model-net-mpi-wrklds.c 44.3 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
	struct codes_workload_op* mpi_op;
57 58 59
	struct mpi_msgs_queue* next;
};

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

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;
101 102
	double elapsed_time;

103
	/* time spent in compute operations */
104 105 106 107 108 109 110 111 112 113
	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;
114 115 116

	/* FIFO for isend messages arrived on destination */
	struct mpi_queue_ptrs* arrival_queue;
117

118 119
	/* FIFO for irecv messages posted but not yet matched with send operations */
	struct mpi_queue_ptrs* pending_recvs_queue;
120

121
	/* list of pending waits (and saved pending wait for reverse computation) */
122 123 124 125
	struct pending_waits* pending_waits;

	/* List of completed send/receive requests */
	struct completed_requests* completed_reqs;
126 127
};

128 129 130 131
/* 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). */
132 133
struct nw_message
{
134 135 136
   int msg_type;
   /* for reverse computation */
   struct codes_workload_op * op;
137

138 139 140 141 142 143 144 145 146 147 148
   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;
149 150
        // for callbacks - time message was received
        double msg_send_time;
151 152 153 154 155 156 157
        int16_t req_id;   
        int tag;
     } msg_info;

     /* required for reverse computation*/
     struct 
      {
158 159 160
	int found_match;
	short matched_op;
	dumpi_req_id saved_matched_req;
161
	struct codes_workload_op* ptr_match_op;
162 163 164 165 166
	struct pending_waits* saved_pending_wait;

	double saved_send_time;
	double saved_recv_time;
	double saved_wait_time;
167 168
      } rc;
  } u;
169 170
};

171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192
/* executes MPI wait operation */
static void codes_exec_mpi_wait(nw_state* s, tw_bf* bf, nw_message* m, tw_lp* lp);

/* reverse of mpi wait function. */
static void codes_exec_mpi_wait_rc(nw_state* s, tw_bf* bf, nw_message* m, tw_lp* lp);

/* executes MPI isend and send operations */
static void codes_exec_mpi_send(nw_state* s, nw_message* m, tw_lp* lp);

/* execute MPI irecv operation */
static void codes_exec_mpi_recv(nw_state* s, nw_message* m, tw_lp* lp);

/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(nw_state* s, nw_message* m, tw_lp* lp);

/* execute the computational delay */
static void codes_exec_comp_delay(nw_state* s, nw_message* m, tw_lp* lp);

/* execute collective operation, currently only skips these operations. */
static void codes_exec_mpi_col(nw_state* s, nw_message* m, tw_lp* lp);

/* gets the next MPI operation from the network-workloads API. */
193 194
static void get_next_mpi_operation(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);

195 196 197 198 199 200 201
/* 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);

///////////////////// HELPER FUNCTIONS FOR MPI MESSAGE QUEUE HANDLING ///////////////
202 203 204 205 206 207 208 209 210 211 212 213
/* upon arrival of local completion message, inserts operation in completed send queue */
static void update_send_completion_queue(nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);

/* reverse of the above function */
static void update_send_completion_queue_rc(nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);

/* upon arrival of an isend operation, updates the arrival queue of the network */
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);

214 215 216 217
/* 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);
static void update_message_time_rc(nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);

218
/* insert MPI operation in the waiting queue*/
219
static void mpi_pending_queue_insert_op(struct mpi_queue_ptrs* mpi_queue, struct codes_workload_op* mpi_op);
220

221
/* remove completed request IDs from the queue for reuse. Reverse of above function. */
222
static void remove_req_id(struct completed_requests** requests, int16_t req_id);
223

224
/* remove MPI operation from the waiting queue.*/
225
static int mpi_queue_remove_matching_op(nw_state* s, tw_lp* lp, struct mpi_queue_ptrs* mpi_queue, nw_message * m);
226

227
/* remove the tail of the MPI operation from waiting queue */
228
static int mpi_queue_remove_tail(tw_lpid lpid, struct mpi_queue_ptrs* mpi_queue);
229

230
/* insert completed MPI requests in the queue. */
231
static void mpi_completed_queue_insert_op(struct completed_requests** mpi_completed_queue, dumpi_req_id req_id);
232

