
	  {
667
	   	  m->found_match = -1;
668
          qlist_add_tail(&recv_op->ql, &s->pending_recvs_queue);
669 670 671 672 673 674 675
	
	       /* 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)
		   {
			codes_issue_next_event(lp);	
			return;
		   }
676
      }
677 678
	else
	  {
679 680
        m->found_match = found_matching_sends;
        free(recv_op);
681 682 683 684 685
        codes_issue_next_event(lp); 
	 }
}

/* executes MPI send and isend operations */
686 687 688 689 690
static void codes_exec_mpi_send(nw_state* s, 
        tw_bf * bf,
        nw_message * m,
        tw_lp* lp, 
        struct codes_workload_op * mpi_op)
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
{
	/* model-net event */
	tw_lpid dest_rank;
	codes_mapping_get_lp_info(lp->gid, lp_group_name, &mapping_grp_id, 
	    lp_type_name, &mapping_type_id, annotation, &mapping_rep_id, &mapping_offset);

	if(net_id == DRAGONFLY) /* special handling for the dragonfly case */
	{
		int num_routers, lps_per_rep, factor;
		num_routers = codes_mapping_get_lp_count("MODELNET_GRP", 1,
                  "dragonfly_router", NULL, 1);
	 	lps_per_rep = (2 * num_nw_lps) + num_routers;	
		factor = mpi_op->u.send.dest_rank / num_nw_lps;
		dest_rank = (lps_per_rep * factor) + (mpi_op->u.send.dest_rank % num_nw_lps);	
	}
	else
	{
		/* other cases like torus/simplenet/loggp etc. */
		codes_mapping_get_lp_id(lp_group_name, lp_type_name, NULL, 1,  
	    	  mpi_op->u.send.dest_rank, mapping_offset, &dest_rank);
	}

	num_bytes_sent += mpi_op->u.send.num_bytes;

	nw_message local_m;
	nw_message remote_m;

718 719 720 721
    local_m.sim_start_time = tw_now(lp);
    local_m.dest_rank = mpi_op->u.send.dest_rank;
    local_m.src_rank = mpi_op->u.send.source_rank;
    local_m.op_type = mpi_op->op_type; 
722
    local_m.msg_type = MPI_SEND_POSTED;
723 724 725
    local_m.tag = mpi_op->u.send.tag;
    local_m.num_bytes = mpi_op->u.send.num_bytes;
    local_m.req_id = mpi_op->u.send.req_id;
726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743

    remote_m = local_m;
	remote_m.msg_type = MPI_SEND_ARRIVED;

	model_net_event(net_id, "test", dest_rank, mpi_op->u.send.num_bytes, 0.0, 
	    sizeof(nw_message), (const void*)&remote_m, sizeof(nw_message), (const void*)&local_m, lp);

	/* isend executed, now get next MPI operation from the queue */ 
	if(mpi_op->op_type == CODES_WK_ISEND)
	   codes_issue_next_event(lp);
}

/* convert seconds to ns */
static tw_stime s_to_ns(tw_stime ns)
{
    return(ns * (1000.0 * 1000.0 * 1000.0));
}

744 745
static void update_completed_queue_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
746 747 748 749
    if(m->wait_completed > 0)
    {
           s->wait_op->num_completed--;
    }
750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797
    if(bf->c0)
    {
       struct qlist_head * ent = qlist_pop_back(&s->completed_reqs);
       completed_requests * req = qlist_entry(ent, completed_requests, ql);
       free(req);
    }
    else if(bf->c1)
    {
       struct pending_waits* wait_elem = rc_stack_pop(s->matched_reqs); 
       s->wait_op = wait_elem;
       s->wait_time = m->saved_wait_time;
       add_completed_reqs(s, s->wait_op->req_ids, s->wait_op->count);
       codes_issue_next_event_rc(lp); 
    }
}

static void update_completed_queue(nw_state* s, 
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp,
        dumpi_req_id req_id)
{
    bf->c0 = 0;
    bf->c1 = 0;
    

    int waiting = notify_posted_wait(s, bf, m, lp, req_id);
 
    if(!waiting)
    {
        bf->c0 = 1;
        completed_requests * req = malloc(sizeof(completed_requests));
        req->req_id = req_id;
        qlist_add_tail(&req->ql, &s->completed_reqs);
    
    }
    else 
     {
            bf->c1 = 1;
            clear_completed_reqs(s, lp, s->wait_op->req_ids, s->wait_op->count);
            m->saved_wait_time = s->wait_time;
            s->wait_time += (tw_now(lp) - s->wait_op->start_time);  
            rc_stack_push(lp, s->wait_op, free, s->matched_reqs);
            s->wait_op = NULL;
            codes_issue_next_event(lp); 
     }
}

