model-net-mpi-wrklds.c 45.2 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

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

23 24 25 26 27 28
/* Doing LP IO*/
static char lp_io_dir[256] = {'\0'};
static lp_io_handle io_handle;
static unsigned int lp_io_use_suffix = 0;
static int do_lp_io = 0;

29 30
typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
31
typedef int16_t dumpi_req_id;
32 33 34 35 36 37

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;
38 39
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;
40 41 42 43 44

/* 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;

45 46 47
/* runtime option for disabling computation time simulation */
static int disable_delay = 0;

48 49 50
/* 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). */
51 52 53 54
enum MPI_NW_EVENTS
{
	MPI_OP_GET_NEXT=1,
	MPI_SEND_ARRIVED,
55
    MPI_SEND_ARRIVED_CB, // for tracking message times on sender
56 57 58
	MPI_SEND_POSTED,
};

59
/* stores pointers of pending MPI operations to be matched with their respective sends/receives. */
60 61
struct mpi_msgs_queue
{
62 63
	struct codes_workload_op * mpi_op;
	struct mpi_msgs_queue * next;
64 65
};

66
/* stores request IDs of completed MPI operations (Isends or Irecvs) */
67 68 69
struct completed_requests
{
	dumpi_req_id req_id;
70
	struct completed_requests * next;
71 72
};

73
/* for wait operations, store the pending operation and number of completed waits so far. */
74 75
struct pending_waits
{
76
	struct codes_workload_op * mpi_op;
77 78 79 80
	int num_completed;
	tw_stime start_time;
};

81
/* 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). */
82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100
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;
101 102 103
	unsigned long num_wait;
	unsigned long num_waitall;
	unsigned long num_waitsome;
104 105

	/* time spent by the LP in executing the app trace*/
106
	double start_time;
107
	double elapsed_time;
108
	/* time spent in compute operations */
109 110 111 112 113 114 115
	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;
116
	/* FIFO for isend messages arrived on destination */
117
	struct mpi_queue_ptrs * arrival_queue;
118
	/* FIFO for irecv messages posted but not yet matched with send operations */
119
	struct mpi_queue_ptrs * pending_recvs_queue;
120
	/* list of pending waits (and saved pending wait for reverse computation) */
121
	struct pending_waits * pending_waits;
122
	/* List of completed send/receive requests */
123
	struct completed_requests * completed_reqs;
124 125
};

126 127 128 129
/* 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). */
130 131
struct nw_message
{
132 133 134
   int msg_type;
   /* for reverse computation */
   struct codes_workload_op * op;
135

136 137 138 139 140 141 142 143 144 145 146
   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;
147 148
        // for callbacks - time message was received
        double msg_send_time;
149 150 151 152 153 154 155
        int16_t req_id;   
        int tag;
     } msg_info;

     /* required for reverse computation*/
     struct 
      {
156 157 158 159 160 161 162 163 164
        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;
165 166
      } rc;
  } u;
167 168
};

169
/* executes MPI wait operation */
170 171
static void codes_exec_mpi_wait(
        nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op);
172
/* reverse of mpi wait function. */
173 174
static void codes_exec_mpi_wait_rc(
        nw_state* s, tw_bf* bf, nw_message* m, tw_lp* lp);
175
/* executes MPI isend and send operations */
176 177
static void codes_exec_mpi_send(
        nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op);
178
/* execute MPI irecv operation */
179 180
static void codes_exec_mpi_recv(
        nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op);
181
/* reverse of mpi recv function. */
182 183
static void codes_exec_mpi_recv_rc(
        nw_state* s, nw_message* m, tw_lp* lp);
184
/* execute the computational delay */
185 186
static void codes_exec_comp_delay(
        nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op);
187
/* execute collective operation, currently only skips these operations. */
188 189
static void codes_exec_mpi_col(
        nw_state* s, tw_lp* lp);
190
/* gets the next MPI operation from the network-workloads API. */
191 192
static void get_next_mpi_operation(
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
193
/* reverse handler of get next mpi operation. */
194 195
static void get_next_mpi_operation_rc(
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
196 197 198 199
/* Makes a call to get_next_mpi_operation. */
static void codes_issue_next_event(tw_lp* lp);

