model-net-mpi-wrklds.c 44.4 KB
Newer Older
1 2 3 4 5 6
/*
 * Copyright (C) 2014 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
 */
#include <ross.h>
7
#include <inttypes.h>
8

9
#include "codes/codes-workload.h"
10 11 12 13 14
#include "codes/codes.h"
#include "codes/configuration.h"
#include "codes/codes_mapping.h"
#include "codes/model-net.h"

15
#define TRACE -1
16 17 18 19 20 21 22 23 24

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

typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
25
typedef int16_t dumpi_req_id;
26 27 28 29 30 31

static int net_id = 0;
static float noise = 5.0;
static int num_net_lps, num_nw_lps;
long long num_bytes_sent=0;
long long num_bytes_recvd=0;
32 33
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;
34 35 36 37 38

/* 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;

39 40 41
/* runtime option for disabling computation time simulation */
static int disable_delay = 0;

42 43 44
/* 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). */
45 46 47 48
enum MPI_NW_EVENTS
{
	MPI_OP_GET_NEXT=1,
	MPI_SEND_ARRIVED,
49
    MPI_SEND_ARRIVED_CB, // for tracking message times on sender
50 51 52
	MPI_SEND_POSTED,
};

53
/* stores pointers of pending MPI operations to be matched with their respective sends/receives. */
54 55
struct mpi_msgs_queue
{
56 57
	struct codes_workload_op * mpi_op;
	struct mpi_msgs_queue * next;
58 59
};

60
/* stores request IDs of completed MPI operations (Isends or Irecvs) */
61 62 63
struct completed_requests
{
	dumpi_req_id req_id;
64
	struct completed_requests * next;
65 66
};

67
/* for wait operations, store the pending operation and number of completed waits so far. */
68 69
struct pending_waits
{
70
	struct codes_workload_op * mpi_op;
71 72 73 74
	int num_completed;
	tw_stime start_time;
};

75
/* maintains the head and tail of the queue, as well as the number of elements currently in queue. Queues are pending_recvs queue (holds unmatched MPI recv operations) and arrival_queue (holds unmatched MPI send messages). */
76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94
struct mpi_queue_ptrs
{
	int num_elems;
	struct mpi_msgs_queue* queue_head;
	struct mpi_msgs_queue* queue_tail;
};

/* 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;

	/* count of sends, receives, collectives and delays */
	unsigned long num_sends;
	unsigned long num_recvs;
	unsigned long num_cols;
	unsigned long num_delays;
95 96 97
	unsigned long num_wait;
	unsigned long num_waitall;
	unsigned long num_waitsome;
98 99

	/* time spent by the LP in executing the app trace*/
100
	double start_time;
101
	double elapsed_time;
102
	/* time spent in compute operations */
103 104 105 106 107 108 109
	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;
110
	/* FIFO for isend messages arrived on destination */
111
	struct mpi_queue_ptrs * arrival_queue;
112
	/* FIFO for irecv messages posted but not yet matched with send operations */
113
	struct mpi_queue_ptrs * pending_recvs_queue;
114
	/* list of pending waits (and saved pending wait for reverse computation) */
115
	struct pending_waits * pending_waits;
116
	/* List of completed send/receive requests */
117
	struct completed_requests * completed_reqs;
118 119
};

120 121 122 123
/* 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). */
124 125
struct nw_message
{
126 127 128
   int msg_type;
   /* for reverse computation */
   struct codes_workload_op * op;
129

130 131 132 133 134 135 136 137 138 139 140
   struct
   {
     /* forward event handler */
     struct
     {
        int op_type;
        tw_lpid src_rank;
        tw_lpid dest_rank;
        int num_bytes;
        int data_type;
        double sim_start_time;
141 142
        // for callbacks - time message was received
        double msg_send_time;
143 144 145 146 147 148 149
        int16_t req_id;   
        int tag;
     } msg_info;

     /* required for reverse computation*/
     struct 
      {
150 151 152 153 154 155 156 157 158
        int found_match;
        short matched_op;
        dumpi_req_id saved_matched_req;
        struct codes_workload_op* ptr_match_op;
        struct pending_waits* saved_pending_wait;

        double saved_send_time;
        double saved_recv_time;
        double saved_wait_time;
159 160
      } rc;
  } u;
161 162
};