233 234
/* notifies the wait operations (if any) about the completed receives and sends requests. */
static int notify_waits(nw_state* s, tw_bf* bf, tw_lp* lp, nw_message* m, dumpi_req_id req_id);
235

236 237
/* reverse 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);
238

239 240
/* conversion from seconds to eanaoseconds */
static tw_stime s_to_ns(tw_stime ns);
241

242 243 244 245 246
/* 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);
}
247

248 249 250 251 252 253 254 255 256 257 258 259
/* 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;
}

260
/* helper function: counts number of elements in the queue */
261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288
static int numQueue(struct mpi_queue_ptrs* mpi_queue)
{
	struct mpi_msgs_queue* tmp = malloc(sizeof(struct mpi_msgs_queue)); 
	assert(tmp);

	tmp = mpi_queue->queue_head;
	int count = 0;

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

/* 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)
	{
289
		if(tmp->mpi_op->op_type == CODES_WK_SEND || tmp->mpi_op->op_type == CODES_WK_ISEND)
290 291 292
			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);
293
		else if(tmp->mpi_op->op_type == CODES_WK_IRECV || tmp->mpi_op->op_type == CODES_WK_RECV)
294 295 296 297 298 299 300 301 302 303 304
			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 */
305
static void mpi_queue_update(struct mpi_queue_ptrs* mpi_queue, struct codes_workload_op* mpi_op, int pos)
306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341
{
	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;
}

342
/* prints the elements of a queue (for debugging purposes). */
343 344 345 346
static void printCompletedQueue(nw_state* s, tw_lp* lp)
{
	   if(TRACE == lp->gid)
	   {
347
	   	printf("\n %lf contents of completed operations queue ", tw_now(lp));
348 349 350 351 352 353 354 355 356
	   	struct completed_requests* current = s->completed_reqs;
	   	while(current)
	    	{
			printf(" %d ",current->req_id);
			current = current->next;
	   	}
	   }
}

357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373
/* 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);
  printCompletedQueue(s, lp);
374
  if(m->u.rc.matched_op == 1)
375 376
	s->pending_waits->num_completed--;
   /* if a wait-elem exists, it means the request ID has been matched*/
377
   if(m->u.rc.matched_op == 2) 
378 379 380 381 382 383
    {
	if(lp->gid == TRACE)
	{
		printf("\n %lf matched req id %d ", tw_now(lp), completed_req);
		printCompletedQueue(s, lp);
	}
384 385
        struct pending_waits* wait_elem = m->u.rc.saved_pending_wait;
	s->wait_time = m->u.rc.saved_wait_time;
386 387 388 389 390 391 392 393 394 395 396 397 398 399
	int count = wait_elem->mpi_op->u.waits.count; 

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

	wait_elem->num_completed--;	
	s->pending_waits = wait_elem;
	tw_rand_reverse_unif(lp->rng);

   }
}

/* 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)
400 401 402 403 404 405 406
{
	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;
407
	m->u.rc.matched_op = 0;
408 409 410
	
	if(lp->gid == TRACE)
		printf("\n %lf notify waits req id %d ", tw_now(lp), completed_req);
411 412 413

	if(!wait_elem)
		return 0;
414

415 416
	int op_type = wait_elem->mpi_op->op_type;

417
	if(op_type == CODES_WK_WAIT)
418 419 420
	{
		if(wait_elem->mpi_op->u.wait.req_id == completed_req)	
		  {
421
			m->u.rc.saved_wait_time = s->wait_time;
422
			s->wait_time += (tw_now(lp) - wait_elem->start_time);
423
                        remove_req_id(&s->completed_reqs, completed_req);
424
	
