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

50
/* stores pointers of pending MPI operations to be matched with their respective sends/receives. */
51 52
struct mpi_msgs_queue
{
53
	struct codes_workload_op* mpi_op;
54 55 56
	struct mpi_msgs_queue* next;
};

57
/* stores request IDs of completed MPI operations (Isends or Irecvs) */
58 59 60 61 62 63
struct completed_requests
{
	dumpi_req_id req_id;
	struct completed_requests* next;
};

64
/* for wait operations, store the pending operation and number of completed waits so far. */
65 66
struct pending_waits
{
67
	struct codes_workload_op* mpi_op;
68 69 70 71
	int num_completed;
	tw_stime start_time;
};

72
/* 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). */
73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91
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;
92 93 94
	unsigned long num_wait;
	unsigned long num_waitall;
	unsigned long num_waitsome;
95 96

	/* time spent by the LP in executing the app trace*/
97
	double start_time;
98 99
	double elapsed_time;

100
	/* time spent in compute operations */
101 102 103 104 105 106 107 108 109 110
	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;
111 112 113

	/* FIFO for isend messages arrived on destination */
	struct mpi_queue_ptrs* arrival_queue;
114

115 116
	/* FIFO for irecv messages posted but not yet matched with send operations */
	struct mpi_queue_ptrs* pending_recvs_queue;
117

118
	/* list of pending waits (and saved pending wait for reverse computation) */
119 120 121 122
	struct pending_waits* pending_waits;

	/* List of completed send/receive requests */
	struct completed_requests* completed_reqs;
123 124
};

125 126 127 128
/* 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). */
129 130
struct nw_message
{
131 132 133
   int msg_type;
   /* for reverse computation */
   struct codes_workload_op * op;
134

135 136 137 138 139 140 141 142 143 144 145
   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;
146 147
        // for callbacks - time message was received
        double msg_send_time;
148 149 150 151 152 153 154
        int16_t req_id;   
        int tag;
     } msg_info;

     /* required for reverse computation*/
     struct 
      {
155 156 157
	int found_match;
	short matched_op;
	dumpi_req_id saved_matched_req;
158
	struct codes_workload_op* ptr_match_op;
159 160 161 162 163
	struct pending_waits* saved_pending_wait;

	double saved_send_time;
	double saved_recv_time;
	double saved_wait_time;
164 165
      } rc;
  } u;
166 167
};

168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189
/* 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. */
190 191
static void get_next_mpi_operation(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);

192 193 194 195 196 197 198
/* 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 ///////////////
199 200 201 202 203 204 205 206 207 208 209 210
/* 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);

211 212 213 214
/* 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);

215
/* insert MPI operation in the waiting queue*/
216
static void mpi_pending_queue_insert_op(struct mpi_queue_ptrs* mpi_queue, struct codes_workload_op* mpi_op);
217

218
/* remove completed request IDs from the queue for reuse. Reverse of above function. */
219
static void remove_req_id(struct completed_requests** requests, int16_t req_id);
220

221
/* remove MPI operation from the waiting queue.*/
222
static int mpi_queue_remove_matching_op(nw_state* s, tw_lp* lp, struct mpi_queue_ptrs* mpi_queue, nw_message * m);
223

224
/* remove the tail of the MPI operation from waiting queue */
225
static int mpi_queue_remove_tail(tw_lpid lpid, struct mpi_queue_ptrs* mpi_queue);
226

227
/* insert completed MPI requests in the queue. */
228
static void mpi_completed_queue_insert_op(struct completed_requests** mpi_completed_queue, dumpi_req_id req_id);
229

230 231
/* 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);
232

233 234
/* 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);
235

236 237
/* conversion from seconds to eanaoseconds */
static tw_stime s_to_ns(tw_stime ns);
238

239 240 241 242 243
/* 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);
}
244

245 246 247 248 249 250 251 252 253 254 255 256
/* 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;
}

257
/* helper function: counts number of elements in the queue */
258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285
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)
	{
286
		if(tmp->mpi_op->op_type == CODES_WK_SEND || tmp->mpi_op->op_type == CODES_WK_ISEND)
287 288 289
			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);
290
		else if(tmp->mpi_op->op_type == CODES_WK_IRECV || tmp->mpi_op->op_type == CODES_WK_RECV)
291 292 293 294 295 296 297 298 299 300 301
			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 */
302
static void mpi_queue_update(struct mpi_queue_ptrs* mpi_queue, struct codes_workload_op* mpi_op, int pos)
303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338
{
	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;
}