163
/* executes MPI wait operation */
164 165
static void codes_exec_mpi_wait(
        nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op);
166
/* reverse of mpi wait function. */
167 168
static void codes_exec_mpi_wait_rc(
        nw_state* s, tw_bf* bf, nw_message* m, tw_lp* lp);
169
/* executes MPI isend and send operations */
170 171
static void codes_exec_mpi_send(
        nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op);
172
/* execute MPI irecv operation */
173 174
static void codes_exec_mpi_recv(
        nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op);
175
/* reverse of mpi recv function. */
176 177
static void codes_exec_mpi_recv_rc(
        nw_state* s, nw_message* m, tw_lp* lp);
178
/* execute the computational delay */
179 180
static void codes_exec_comp_delay(
        nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op);
181
/* execute collective operation, currently only skips these operations. */
182 183
static void codes_exec_mpi_col(
        nw_state* s, tw_lp* lp);
184
/* gets the next MPI operation from the network-workloads API. */
185 186
static void get_next_mpi_operation(
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
187
/* reverse handler of get next mpi operation. */
188 189
static void get_next_mpi_operation_rc(
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
190 191 192 193
/* Makes a call to get_next_mpi_operation. */
static void codes_issue_next_event(tw_lp* lp);

///////////////////// HELPER FUNCTIONS FOR MPI MESSAGE QUEUE HANDLING ///////////////
194
/* upon arrival of local completion message, inserts operation in completed send queue */
195 196
static void update_send_completion_queue(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
197
/* reverse of the above function */
198 199
static void update_send_completion_queue_rc(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
200
/* upon arrival of an isend operation, updates the arrival queue of the network */
201 202
static void update_arrival_queue(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
203
/* reverse of the above function */
204 205
static void update_arrival_queue_rc(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
206
/* callback to a message sender for computing message time */
207 208 209 210 211
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);
212

213
/* insert MPI operation in the waiting queue*/
214 215
static void mpi_pending_queue_insert_op(
        struct mpi_queue_ptrs* mpi_queue, struct codes_workload_op* mpi_op);
216
/* remove completed request IDs from the queue for reuse. Reverse of above function. */
217 218
static void remove_req_id(
        struct completed_requests** requests, int16_t req_id);
219
/* remove MPI operation from the waiting queue.*/
220 221
static int mpi_queue_remove_matching_op(
        nw_state* s, tw_lp* lp, nw_message * m, struct mpi_queue_ptrs* mpi_queue, struct codes_workload_op * mpi_op);
222
/* remove the tail of the MPI operation from waiting queue */
223 224
static int mpi_queue_remove_tail(
        tw_lpid lpid, struct mpi_queue_ptrs* mpi_queue);
225
/* insert completed MPI requests in the queue. */
226 227
static void mpi_completed_queue_insert_op(
        struct completed_requests** mpi_completed_queue, dumpi_req_id req_id);
228
/* notifies the wait operations (if any) about the completed receives and sends requests. */
229 230
static int notify_waits(
        nw_state* s, tw_bf* bf, tw_lp* lp, nw_message* m, dumpi_req_id req_id);
231
/* reverse of notify waits function. */
232 233
static void notify_waits_rc(
        nw_state* s, tw_bf* bf, tw_lp* lp, nw_message* m, dumpi_req_id completed_req);
234 235
/* conversion from seconds to eanaoseconds */
static tw_stime s_to_ns(tw_stime ns);
236

237 238 239 240 241
/* 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);
}
242

243 244 245 246 247 248 249 250 251 252 253 254
/* initializes the queue and allocates memory */
static struct mpi_queue_ptrs* queue_init()
{
	struct mpi_queue_ptrs* mpi_queue = malloc(sizeof(struct mpi_queue_ptrs));

	mpi_queue->num_elems = 0;
	mpi_queue->queue_head = NULL;
	mpi_queue->queue_tail = NULL;
	
	return mpi_queue;
}

255
/* helper function: counts number of elements in the queue */
256 257
static int numQueue(struct mpi_queue_ptrs* mpi_queue)
{
258
	struct mpi_msgs_queue* tmp = mpi_queue->queue_head;
259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279
	int count = 0;