798
/* reverse handler for updating arrival queue function */
799 800 801
static void update_arrival_queue_rc(nw_state* s, 
        tw_bf * bf, 
        nw_message * m, tw_lp * lp)
802
{
803
	s->recv_time = m->saved_recv_time;
804
    codes_local_latency_reverse(lp);
805
  
806
    if(m->found_match >= 0)
807 808
	{
        mpi_msgs_queue * qi = rc_stack_pop(s->matched_qitems);	
809 810
        int queue_count = qlist_count(&s->pending_recvs_queue); 

811 812
        
        if(!m->found_match)
813 814 815
        {
            qlist_add(&qi->ql, &s->pending_recvs_queue);
        }
816
        else if(m->found_match >= queue_count)
817
        {
818 819
            qlist_add_tail(&qi->ql, &s->pending_recvs_queue);
        }
820
        else if(m->found_match > 0 && m->found_match < queue_count)
821 822 823 824 825
        {
            int index = 1;
            struct qlist_head * ent = NULL;
            qlist_for_each(ent, &s->pending_recvs_queue)
            {
826
               if(index == m->found_match)
827 828 829 830 831 832
               {
                 qlist_add(&qi->ql, ent);
                 break;
               }
               index++; 
            }
833
        }
834 835
        if(qi->op_type == CODES_WK_IRECV)
            update_completed_queue_rc(s, bf, m, lp);
836
    }
837
	else if(m->found_match < 0)
838 839 840
	{
	    struct qlist_head * ent = qlist_pop_back(&s->arrival_queue); 
        mpi_msgs_queue * qi = qlist_entry(ent, mpi_msgs_queue, ql);
841
        
842 843 844 845 846 847 848
        free(qi);
    }
}

/* once an isend operation arrives, the pending receives queue is checked to find out if there is a irecv that has already been posted. If no isend has been posted, */
static void update_arrival_queue(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
849
	m->saved_recv_time = s->recv_time;
850 851 852 853


    // send a callback to the sender to increment times
    tw_event *e_callback =
854
        tw_event_new(rank_to_lpid(m->src_rank),
855 856 857
                codes_local_latency(lp), lp);
    nw_message *m_callback = tw_event_data(e_callback);
    m_callback->msg_type = MPI_SEND_ARRIVED_CB;
858
    m_callback->msg_send_time = tw_now(lp) - m->sim_start_time;
859 860 861 862
    tw_event_send(e_callback);

    /* Now reconstruct the queue item */
    mpi_msgs_queue * arrived_op = (mpi_msgs_queue *) malloc(sizeof(mpi_msgs_queue));
863 864 865 866 867 868 869
    arrived_op->req_init_time = m->sim_start_time;
    arrived_op->op_type = m->op_type;
    arrived_op->source_rank = m->src_rank;
    arrived_op->dest_rank = m->dest_rank;
    arrived_op->num_bytes = m->num_bytes;
    arrived_op->tag = m->tag;
    arrived_op->req_id = m->req_id;
870

871
    int found_matching_recv = rm_matching_rcv(s, bf, m, lp, arrived_op);
872 873 874

    if(found_matching_recv < 0)
    {
875
        m->found_match = -1;
876 877 878 879
        qlist_add_tail(&arrived_op->ql, &s->arrival_queue);
    }
    else
    {
880 881
        m->found_match = found_matching_recv;
        free(arrived_op);
882 883 884 885 886 887 888 889
    }
}
static void update_message_time(
        nw_state * s,
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp)
{
890 891
    m->saved_send_time = s->send_time;
    s->send_time += m->msg_send_time;
892 893 894 895 896 897 898 899
}

static void update_message_time_rc(
        nw_state * s,
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp)
{
900
    s->send_time = m->saved_send_time;
901 902 903 904 905 906 907 908 909 910
}

/* initializes the network node LP, loads the trace file in the structs, calls the first MPI operation to be executed */
void nw_test_init(nw_state* s, tw_lp* lp)
{
   /* initialize the LP's and load the data */
   char * params = NULL;
   dumpi_trace_params params_d;
  
   memset(s, 0, sizeof(*s));
911
   s->nw_id = codes_mapping_get_lp_relative_id(lp->gid, 0, 0);
912 913 914 915

   if(!num_net_traces) 
	num_net_traces = num_net_lps;

916 917
   assert(num_net_traces <= num_net_lps);

918 919 920 921 922 923 924 925 926 927 928 929 930
   if (strcmp(workload_type, "dumpi") == 0){
       strcpy(params_d.file_name, workload_file);
       params_d.num_net_traces = num_net_traces;

       params = (char*)&params_d;
   }
  /* In this case, the LP will not generate any workload related events*/
   if(s->nw_id >= params_d.num_net_traces)
	    return;