425
			m->u.rc.saved_pending_wait = wait_elem;			
426 427 428 429 430 431
			s->pending_waits = NULL;
			codes_issue_next_event(lp);	
			return 0;
		 }
	}
	else
432
	if(op_type == CODES_WK_WAITALL)
433
	{
434 435
	   int required_count = wait_elem->mpi_op->u.waits.count;
	  for(i = 0; i < required_count; i++)
436 437
	   {
	    if(wait_elem->mpi_op->u.waits.req_ids[i] == completed_req)
438 439 440
		{
			if(lp->gid == TRACE)
				printCompletedQueue(s, lp);
441
			m->u.rc.matched_op = 1;
442
			wait_elem->num_completed++;	
443
		}
444 445
	   }
	   
446
	    if(wait_elem->num_completed == required_count)
447
	     {
448 449 450 451 452
		if(lp->gid == TRACE)
		{
			printf("\n %lf req %d completed %d", tw_now(lp), completed_req, wait_elem->num_completed);
			printCompletedQueue(s, lp);
		}
453 454
		m->u.rc.matched_op = 2;
		m->u.rc.saved_wait_time = s->wait_time;
455
		s->wait_time += (tw_now(lp) - wait_elem->start_time);
456
		m->u.rc.saved_pending_wait = wait_elem;
457
		s->pending_waits = NULL; 
458
		for(i = 0; i < required_count; i++)
459 460 461
			remove_req_id(&s->completed_reqs, wait_elem->mpi_op->u.waits.req_ids[i]);	
		codes_issue_next_event(lp); //wait completed
	    }
462
       }
463 464 465
	return 0;
}

466 467 468 469 470 471 472 473 474 475
/* 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;
	return;
     }
   else
    {
476
 	mpi_completed_queue_insert_op(&s->completed_reqs, m->op->u.wait.req_id);	
477 478 479
	tw_rand_reverse_unif(lp->rng);		
    }
}
480 481

/* execute MPI wait operation */
482
static void codes_exec_mpi_wait(nw_state* s, tw_bf* bf, nw_message* m, tw_lp* lp)
483
{
Jonathan Jenkins's avatar
Jonathan Jenkins committed
484 485 486 487 488 489 490 491 492 493 494 495 496
    /* check in the completed receives queue if the request ID has already been completed.*/
    assert(!s->pending_waits);
    dumpi_req_id req_id = m->op->u.wait.req_id;

    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;
497 498
    }

Jonathan Jenkins's avatar
Jonathan Jenkins committed
499
    /* If not, add the wait operation in the pending 'waits' list. */
500
    struct pending_waits* wait_op = malloc(sizeof(struct pending_waits));
Jonathan Jenkins's avatar
Jonathan Jenkins committed
501 502 503
    wait_op->mpi_op = m->op;
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
504
    s->pending_waits = wait_op;
505 506
}

507
static void codes_exec_mpi_wait_all_rc(nw_state* s, tw_bf* bf, nw_message* m, tw_lp* lp)
508
{
509 510
   if(lp->gid == TRACE)
  {
511
   printf("\n %lf codes exec mpi waitall reverse %d ", tw_now(lp), m->u.rc.found_match);
512 513
   printCompletedQueue(s, lp); 
  } 
514
  if(m->u.rc.found_match)
515 516
    {
   	int i;
517
	int count = m->op->u.waits.count;
518 519 520 521
	dumpi_req_id req_id[count];

	for( i = 0; i < count; i++)
	{
522
		req_id[i] = m->op->u.waits.req_ids[i];
523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539
		mpi_completed_queue_insert_op(&s->completed_reqs, req_id[i]);
	}
	tw_rand_reverse_unif(lp->rng);
    }
    else
    {
	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));
    }
}
static void codes_exec_mpi_wait_all(nw_state* s, tw_bf* bf, nw_message* m, tw_lp* lp)
{
  //assert(!s->pending_waits);
540
  int count = m->op->u.waits.count;
541 542 543 544 545 546 547
  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)
    {
548
  	printf(" \n (%lf) MPI waitall posted %d count", tw_now(lp), m->op->u.waits.count);
549
	for(i = 0; i < count; i++)
550
		printf(" %d ", (int)m->op->u.waits.req_ids[i]);
551 552 553 554 555 556
   	printCompletedQueue(s, lp);	 
   }
  while(current) 
   {
	  for(i = 0; i < count; i++)
	   {
557
	     req_id[i] = m->op->u.waits.req_ids[i];
558 559 560 561 562 563 564
	     if(req_id[i] == current->req_id)
 		 num_completed++;
   	  }
	 current = current->next;
   }

  if(TRACE== lp->gid)
565
	  printf("\n %lf Num completed %d count %d ", tw_now(lp), num_completed, count);
566