	while(tmp)
	{
		++count;
		tmp = tmp->next;
	}
	return count;
}

/* prints elements in a send/recv queue */
static void printQueue(tw_lpid lpid, struct mpi_queue_ptrs* mpi_queue, char* msg)
{
	printf("\n ************ Printing the queue %s *************** ", msg);
	struct mpi_msgs_queue* tmp = malloc(sizeof(struct mpi_msgs_queue));
	assert(tmp);

	tmp = mpi_queue->queue_head;
	
	while(tmp)
	{
280
		if(tmp->mpi_op->op_type == CODES_WK_SEND || tmp->mpi_op->op_type == CODES_WK_ISEND)
281 282 283
			printf("\n lpid %ld send operation data type %d count %d tag %d source %d", 
				    lpid, tmp->mpi_op->u.send.data_type, tmp->mpi_op->u.send.count, 
				     tmp->mpi_op->u.send.tag, tmp->mpi_op->u.send.source_rank);
284
		else if(tmp->mpi_op->op_type == CODES_WK_IRECV || tmp->mpi_op->op_type == CODES_WK_RECV)
285 286 287 288 289 290 291 292 293 294 295
			printf("\n lpid %ld recv operation data type %d count %d tag %d source %d", 
				   lpid, tmp->mpi_op->u.recv.data_type, tmp->mpi_op->u.recv.count, 
				    tmp->mpi_op->u.recv.tag, tmp->mpi_op->u.recv.source_rank );
		else
			printf("\n Invalid data type in the queue %d ", tmp->mpi_op->op_type);
		tmp = tmp->next;
	}
	free(tmp);
}

/* re-insert element in the queue at the index --- maintained for reverse computation */
296
static void mpi_queue_update(struct mpi_queue_ptrs* mpi_queue, struct codes_workload_op* mpi_op, int pos)
297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332
{
	struct mpi_msgs_queue* elem = malloc(sizeof(struct mpi_msgs_queue));
	assert(elem);
	elem->mpi_op = mpi_op;
	
	/* inserting at the head */
	if(pos == 0)
	{
	   if(!mpi_queue->queue_tail)
		mpi_queue->queue_tail = elem;
	   elem->next = mpi_queue->queue_head;
	   mpi_queue->queue_head = elem;
	   mpi_queue->num_elems++;
	   return;
	}

	int index = 0;
	struct mpi_msgs_queue* tmp = mpi_queue->queue_head;
	while(index < pos - 1)
	{
		tmp = tmp->next;
		++index;
	}

	if(!tmp)
		printf("\n Invalid index! %d pos %d size %d ", index, pos, numQueue(mpi_queue));
	if(tmp == mpi_queue->queue_tail)
	    mpi_queue->queue_tail = elem;

	elem->next = tmp->next;
	tmp->next = elem;
	mpi_queue->num_elems++;

	return;
}

333
/* prints the elements of a queue (for debugging purposes). */
334 335 336 337
static void printCompletedQueue(nw_state* s, tw_lp* lp)
{
	   if(TRACE == lp->gid)
	   {
338
	   	printf("\n %lf contents of completed operations queue ", tw_now(lp));
339 340 341 342 343 344 345 346 347
	   	struct completed_requests* current = s->completed_reqs;
	   	while(current)
	    	{
			printf(" %d ",current->req_id);
			current = current->next;
	   	}
	   }
}

348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363
/* reverse handler of notify_waits function. */
static void notify_waits_rc(nw_state* s, tw_bf* bf, tw_lp* lp, nw_message* m, dumpi_req_id completed_req)
{
   int i;

   /*if(bf->c1)
    {*/
	/* if pending wait is still present and is of type MPI_WAIT then do nothing*/
/*	s->wait_time = s->saved_wait_time; 	
	mpi_completed_queue_insert_op(&s->completed_reqs, completed_req);	
	s->pending_waits = wait_elem;
	s->saved_pending_wait = NULL;
    }
*/
  if(lp->gid == TRACE)
	  printf("\n %lf reverse -- notify waits req id %d ", tw_now(lp), completed_req);
364
  