339
/* prints the elements of a queue (for debugging purposes). */
340 341 342 343
static void printCompletedQueue(nw_state* s, tw_lp* lp)
{
	   if(TRACE == lp->gid)
	   {
344
	   	printf("\n %lf contents of completed operations queue ", tw_now(lp));
345 346 347 348 349 350 351 352 353
	   	struct completed_requests* current = s->completed_reqs;
	   	while(current)
	    	{
			printf(" %d ",current->req_id);
			current = current->next;
	   	}
	   }
}

354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370
/* 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);
371
  if(m->u.rc.matched_op == 1)
372 373
	s->pending_waits->num_completed--;
   /* if a wait-elem exists, it means the request ID has been matched*/
374
   if(m->u.rc.matched_op == 2) 
375 376 377 378 379 380
    {
	if(lp->gid == TRACE)
	{
		printf("\n %lf matched req id %d ", tw_now(lp), completed_req);
		printCompletedQueue(s, lp);
	}
381 382
        struct pending_waits* wait_elem = m->u.rc.saved_pending_wait;
	s->wait_time = m->u.rc.saved_wait_time;
383 384 385 386 387 388 389 390 391 392 393 394 395 396
	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)
397 398 399 400 401 402 403
{
	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;
404
	m->u.rc.matched_op = 0;
405 406 407
	
	if(lp->gid == TRACE)
		printf("\n %lf notify waits req id %d ", tw_now(lp), completed_req);
408 409 410

	if(!wait_elem)
		return 0;
411

412 413
	int op_type = wait_elem->mpi_op->op_type;

414
	if(op_type == CODES_WK_WAIT)
415 416 417
	{
		if(wait_elem->mpi_op->u.wait.req_id == completed_req)	
		  {
418
			m->u.rc.saved_wait_time = s->wait_time;
419
			s->wait_time += (tw_now(lp) - wait_elem->start_time);
420
                        remove_req_id(&s->completed_reqs, completed_req);
421
	
422
			m->u.rc.saved_pending_wait = wait_elem;			
423 424 425 426 427 428
			s->pending_waits = NULL;
			codes_issue_next_event(lp);	
			return 0;
		 }
	}
	else
429
	if(op_type == CODES_WK_WAITALL)
430
	{
431 432
	   int required_count = wait_elem->mpi_op->u.waits.count;
	  for(i = 0; i < required_count; i++)
433 434
	   {
	    if(wait_elem->mpi_op->u.waits.req_ids[i] == completed_req)
435 436 437
		{
			if(lp->gid == TRACE)
				printCompletedQueue(s, lp);
438
			m->u.rc.matched_op = 1;
439
			wait_elem->num_completed++;	
440
		}
441 442
	   }
	   
443
	    if(wait_elem->num_completed == required_count)
444
	     {
445 446 447 448 449
		if(lp->gid == TRACE)
		{
			printf("\n %lf req %d completed %d", tw_now(lp), completed_req, wait_elem->num_completed);
			printCompletedQueue(s, lp);
		}
450 451
		m->u.rc.matched_op = 2;
		m->u.rc.saved_wait_time = s->wait_time;
452
		s->wait_time += (tw_now(lp) - wait_elem->start_time);
453
		m->u.rc.saved_pending_wait = wait_elem;
454
		s->pending_waits = NULL; 
455
		for(i = 0; i < required_count; i++)
456 457 458
			remove_req_id(&s->completed_reqs, wait_elem->mpi_op->u.waits.req_ids[i]);	
		codes_issue_next_event(lp); //wait completed
	    }
459
       }
460 461 462
	return 0;
}

463 464 465 466 467 468 469 470 471 472
/* 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
    {
473
 	mpi_completed_queue_insert_op(&s->completed_reqs, m->op->u.wait.req_id);	
474 475 476
	tw_rand_reverse_unif(lp->rng);		
    }
}
477 478

/* execute MPI wait operation */
479
static void codes_exec_mpi_wait(nw_state* s, tw_bf* bf, nw_message* m, tw_lp* lp)
480
{
Jonathan Jenkins's avatar
Jonathan Jenkins committed
481 482 483 484 485 486 487 488 489 490 491 492 493
    /* 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;
494 495
    }

Jonathan Jenkins's avatar
Jonathan Jenkins committed
496
    /* If not, add the wait operation in the pending 'waits' list. */
497
    struct pending_waits* wait_op = malloc(sizeof(struct pending_waits));
Jonathan Jenkins's avatar
Jonathan Jenkins committed
498 499 500
    wait_op->mpi_op = m->op;
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
501
    s->pending_waits = wait_op;
502 503
}