   /* Initialize the RC stack */
   rc_stack_create(&s->processed_ops);
   rc_stack_create(&s->matched_qitems);
931
   rc_stack_create(&s->matched_reqs);
932 933 934

   assert(s->processed_ops != NULL);
   assert(s->matched_qitems != NULL);
935
   assert(s->matched_reqs != NULL);
936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952

   wrkld_id = codes_workload_load("dumpi-trace-workload", params, 0, (int)s->nw_id);

   INIT_QLIST_HEAD(&s->arrival_queue);
   INIT_QLIST_HEAD(&s->pending_recvs_queue);
   INIT_QLIST_HEAD(&s->completed_reqs);

   /* clock starts when the first event is processed */
   s->start_time = tw_now(lp);
   codes_issue_next_event(lp);

   return;
}

void nw_test_event_handler(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
	*(int *)bf = (int)0;
953
    //rc_stack_gc(lp, s->matched_reqs);
954 955 956 957 958 959 960 961 962 963 964 965 966
    rc_stack_gc(lp, s->processed_ops);
    rc_stack_gc(lp, s->matched_qitems);

    switch(m->msg_type)
	{
		case MPI_SEND_ARRIVED:
			update_arrival_queue(s, bf, m, lp);
		break;

		case MPI_SEND_ARRIVED_CB:
			update_message_time(s, bf, m, lp);
		break;

967 968 969 970 971 972 973 974 975 976 977
        case MPI_SEND_POSTED:
        {
           if(m->op_type == CODES_WK_SEND)
               codes_issue_next_event(lp);
           else
            if(m->op_type == CODES_WK_ISEND)
            {
              update_completed_queue(s, bf, m, lp, m->req_id);  
            }
        }
        break;
978 979 980 981 982 983 984 985
		case MPI_OP_GET_NEXT:
			get_next_mpi_operation(s, bf, m, lp);	
		break; 
	}
}

static void get_next_mpi_operation_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
986 987 988
//    struct codes_workload_op * mpi_op = (struct codes_workload_op *)rc_stack_pop(s->processed_ops);

    struct codes_workload_op * mpi_op = m->op;
989 990 991
    codes_workload_get_next_rc(wrkld_id, 0, (int)s->nw_id, mpi_op);

	if(mpi_op->op_type == CODES_WK_END)
992
    {
993
		return;
994
    }
995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010
	switch(mpi_op->op_type)
	{
		case CODES_WK_SEND:
		case CODES_WK_ISEND:
		{
			model_net_event_rc(net_id, lp, mpi_op->u.send.num_bytes);
			if(mpi_op->op_type == CODES_WK_ISEND)
				codes_issue_next_event_rc(lp);
			s->num_sends--;
			num_bytes_sent -= mpi_op->u.send.num_bytes;
		}
		break;

		case CODES_WK_IRECV:
		case CODES_WK_RECV:
		{
1011
			codes_exec_mpi_recv_rc(s, bf, m, lp, mpi_op);
1012 1013 1014
			s->num_recvs--;
		}
		break;
1015 1016 1017
        
		
        case CODES_WK_DELAY:
1018 1019
		{
			s->num_delays--;
1020
            tw_rand_reverse_unif(lp->rng);
1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040
                        
            if (!disable_delay) {
                     s->compute_time -= s_to_ns(mpi_op->u.delay.seconds);
            }
		}
		break;
		case CODES_WK_BCAST:
		case CODES_WK_ALLGATHER:
		case CODES_WK_ALLGATHERV:
		case CODES_WK_ALLTOALL:
		case CODES_WK_ALLTOALLV:
		case CODES_WK_REDUCE:
		case CODES_WK_ALLREDUCE:
		case CODES_WK_COL:
		{
			s->num_cols--;
		    codes_issue_next_event_rc(lp);
        }
		break;
	
1041 1042 1043 1044 1045 1046 1047 1048
		case CODES_WK_WAITSOME:
		case CODES_WK_WAITANY:
        {
           s->num_waitsome--;
           codes_issue_next_event_rc(lp); 
        }
        break;

1049 1050 1051
		case CODES_WK_WAIT:
		{
			s->num_wait--;
1052
			codes_exec_mpi_wait_rc(s, lp, mpi_op);
1053 1054 1055 1056 1057
		}
		break;
		case CODES_WK_WAITALL:
		{
			s->num_waitall--;
1058
			codes_exec_mpi_wait_all_rc(s, bf, m, lp, mpi_op);
1059 1060 1061 1062 1063
		}
		break;
		default:
			printf("\n Invalid op type %d ", mpi_op->op_type);
	}
1064
    free(mpi_op);
1065 1066 1067 1068 1069 1070
}

static void get_next_mpi_operation(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
		struct codes_workload_op * mpi_op = malloc(sizeof(struct codes_workload_op));
        codes_workload_get_next(wrkld_id, 0, (int)s->nw_id, mpi_op);
1071
       