365
  printCompletedQueue(s, lp);
366
  
367
  if(m->u.rc.matched_op == 1)
368 369
	s->pending_waits->num_completed--;
   /* if a wait-elem exists, it means the request ID has been matched*/
370
   if(m->u.rc.matched_op == 2) 
371
    {
372 373 374 375 376
        if(lp->gid == TRACE)
        {
            printf("\n %lf matched req id %d ", tw_now(lp), completed_req);
            printCompletedQueue(s, lp);
        }
377
        struct pending_waits* wait_elem = m->u.rc.saved_pending_wait;
378 379
        s->wait_time = m->u.rc.saved_wait_time;
        int count = wait_elem->mpi_op->u.waits.count; 
380

381 382
        for( i = 0; i < count; i++ )
            mpi_completed_queue_insert_op(&s->completed_reqs, wait_elem->mpi_op->u.waits.req_ids[i]);
383

384 385 386
        wait_elem->num_completed--;	
        s->pending_waits = wait_elem;
        tw_rand_reverse_unif(lp->rng);
387 388 389 390 391 392

   }
}

/* notify the completed send/receive request to the wait operation. */
static int notify_waits(nw_state* s, tw_bf* bf, tw_lp* lp, nw_message* m, dumpi_req_id completed_req)
393 394 395 396 397 398 399
{
	int i;
	/* traverse the pending waits list and look what type of wait operations are 
	there. If its just a single wait and the request ID has just been completed, 
	then the network node LP can go on with fetching the next operation from the log.
	If its waitall then wait for all pending requests to complete and then proceed. */
	struct pending_waits* wait_elem = s->pending_waits;
400
	m->u.rc.matched_op = 0;
401 402 403
	
	if(lp->gid == TRACE)
		printf("\n %lf notify waits req id %d ", tw_now(lp), completed_req);
404 405 406

	if(!wait_elem)
		return 0;
407

408 409
	int op_type = wait_elem->mpi_op->op_type;

410
	if(op_type == CODES_WK_WAIT)
411 412 413
	{
		if(wait_elem->mpi_op->u.wait.req_id == completed_req)	
		  {
414
			m->u.rc.saved_wait_time = s->wait_time;
415
			s->wait_time += (tw_now(lp) - wait_elem->start_time);
416
                        remove_req_id(&s->completed_reqs, completed_req);
417
	
418
			m->u.rc.saved_pending_wait = wait_elem;			
419 420 421 422 423
			s->pending_waits = NULL;
			codes_issue_next_event(lp);	
			return 0;
		 }
	}
424
	else if(op_type == CODES_WK_WAITALL)
425
	{
426 427
	   int required_count = wait_elem->mpi_op->u.waits.count;
	  for(i = 0; i < required_count; i++)
428 429
	   {
	    if(wait_elem->mpi_op->u.waits.req_ids[i] == completed_req)
430 431 432
		{
			if(lp->gid == TRACE)
				printCompletedQueue(s, lp);
433
			m->u.rc.matched_op = 1;
434
			wait_elem->num_completed++;	
435
		}
436 437
	   }
	   
438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456
	  if(wait_elem->num_completed == required_count)
	   {
            if(lp->gid == TRACE)
            {
                printf("\n %lf req %d completed %d", tw_now(lp), completed_req, wait_elem->num_completed);
                printCompletedQueue(s, lp);
            }
            m->u.rc.matched_op = 2;
            m->u.rc.saved_wait_time = s->wait_time;
            s->wait_time += (tw_now(lp) - wait_elem->start_time);
            m->u.rc.saved_pending_wait = wait_elem;
            s->pending_waits = NULL; 
            
            for(i = 0; i < required_count; i++)
                remove_req_id(&s->completed_reqs, wait_elem->mpi_op->u.waits.req_ids[i]);	
            
            codes_issue_next_event(lp); //wait completed
          }
    }
457 458 459
	return 0;
}

460 461 462 463 464 465
/* reverse handler of MPI wait operation */
static void codes_exec_mpi_wait_rc(nw_state* s, tw_bf* bf, nw_message* m, tw_lp* lp)
{
    if(s->pending_waits)
     {
    	s->pending_waits = NULL;
466
	    return;
467 468 469
     }
   else
    {
470 471
 	    mpi_completed_queue_insert_op(&s->completed_reqs, m->op->u.wait.req_id);	
	    tw_rand_reverse_unif(lp->rng);		
472 473
    }
}
474 475