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

	for( i = 0; i < count; i++)
	{
519
		req_id[i] = m->op->u.waits.req_ids[i];
520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536
		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);
537
  int count = m->op->u.waits.count;
538 539 540 541 542 543 544
  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)
    {
545
  	printf(" \n (%lf) MPI waitall posted %d count", tw_now(lp), m->op->u.waits.count);
546
	for(i = 0; i < count; i++)
547
		printf(" %d ", (int)m->op->u.waits.req_ids[i]);
548 549 550 551 552 553
   	printCompletedQueue(s, lp);	 
   }
  while(current) 
   {
	  for(i = 0; i < count; i++)
	   {
554
	     req_id[i] = m->op->u.waits.req_ids[i];
555 556 557 558 559 560 561
	     if(req_id[i] == current->req_id)
 		 num_completed++;
   	  }
	 current = current->next;
   }

  if(TRACE== lp->gid)
562
	  printf("\n %lf Num completed %d count %d ", tw_now(lp), num_completed, count);
563

564
  m->u.rc.found_match = 0;
565
  if(count == num_completed)
566
  {
567
	m->u.rc.found_match = 1;
568
	for( i = 0; i < count; i++)	
569
		remove_req_id(&s->completed_reqs, req_id[i]);
570

571 572 573 574 575 576
	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));
577
	  wait_op->mpi_op = m->op;  
578
	  wait_op->num_completed = num_completed;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
579
	  wait_op->start_time = tw_now(lp);
580 581 582 583 584 585 586 587 588 589
	  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)
590 591 592
		tw_error(TW_LOC, "\n REQ ID DOES NOT EXIST");
	
       if(current->req_id == req_id)
593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614
	{
		*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 */
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
			reqs->next = NULL;
625
			*mpi_completed_queue = reqs;
626
			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
static void mpi_pending_queue_insert_op(struct mpi_queue_ptrs* mpi_queue, struct codes_workload_op* mpi_op)
635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655
{
	/* 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 */
656
static int match_receive(nw_state* s, tw_lp* lp, tw_lpid lpid, struct codes_workload_op* op1, struct codes_workload_op* op2)
657
{
658 659 660 661 662 663 664 665 666
        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);
667
                        s->recv_time += tw_now(lp) - op1->sim_start_time;
668 669 670 671
                        mpi_completed_queue_insert_op(&s->completed_reqs, op1->u.recv.req_id);
                        return 1;
                   }
        return -1;
672 673 674
}

/* used for reverse computation. removes the tail of the queue */
675
static int mpi_queue_remove_tail(tw_lpid lpid, struct mpi_queue_ptrs* mpi_queue)
676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709
{
	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. */
710
static int mpi_queue_remove_matching_op(nw_state* s, tw_lp* lp, struct mpi_queue_ptrs* mpi_queue, nw_message * m)
711
{
712 713
       struct codes_workload_op * mpi_op = m->op;
 
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
	{
735 736
		/* TODO: fix RC */
		/*memcpy(&m->u.rc.ptr_match_op, &tmp->mpi_op, sizeof(struct codes_workload_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
		rcv_val = match_receive(s, lp, lp->gid, mpi_op, elem->mpi_op);
768
		m->u.rc.saved_matched_req = mpi_op->u.recv.req_id;
769 770
	     }
   	     if(rcv_val >= 0)
771
		{
772 773
		    /* TODO: fix RC */
		    /*memcpy(&m->u.rc.ptr_match_op, &elem->mpi_op, sizeof(struct codes_workload_op));*/
774 775
		    if(elem == mpi_queue->queue_tail)
			mpi_queue->queue_tail = tmp;
776
		    
777 778 779 780
		    tmp->next = elem->next;

		    free(elem);
		    mpi_queue->num_elems--;
781
		
782 783 784
		    return indx;
		}
	   tmp = tmp->next;
785
        }
786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806
	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)
{
807
	struct codes_workload_op* mpi_op = m->op;
808 809 810 811
	tw_event* e;
	tw_stime ts;
	nw_message* msg;