567
  m->u.rc.found_match = 0;
568
  if(count == num_completed)
569
  {
570
	m->u.rc.found_match = 1;
571
	for( i = 0; i < count; i++)	
572
		remove_req_id(&s->completed_reqs, req_id[i]);
573

574 575 576 577 578 579
	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));
580
	  wait_op->mpi_op = m->op;  
581
	  wait_op->num_completed = num_completed;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
582
	  wait_op->start_time = tw_now(lp);
583 584 585 586 587 588 589 590 591 592
	  s->pending_waits = wait_op;
  }
}

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

	if(!current)
593 594 595
		tw_error(TW_LOC, "\n REQ ID DOES NOT EXIST");
	
       if(current->req_id == req_id)
596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617
	{
		*mpi_completed_queue = current->next;
		free(current);
		return;
	}
	
	struct completed_requests* elem;
	while(current->next)
	{
	   elem = current->next;
	   if(elem->req_id == req_id)	
	     {
		current->next = elem->next;
		free(elem);
		return;
	     }
	   current = current->next;	
	}
	return;
}

/* inserts mpi operation in the completed requests queue */
618
static void mpi_completed_queue_insert_op(struct completed_requests** mpi_completed_queue, dumpi_req_id req_id)
619 620 621 622 623 624 625 626
{
	struct completed_requests* reqs = malloc(sizeof(struct completed_requests));
	assert(reqs);

	reqs->req_id = req_id;

	if(!(*mpi_completed_queue))	
	{
627
			reqs->next = NULL;
628
			*mpi_completed_queue = reqs;
629
			return;
630 631 632
	}
	reqs->next = *mpi_completed_queue;
	*mpi_completed_queue = reqs;
633
	return;
634 635
}

636
/* 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. */
637
static void mpi_pending_queue_insert_op(struct mpi_queue_ptrs* mpi_queue, struct codes_workload_op* mpi_op)
638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658
{
	/* insert mpi operation */
	struct mpi_msgs_queue* elem = malloc(sizeof(struct mpi_msgs_queue));
	assert(elem);

	elem->mpi_op = mpi_op;
     	elem->next = NULL;

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

	if(mpi_queue->queue_tail)
	    mpi_queue->queue_tail->next = elem;
	
        mpi_queue->queue_tail = elem;
	mpi_queue->num_elems++;

	return;
}

/* match the send/recv operations */
659
static int match_receive(nw_state* s, tw_lp* lp, tw_lpid lpid, struct codes_workload_op* op1, struct codes_workload_op* op2)
660
{
661 662 663 664 665 666 667 668 669
        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);
670
                        s->recv_time += tw_now(lp) - op1->sim_start_time;
671 672 673 674
                        mpi_completed_queue_insert_op(&s->completed_reqs, op1->u.recv.req_id);
                        return 1;
                   }
        return -1;
675 676 677
}

/* used for reverse computation. removes the tail of the queue */
678
static int mpi_queue_remove_tail(tw_lpid lpid, struct mpi_queue_ptrs* mpi_queue)
679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712
{
	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. */
713
static int mpi_queue_remove_matching_op(nw_state* s, tw_lp* lp, struct mpi_queue_ptrs* mpi_queue, nw_message * m)
714
{
715 716
       struct codes_workload_op * mpi_op = m->op;
 