/* execute MPI wait operation */
476
static void codes_exec_mpi_wait(nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op)
477
{
Jonathan Jenkins's avatar
Jonathan Jenkins committed
478 479
    /* check in the completed receives queue if the request ID has already been completed.*/
    assert(!s->pending_waits);
480
    dumpi_req_id req_id = mpi_op->u.wait.req_id;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
481 482 483 484 485 486 487 488 489 490

    struct completed_requests* current = s->completed_reqs;
    while(current) {
        if(current->req_id == req_id) {
            remove_req_id(&s->completed_reqs, req_id);
            m->u.rc.saved_wait_time = s->wait_time;
            codes_issue_next_event(lp);
            return;
        }
        current = current->next;
491 492
    }

Jonathan Jenkins's avatar
Jonathan Jenkins committed
493
    /* If not, add the wait operation in the pending 'waits' list. */
494
    struct pending_waits* wait_op = malloc(sizeof(struct pending_waits));
495
    wait_op->mpi_op = mpi_op;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
496 497
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
498
    s->pending_waits = wait_op;
499 500
}

501
static void codes_exec_mpi_wait_all_rc(nw_state* s, tw_bf* bf, nw_message* m, tw_lp* lp)
502
{
503 504 505 506 507
  if(lp->gid == TRACE)
   {
       printf("\n %lf codes exec mpi waitall reverse %d ", tw_now(lp), m->u.rc.found_match);
       printCompletedQueue(s, lp); 
   } 
508
  if(m->u.rc.found_match)
509
    {
510 511 512 513 514 515 516 517 518 519 520
        int i;
        int count = m->op->u.waits.count;
        dumpi_req_id req_id[count];

        for( i = 0; i < count; i++)
        {
            req_id[i] = m->op->u.waits.req_ids[i];
            mpi_completed_queue_insert_op(&s->completed_reqs, req_id[i]);
        }
        tw_rand_reverse_unif(lp->rng);
   }
521 522
    else
    {
523 524 525 526 527 528 529
        struct pending_waits* wait_op = s->pending_waits;
        free(wait_op);
        s->pending_waits = NULL;
        assert(!s->pending_waits);
        if(lp->gid == TRACE)
            printf("\n %lf Nullifying codes waitall ", tw_now(lp));
   }
530
}
531 532
static void codes_exec_mpi_wait_all(
        nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op)
533 534
{
  //assert(!s->pending_waits);
535
  int count = mpi_op->u.waits.count;
536 537 538 539 540 541 542
  int i, num_completed = 0;
  dumpi_req_id req_id[count];
  struct completed_requests* current = s->completed_reqs;

  /* check number of completed irecvs in the completion queue */ 
  if(lp->gid == TRACE)
    {
543
  	printf(" \n (%lf) MPI waitall posted %d count", tw_now(lp), mpi_op->u.waits.count);
544
	for(i = 0; i < count; i++)
545
		printf(" %d ", (int)mpi_op->u.waits.req_ids[i]);
546 547 548 549 550 551
   	printCompletedQueue(s, lp);	 
   }
  while(current) 
   {
	  for(i = 0; i < count; i++)
	   {
552
	     req_id[i] = mpi_op->u.waits.req_ids[i];
553
	     if(req_id[i] == current->req_id)
554
 		    num_completed++;
555 556 557 558 559
   	  }
	 current = current->next;
   }

  if(TRACE== lp->gid)
560
	  printf("\n %lf Num completed %d count %d ", tw_now(lp), num_completed, count);
561

562
  m->u.rc.found_match = 0;
563
  if(count == num_completed)
564
  {
565
	m->u.rc.found_match = 1;
566
	for( i = 0; i < count; i++)	
567
		remove_req_id(&s->completed_reqs, req_id[i]);
568

569 570 571 572 573 574
	codes_issue_next_event(lp);
  }
  else
  {
 	/* If not, add the wait operation in the pending 'waits' list. */
	  struct pending_waits* wait_op = malloc(sizeof(struct pending_waits));
575
	  wait_op->mpi_op = mpi_op;  
576
	  wait_op->num_completed = num_completed;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
577
	  wait_op->start_time = tw_now(lp);
578 579 580 581 582
	  s->pending_waits = wait_op;
  }
}

/* request ID is being reused so delete it from the list once the matching is done */
583 584
static void remove_req_id(
        struct completed_requests** mpi_completed_queue, dumpi_req_id req_id)
585 586 587 588
{
	struct completed_requests* current = *mpi_completed_queue;

	if(!current)
589 590
		tw_error(TW_LOC, "\n REQ ID DOES NOT EXIST");
	