812 813
	s->compute_time += s_to_ns(mpi_op->u.delay.seconds);
	ts = s_to_ns(mpi_op->u.delay.seconds) + g_tw_lookahead + 0.1;
814 815 816 817 818 819 820 821 822 823
	ts += 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 */
824
static void codes_exec_mpi_recv_rc(nw_state* s, nw_message* m, tw_lp* lp)
825
{
826 827 828
	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)
829
	  {
830 831 832
		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);
833
		tw_rand_reverse_unif(lp->rng);
834
	  }
835
	else if(m->u.rc.found_match < 0)
836
	    {
837 838
		mpi_queue_remove_tail(lp->gid, s->pending_recvs_queue);
		if(m->op->op_type == CODES_WK_IRECV)
839
			tw_rand_reverse_unif(lp->rng);
840 841 842 843
	    }
}

/* Execute MPI Irecv operation (non-blocking receive) */ 
844
static void codes_exec_mpi_recv(nw_state* s, nw_message* m, tw_lp* lp)
845 846 847 848 849
{
/* 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. */

850 851
	m->u.rc.saved_recv_time = s->recv_time;
	struct codes_workload_op* mpi_op = m->op;
852
	mpi_op->sim_start_time = tw_now(lp);
853
	num_bytes_recvd += mpi_op->u.recv.num_bytes;
854 855

	if(lp->gid == TRACE)
856
		printf("\n %lf codes exec mpi recv req id %d", tw_now(lp), (int)mpi_op->u.recv.req_id);
857 858
	
	dumpi_req_id req_id;
859
	int found_matching_sends = mpi_queue_remove_matching_op(s, lp, s->arrival_queue, m);
860 861 862
	
	/* save the req id inserted in the completed queue for reverse computation. */
	//m->matched_recv = req_id;
863 864 865

	if(found_matching_sends < 0)
	  {
866
		m->u.rc.found_match = -1;
867 868 869
		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. */
870
		if(mpi_op->op_type == CODES_WK_IRECV)
871 872 873 874 875 876
		   {
			codes_issue_next_event(lp);	
			return;
		   }
		else
			printf("\n CODES MPI RECV OPERATION!!! ");
877
	  }
878
	else
879 880
	  {
		/*if(lp->gid == TRACE)
881 882 883 884
			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));
885
	   	m->u.rc.found_match = found_matching_sends;
886
		codes_issue_next_event(lp); 
887 888 889 890
	 }
}

/* executes MPI send and isend operations */
891
static void codes_exec_mpi_send(nw_state* s, nw_message * m, tw_lp* lp)
892
{
893
        struct codes_workload_op * mpi_op = m->op; 
894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921
	/* 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);

922 923 924 925 926 927 928 929 930 931
        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));
932
	remote_m->msg_type = MPI_SEND_ARRIVED;
933

934 935
	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);
936

937 938 939
	/*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));
	*/
940
	/* isend executed, now get next MPI operation from the queue */ 
941
	if(mpi_op->op_type == CODES_WK_ISEND)
942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960
	   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);
961
	if(m->u.msg_info.op_type == CODES_WK_SEND)
962
		tw_rand_reverse_unif(lp->rng);	
963

964
	if(m->u.msg_info.op_type == CODES_WK_ISEND)
965
	  {
966 967
		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);
968
	 }
969 970 971 972 973
}

/* 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)
{
974
	if(TRACE == lp->gid)
975 976
		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)
977
	   {	
978 979
		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);
980 981
	   }  
	
982
	/* blocking send operation */
983
	if(m->u.msg_info.op_type == CODES_WK_SEND)
984 985 986 987 988 989 990 991
		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)
{
992 993 994
	s->recv_time = m->u.rc.saved_recv_time;

        codes_local_latency_reverse(lp);
995

996
	if(m->u.rc.found_match >= 0)
997
	{
998 999
		// TODO: Modify for recvs
		if(lp->gid == TRACE)
1000 1001 1002
			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);
1003
		//int count = numQueue(s->pending_recvs_queue);
1004 1005
		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);
1006
	
1007 1008 1009
		/*if(lp->gid == TRACE)
			printf("\n Reverse: after adding pending recvs queue %d ", s->pending_recvs_queue->num_elems);*/
	}
1010
	else if(m->u.rc.found_match < 0)
1011
	{
1012
		mpi_queue_remove_tail(lp->gid, s->arrival_queue);	
1013 1014 1015 1016 1017 1018 1019 1020
		/*if(lp->gid == TRACE)
			printf("\n Reverse: after removing arrivals queue %d ", s->arrival_queue->num_elems);*/
	}
}

/* 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)
{
1021 1022
	//int count_before = numQueue(s->pending_recvs_queue);