717 718 719 720 721 722 723 724
	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 */
725
	int rcv_val = 0;
726
	if(mpi_op->op_type == CODES_WK_SEND || mpi_op->op_type == CODES_WK_ISEND)
727
	  {
728
		rcv_val = match_receive(s, lp, lp->gid, tmp->mpi_op, mpi_op);
729
		m->u.rc.saved_matched_req = tmp->mpi_op->u.recv.req_id;  
730
	 }
731
	else if(mpi_op->op_type == CODES_WK_RECV || mpi_op->op_type == CODES_WK_IRECV)
732
	  {
733
		rcv_val = match_receive(s, lp, lp->gid, mpi_op, tmp->mpi_op);
734
	  	m->u.rc.saved_matched_req = mpi_op->u.recv.req_id;
735 736
	  }
	if(rcv_val >= 0)
737
	{
738 739
		/* TODO: fix RC */
		/*memcpy(&m->u.rc.ptr_match_op, &tmp->mpi_op, sizeof(struct codes_workload_op));*/
740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761
		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;
762
	   
763
	    if(mpi_op->op_type == CODES_WK_SEND || mpi_op->op_type == CODES_WK_ISEND)
764
	     {
765
		rcv_val = match_receive(s, lp, lp->gid, elem->mpi_op, mpi_op);
766
	     	m->u.rc.saved_matched_req = elem->mpi_op->u.recv.req_id; 
767
	     }
768
	    else if(mpi_op->op_type == CODES_WK_RECV || mpi_op->op_type == CODES_WK_IRECV)
769
	     {
770
		rcv_val = match_receive(s, lp, lp->gid, mpi_op, elem->mpi_op);
771
		m->u.rc.saved_matched_req = mpi_op->u.recv.req_id;
772 773
	     }
   	     if(rcv_val >= 0)
774
		{
775 776
		    /* TODO: fix RC */
		    /*memcpy(&m->u.rc.ptr_match_op, &elem->mpi_op, sizeof(struct codes_workload_op));*/
777 778
		    if(elem == mpi_queue->queue_tail)
			mpi_queue->queue_tail = tmp;
779
		    
780 781 782 783
		    tmp->next = elem->next;

		    free(elem);
		    mpi_queue->num_elems--;
784
		
785 786 787
		    return indx;
		}
	   tmp = tmp->next;
788
        }
789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809
	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 */
static void codes_exec_comp_delay(nw_state* s, nw_message* m, tw_lp* lp)
{
810
	struct codes_workload_op* mpi_op = m->op;
811 812 813 814
	tw_event* e;
	tw_stime ts;
	nw_message* msg;

815 816 817 818 819 820 821 822 823
        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);
824 825 826 827 828 829 830 831 832
	
	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 */
833
static void codes_exec_mpi_recv_rc(nw_state* s, nw_message* m, tw_lp* lp)
834
{
835 836 837
	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)
838
	  {
839 840 841
		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);
842
		tw_rand_reverse_unif(lp->rng);
843
	  }
844
	else if(m->u.rc.found_match < 0)
845
	    {
846 847
		mpi_queue_remove_tail(lp->gid, s->pending_recvs_queue);
		if(m->op->op_type == CODES_WK_IRECV)
848
			tw_rand_reverse_unif(lp->rng);
849 850 851 852
	    }
}

/* Execute MPI Irecv operation (non-blocking receive) */ 
853
static void codes_exec_mpi_recv(nw_state* s, nw_message* m, tw_lp* lp)
854 855 856 857 858
{
/* 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. */

859 860
	m->u.rc.saved_recv_time = s->recv_time;
	struct codes_workload_op* mpi_op = m->op;
861
	mpi_op->sim_start_time = tw_now(lp);
862
	num_bytes_recvd += mpi_op->u.recv.num_bytes;
863 864

	if(lp->gid == TRACE)
865
		printf("\n %lf codes exec mpi recv req id %d", tw_now(lp), (int)mpi_op->u.recv.req_id);
866 867
	
	dumpi_req_id req_id;
868
	int found_matching_sends = mpi_queue_remove_matching_op(s, lp, s->arrival_queue, m);
869 870 871
	