591
    if(current->req_id == req_id)
592 593 594 595 596 597 598 599 600 601 602 603
	{
		*mpi_completed_queue = current->next;
		free(current);
		return;
	}
	
	struct completed_requests* elem;
	while(current->next)
	{
	   elem = current->next;
	   if(elem->req_id == req_id)	
	     {
604 605 606 607
            current->next = elem->next;
            free(elem);
            return;
         }
608 609 610 611 612 613
	   current = current->next;	
	}
	return;
}

/* inserts mpi operation in the completed requests queue */
614 615
static void mpi_completed_queue_insert_op(
        struct completed_requests** mpi_completed_queue, dumpi_req_id req_id)
616 617 618 619 620 621 622 623
{
	struct completed_requests* reqs = malloc(sizeof(struct completed_requests));
	assert(reqs);

	reqs->req_id = req_id;

	if(!(*mpi_completed_queue))	
	{
624 625 626
        reqs->next = NULL;
        *mpi_completed_queue = reqs;
        return;
627 628 629
	}
	reqs->next = *mpi_completed_queue;
	*mpi_completed_queue = reqs;
630
	return;
631 632
}

633
/* insert MPI send or receive operation in the queues starting from tail. Unmatched sends go to arrival queue and unmatched receives go to pending receives queues. */
634 635
static void mpi_pending_queue_insert_op(
        struct mpi_queue_ptrs* mpi_queue, struct codes_workload_op* mpi_op)
636 637 638 639 640 641
{
	/* insert mpi operation */
	struct mpi_msgs_queue* elem = malloc(sizeof(struct mpi_msgs_queue));
	assert(elem);

	elem->mpi_op = mpi_op;
642
    elem->next = NULL;
643 644 645 646 647 648 649

	if(!mpi_queue->queue_head)
	  mpi_queue->queue_head = elem;

	if(mpi_queue->queue_tail)
	    mpi_queue->queue_tail->next = elem;
	
650
    mpi_queue->queue_tail = elem;
651 652 653 654 655 656
	mpi_queue->num_elems++;

	return;
}

/* match the send/recv operations */
657 658
static int match_receive(
        nw_state* s, tw_lp* lp, tw_lpid lpid, struct codes_workload_op* op1, struct codes_workload_op* op2)
659
{
660 661 662 663 664 665 666 667 668
        assert(op1->op_type == CODES_WK_IRECV || op1->op_type == CODES_WK_RECV);
        assert(op2->op_type == CODES_WK_SEND || op2->op_type == CODES_WK_ISEND);

        if((op1->u.recv.num_bytes >= op2->u.send.num_bytes) &&
                   ((op1->u.recv.tag == op2->u.send.tag) || op1->u.recv.tag == -1) &&
                   ((op1->u.recv.source_rank == op2->u.send.source_rank) || op1->u.recv.source_rank == -1))
                   {
                        if(lp->gid == TRACE)
                           printf("\n op1 rank %d bytes %d ", op1->u.recv.source_rank, op1->u.recv.num_bytes);
669
                        s->recv_time += tw_now(lp) - op1->sim_start_time;
670 671 672 673
                        mpi_completed_queue_insert_op(&s->completed_reqs, op1->u.recv.req_id);
                        return 1;
                   }
        return -1;
674 675 676
}

/* used for reverse computation. removes the tail of the queue */
677
static int mpi_queue_remove_tail(tw_lpid lpid, struct mpi_queue_ptrs* mpi_queue)
678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711
{
	assert(mpi_queue->queue_tail);
	if(mpi_queue->queue_tail == NULL)
	{
		printf("\n Error! tail not updated ");	
		return 0;
	}
	struct mpi_msgs_queue* tmp = mpi_queue->queue_head;

	if(mpi_queue->queue_head == mpi_queue->queue_tail)
	{
		mpi_queue->queue_head = NULL;
		mpi_queue->queue_tail = NULL;
		free(tmp);
		mpi_queue->num_elems--;
		 return 1;
	}

	struct mpi_msgs_queue* elem = mpi_queue->queue_tail;

	while(tmp->next != mpi_queue->queue_tail)
		tmp = tmp->next;

	mpi_queue->queue_tail = tmp;
	mpi_queue->queue_tail->next = NULL;
	mpi_queue->num_elems--;

	free(elem);
	return 1;
}

/* 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. */
712
static int mpi_queue_remove_matching_op(nw_state* s, tw_lp* lp, nw_message * m, struct mpi_queue_ptrs* mpi_queue, struct codes_workload_op * mpi_op)
713 714 715 716 717 718 719 720 721
{
	if(mpi_queue->queue_head == NULL)
		return -1;