1072
        m->op = mpi_op;
1073 1074 1075 1076

        if(mpi_op->op_type == CODES_WK_END)
        {
            s->elapsed_time = tw_now(lp) - s->start_time;
1077
            //rc_stack_push(lp, mpi_op, free, s->processed_ops);
1078 1079 1080 1081 1082 1083 1084 1085
            return;
        }
		switch(mpi_op->op_type)
		{
			case CODES_WK_SEND:
			case CODES_WK_ISEND:
			 {
				s->num_sends++;
1086
				codes_exec_mpi_send(s, bf, m, lp, mpi_op);
1087 1088 1089 1090 1091 1092 1093
			 }
			break;
	
			case CODES_WK_RECV:
			case CODES_WK_IRECV:
			  {
				s->num_recvs++;
1094
				codes_exec_mpi_recv(s, bf, m, lp,  mpi_op);
1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106
			  }
			break;

			case CODES_WK_DELAY:
			  {
				s->num_delays++;
				codes_exec_comp_delay(s, lp, mpi_op);
			  }
			break;

            case CODES_WK_WAITSOME:
            case CODES_WK_WAITANY:
1107 1108 1109 1110 1111 1112
            {
                s->num_waitsome++;
                codes_issue_next_event(lp);
            }
            break;

1113
			case CODES_WK_WAITALL:
1114
			  {
1115 1116 1117
				s->num_waitall++;
			    codes_exec_mpi_wait_all(s, bf, m, lp, mpi_op);
              }
1118 1119 1120 1121
			break;
			case CODES_WK_WAIT:
			{
				s->num_wait++;
1122
                codes_exec_mpi_wait(s, lp, mpi_op);
1123 1124
			}
			break;
1125 1126 1127 1128 1129 1130 1131 1132
			case CODES_WK_BCAST:
			case CODES_WK_ALLGATHER:
			case CODES_WK_ALLGATHERV:
			case CODES_WK_ALLTOALL:
			case CODES_WK_ALLTOALLV:
			case CODES_WK_REDUCE:
			case CODES_WK_ALLREDUCE:
			case CODES_WK_COL:
1133
			{
1134
				s->num_cols++;
1135 1136 1137 1138 1139 1140
			    codes_issue_next_event(lp);
            }
			break;
			default:
				printf("\n Invalid op type %d ", mpi_op->op_type);
		}
1141
//        rc_stack_push(lp, mpi_op, free, s->processed_ops);
1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175
        return;
}

void nw_test_finalize(nw_state* s, tw_lp* lp)
{
	if(s->nw_id < num_net_traces)
	{
		int count_irecv = qlist_count(&s->pending_recvs_queue);
        int count_isend = qlist_count(&s->arrival_queue);
		printf("\n LP %ld unmatched irecvs %d unmatched sends %d Total sends %ld receives %ld collectives %ld delays %ld wait alls %ld waits %ld send time %lf wait %lf", 
			lp->gid, count_irecv, count_isend, s->num_sends, s->num_recvs, s->num_cols, s->num_delays, s->num_waitall, s->num_wait, s->send_time, s->wait_time);

		if(s->elapsed_time - s->compute_time > max_comm_time)
			max_comm_time = s->elapsed_time - s->compute_time;
		
		if(s->elapsed_time > max_time )
			max_time = s->elapsed_time;

		/*if(s->wait_time > max_wait_time)
			max_wait_time = s->wait_time;
        */
		if(s->send_time > max_send_time)
			max_send_time = s->send_time;

		if(s->recv_time > max_recv_time)
			max_recv_time = s->recv_time;

		avg_time += s->elapsed_time;
		avg_comm_time += (s->elapsed_time - s->compute_time);
		avg_wait_time += s->wait_time;
		avg_send_time += s->send_time;
		 avg_recv_time += s->recv_time;

		//printf("\n LP %ld Time spent in communication %llu ", lp->gid, total_time - s->compute_time);
1176
	    rc_stack_destroy(s->matched_reqs);    
1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193
	    rc_stack_destroy(s->matched_qitems);    
	    rc_stack_destroy(s->processed_ops);    
    }
}

void nw_test_event_handler_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
	switch(m->msg_type)
	{
		case MPI_SEND_ARRIVED:
			update_arrival_queue_rc(s, bf, m, lp);
		break;

		case MPI_SEND_ARRIVED_CB:
			update_message_time_rc(s, bf, m, lp);
		break;

1194 1195 1196 1197 1198 1199 1200 1201 1202 1203
        case MPI_SEND_POSTED:
        {
         if(m->op_type == CODES_WK_SEND) 
             codes_issue_next_event_rc(lp);
         else if(m->op_type == CODES_WK_ISEND)
            update_completed_queue_rc(s, bf, m, lp);  
        }
        break;
		
        case MPI_OP_GET_NEXT:
1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276