	/* save the req id inserted in the completed queue for reverse computation. */
	//m->matched_recv = req_id;
872 873 874

	if(found_matching_sends < 0)
	  {
875
		m->u.rc.found_match = -1;
876 877 878
		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. */
879
		if(mpi_op->op_type == CODES_WK_IRECV)
880 881 882 883 884 885
		   {
			codes_issue_next_event(lp);	
			return;
		   }
		else
			printf("\n CODES MPI RECV OPERATION!!! ");
886
	  }
887
	else
888 889
	  {
		/*if(lp->gid == TRACE)
890 891 892 893
			printf("\n Matched after removing: arrival queue num_elems %d ", s->arrival_queue->num_elems);*/
		/* update completed requests list */
		//int count_after = numQueue(s->arrival_queue);
		//assert(count_before == (count_after+1));
894
	   	m->u.rc.found_match = found_matching_sends;
895
		codes_issue_next_event(lp); 
896 897 898 899
	 }
}

/* executes MPI send and isend operations */
900
static void codes_exec_mpi_send(nw_state* s, nw_message * m, tw_lp* lp)
901
{
902
        struct codes_workload_op * mpi_op = m->op; 
903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930
	/* 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 = malloc(sizeof(nw_message));
	nw_message* remote_m = malloc(sizeof(nw_message));
	assert(local_m && remote_m);

931 932 933 934 935 936 937 938 939 940
        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;

        memcpy(remote_m, local_m, sizeof(nw_message));
941
	remote_m->msg_type = MPI_SEND_ARRIVED;
942

943 944
	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);
945

946 947 948
	/*if(TRACE == lp->gid)	
		printf("\n !!! %lf send req id %d dest %d nw_message %d ", tw_now(lp), (int)mpi_op->u.send.req_id, (int)dest_rank, sizeof(nw_message));
	*/
949
	/* isend executed, now get next MPI operation from the queue */ 
950
	if(mpi_op->op_type == CODES_WK_ISEND)
951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969
	   codes_issue_next_event(lp);
}

/* MPI collective operations */
static void codes_exec_mpi_col(nw_state* s, nw_message* m, tw_lp* lp)
{
	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);
970
	if(m->u.msg_info.op_type == CODES_WK_SEND)
971
		tw_rand_reverse_unif(lp->rng);	
972

973
	if(m->u.msg_info.op_type == CODES_WK_ISEND)
974
	  {
975 976
		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);
977
	 }
978 979 980 981 982
}

/* 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)
{
983
	if(TRACE == lp->gid)
984 985
		printf("\n %lf isend operation completed req id %d ", tw_now(lp), m->u.msg_info.req_id);
	if(m->u.msg_info.op_type == CODES_WK_ISEND)
986
	   {	
987 988
		mpi_completed_queue_insert_op(&s->completed_reqs, m->u.msg_info.req_id);
	   	notify_waits(s, bf, lp, m, m->u.msg_info.req_id);
989 990
	   }  
	
991
	/* blocking send operation */
992
	if(m->u.msg_info.op_type == CODES_WK_SEND)
993 994 995 996 997 998 999 1000
		codes_issue_next_event(lp);	

	 return;
}

/* reverse handler for updating arrival queue function */
static void update_arrival_queue_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
1001 1002 1003
	s->recv_time = m->u.rc.saved_recv_time;

        codes_local_latency_reverse(lp);
1004

1005
	if(m->u.rc.found_match >= 0)
1006
	{
1007 1008
		// TODO: Modify for recvs
		if(lp->gid == TRACE)
1009 1010 1011
			printf("\n %lf reverse-- update arrival queue req ID %d", tw_now(lp), (int) m->u.rc.saved_matched_req);
		dumpi_req_id req_id = m->u.rc.saved_matched_req;
		notify_waits_rc(s, bf, lp, m, m->u.rc.saved_matched_req);
1012
		//int count = numQueue(s->pending_recvs_queue);
1013 1014
		mpi_queue_update(s->pending_recvs_queue, m->u.rc.ptr_match_op, m->u.rc.found_match);
		remove_req_id(&s->completed_reqs, m->u.rc.saved_matched_req);
1015
	
1016 1017 1018
		/*if(lp->gid == TRACE)
			printf("\n Reverse: after adding pending recvs queue %d ", s->pending_recvs_queue->num_elems);*/
	}
1019
	else if(m->u.rc.found_match < 0)
1020
	{
1021
		mpi_queue_remove_tail(lp->gid, s->arrival_queue);	
1022 1023