	/* remove mpi operation */
	struct mpi_msgs_queue* tmp = mpi_queue->queue_head;
	int indx = 0;

	/* if head of the list has the required mpi op to be deleted */
722
	int rcv_val = 0;
723
	if(mpi_op->op_type == CODES_WK_SEND || mpi_op->op_type == CODES_WK_ISEND)
724
	  {
725
		rcv_val = match_receive(s, lp, lp->gid, tmp->mpi_op, mpi_op);
726
		m->u.rc.saved_matched_req = tmp->mpi_op->u.recv.req_id;  
727
	 }
728
	else if(mpi_op->op_type == CODES_WK_RECV || mpi_op->op_type == CODES_WK_IRECV)
729
	  {
730
		rcv_val = match_receive(s, lp, lp->gid, mpi_op, tmp->mpi_op);
731
	  	m->u.rc.saved_matched_req = mpi_op->u.recv.req_id;
732 733
	  }
	if(rcv_val >= 0)
734
	{
735
		/* TODO: fix RC */
736
		m->u.rc.ptr_match_op = tmp->mpi_op;
737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758
		if(mpi_queue->queue_head == mpi_queue->queue_tail)
		   {
			mpi_queue->queue_tail = NULL;
			mpi_queue->queue_head = NULL;
			 free(tmp);
		   }
		 else
		   {
			mpi_queue->queue_head = tmp->next;
			free(tmp);	
		   }
		mpi_queue->num_elems--;
		return indx;
	}

	/* record the index where matching operation has been found */
	struct mpi_msgs_queue* elem;

	while(tmp->next)	
	{
	   indx++;
	   elem = tmp->next;
759
	   
760
	    if(mpi_op->op_type == CODES_WK_SEND || mpi_op->op_type == CODES_WK_ISEND)
761
	     {
762
		    rcv_val = match_receive(s, lp, lp->gid, elem->mpi_op, mpi_op);
763
	     	m->u.rc.saved_matched_req = elem->mpi_op->u.recv.req_id; 
764
	     }
765
	    else if(mpi_op->op_type == CODES_WK_RECV || mpi_op->op_type == CODES_WK_IRECV)
766
	     {
767 768 769
            rcv_val = match_receive(s, lp, lp->gid, mpi_op, elem->mpi_op);
            m->u.rc.saved_matched_req = mpi_op->u.recv.req_id;
         }
770
   	     if(rcv_val >= 0)
771 772
		 {
		    m->u.rc.ptr_match_op = elem->mpi_op;
773
		    if(elem == mpi_queue->queue_tail)
774
			    mpi_queue->queue_tail = tmp;
775
		    
776 777 778 779
		    tmp->next = elem->next;

		    free(elem);
		    mpi_queue->num_elems--;
780
		
781
		    return indx;
782
		 }
783
	   tmp = tmp->next;
784
    }
785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803
	return -1;
}
/* 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 */
804 805
static void codes_exec_comp_delay(
        nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op)
806 807 808 809 810
{
	tw_event* e;
	tw_stime ts;
	nw_message* msg;

811 812 813 814 815 816 817 818 819
        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);
820 821 822 823 824 825 826 827 828
	
	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 */
829
static void codes_exec_mpi_recv_rc(nw_state* s, nw_message* m, tw_lp* lp)
830
{
831 832 833
	num_bytes_recvd -= m->op->u.recv.num_bytes;
	s->recv_time = m->u.rc.saved_recv_time;
	if(m->u.rc.found_match >= 0)
834
	  {
835 836 837
		s->recv_time = m->u.rc.saved_recv_time;
		mpi_queue_update(s->arrival_queue, m->u.rc.ptr_match_op, m->u.rc.found_match);
		remove_req_id(&s->completed_reqs, m->op->u.recv.req_id);
838
		tw_rand_reverse_unif(lp->rng);
839
	  }
840
	else if(m->u.rc.found_match < 0)
841
	    {
842 843
		mpi_queue_remove_tail(lp->gid, s->pending_recvs_queue);
		if(m->op->op_type == CODES_WK_IRECV)
844
			tw_rand_reverse_unif(lp->rng);
845 846 847 848
	    }
}

/* Execute MPI Irecv operation (non-blocking receive) */ 
849
static void codes_exec_mpi_recv(nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op)
850 851 852 853 854
{
/* 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. */

855
	m->u.rc.saved_recv_time = s->recv_time;
856
	mpi_op->sim_start_time = tw_now(lp);
857
	num_bytes_recvd += mpi_op->u.recv.num_bytes;
858 859

	if(lp->gid == TRACE)
860
		printf("\n %lf codes exec mpi recv req id %d", tw_now(lp), (int)mpi_op->u.recv.req_id);
861 862
	
	dumpi_req_id req_id;
863
	int found_matching_sends = mpi_queue_remove_matching_op(s, lp, m, s->arrival_queue, mpi_op);
864 865 866
	
	/* save the req id inserted in the completed queue for reverse computation. */
	//m->matched_recv = req_id;
867 868 869

	if(found_matching_sends < 0)
	  {
870
		m->u.rc.found_match = -1;
871 872 873
		mpi_pending_queue_insert_op(s->pending_recvs_queue, mpi_op);
	
	       /* for mpi irecvs, this is a non-blocking receive so just post it and move on with the trace read. */
874
		if(mpi_op->op_type == CODES_WK_IRECV)
875 876 877 878 879 880
		   {
			codes_issue_next_event(lp);	
			return;
		   }
		else
			printf("\n CODES MPI RECV OPERATION!!! ");
881
	  }
882
	else
883
	  {
884
	   	m->u.rc.found_match = found_matching_sends;
885
		codes_issue_next_event(lp); 
886 887 888 889
	 }
}

/* executes MPI send and isend operations */
890
static void codes_exec_mpi_send(nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op)
891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915
{
	/* 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;

916 917
	nw_message local_m;
	nw_message remote_m;
918

919 920 921 922 923 924 925 926
    local_m.u.msg_info.sim_start_time = tw_now(lp);
    local_m.u.msg_info.dest_rank = mpi_op->u.send.dest_rank;
    local_m.u.msg_info.src_rank = mpi_op->u.send.source_rank;
    local_m.u.msg_info.op_type = mpi_op->op_type; 
    local_m.msg_type = MPI_SEND_POSTED;
    local_m.u.msg_info.tag = mpi_op->u.send.tag;
    local_m.u.msg_info.num_bytes = mpi_op->u.send.num_bytes;
    local_m.u.msg_info.req_id = mpi_op->u.send.req_id;
927

928 929
    remote_m = local_m;
	remote_m.msg_type = MPI_SEND_ARRIVED;
930

931
	model_net_event(net_id, "test", dest_rank, mpi_op->u.send.num_bytes, 0.0, 
932
	    sizeof(nw_message), (const void*)&remote_m, sizeof(nw_message), (const void*)&local_m, lp);
933

934
	/* isend executed, now get next MPI operation from the queue */ 
935
	if(mpi_op->op_type == CODES_WK_ISEND)
936 937 938 939
	   codes_issue_next_event(lp);
}

/* MPI collective operations */
940
static void codes_exec_mpi_col(nw_state* s, tw_lp* lp)
941 942 943 944 945 946 947 948 949 950 951 952 953 954
{
	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));
}


static void update_send_completion_queue_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
	//mpi_queue_remove_matching_op(&s->completed_isend_queue_head, &s->completed_isend_queue_tail, &m->op, SEND);
955
	if(m->u.msg_info.op_type == CODES_WK_SEND)
956
		tw_rand_reverse_unif(lp->rng);	
957

958
	if(m->u.msg_info.op_type == CODES_WK_ISEND)
959
	  {
960 961
		notify_waits_rc(s, bf, lp, m, m->u.msg_info.req_id);
		remove_req_id(&s->completed_reqs, m->u.msg_info.req_id);
962
	 }
963 964 965 966 967
}

/* completed isends are added in the list */
static void update_send_completion_queue(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp)
{