///////////////////// HELPER FUNCTIONS FOR MPI MESSAGE QUEUE HANDLING ///////////////
200
/* upon arrival of local completion message, inserts operation in completed send queue */
201 202
static void update_send_completion_queue(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
203
/* reverse of the above function */
204 205
static void update_send_completion_queue_rc(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
206
/* upon arrival of an isend operation, updates the arrival queue of the network */
207 208
static void update_arrival_queue(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
209
/* reverse of the above function */
210 211
static void update_arrival_queue_rc(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
212
/* callback to a message sender for computing message time */
213 214 215 216 217
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);
218

219
/* insert MPI operation in the waiting queue*/
220 221
static void mpi_pending_queue_insert_op(
        struct mpi_queue_ptrs* mpi_queue, struct codes_workload_op* mpi_op);
222
/* remove completed request IDs from the queue for reuse. Reverse of above function. */
223 224
static void remove_req_id(
        struct completed_requests** requests, int16_t req_id);
225
/* remove MPI operation from the waiting queue.*/
226 227
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);
228
/* remove the tail of the MPI operation from waiting queue */
229 230
static int mpi_queue_remove_tail(
        tw_lpid lpid, struct mpi_queue_ptrs* mpi_queue);
231
/* insert completed MPI requests in the queue. */
232 233
static void mpi_completed_queue_insert_op(
        struct completed_requests** mpi_completed_queue, dumpi_req_id req_id);
234
/* notifies the wait operations (if any) about the completed receives and sends requests. */
235 236
static int notify_waits(
        nw_state* s, tw_bf* bf, tw_lp* lp, nw_message* m, dumpi_req_id req_id);
237
/* reverse of notify waits function. */
238 239
static void notify_waits_rc(
        nw_state* s, tw_bf* bf, tw_lp* lp, nw_message* m, dumpi_req_id completed_req);
240 241
/* conversion from seconds to eanaoseconds */
static tw_stime s_to_ns(tw_stime ns);
242

243 244 245 246 247
/* 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);
}
248

249 250 251 252 253 254 255 256 257 258 259 260
/* 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;
}

261
/* helper function: counts number of elements in the queue */
262 263
static int numQueue(struct mpi_queue_ptrs* mpi_queue)
{
264
	struct mpi_msgs_queue* tmp = mpi_queue->queue_head;
265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285
	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)
	{
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
/* 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);
370
  
371
  printCompletedQueue(s, lp);
372
  
373
  if(m->u.rc.matched_op == 1)
374 375
	s->pending_waits->num_completed--;
   /* if a wait-elem exists, it means the request ID has been matched*/
376
   if(m->u.rc.matched_op == 2) 
377
    {
378 379 380 381 382
        if(lp->gid == TRACE)
        {
            printf("\n %lf matched req id %d ", tw_now(lp), completed_req);
            printCompletedQueue(s, lp);
        }
383
        struct pending_waits* wait_elem = m->u.rc.saved_pending_wait;
384 385
        s->wait_time = m->u.rc.saved_wait_time;
        int count = wait_elem->mpi_op->u.waits.count; 
386

387 388
        for( i = 0; i < count; i++ )
            mpi_completed_queue_insert_op(&s->completed_reqs, wait_elem->mpi_op->u.waits.req_ids[i]);
389

390 391 392
        wait_elem->num_completed--;	
        s->pending_waits = wait_elem;
        tw_rand_reverse_unif(lp->rng);
393 394 395 396 397 398

   }
}

/* 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)
399 400 401 402 403 404 405
{
	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;
406
	m->u.rc.matched_op = 0;
407 408 409
	
	if(lp->gid == TRACE)
		printf("\n %lf notify waits req id %d ", tw_now(lp), completed_req);
410 411 412

	if(!wait_elem)
		return 0;
413

414 415
	int op_type = wait_elem->mpi_op->op_type;

416
	if(op_type == CODES_WK_WAIT)
417 418 419
	{
		if(wait_elem->mpi_op->u.wait.req_id == completed_req)	
		  {
420
			m->u.rc.saved_wait_time = s->wait_time;
421
			s->wait_time += (tw_now(lp) - wait_elem->start_time);
422
                        remove_req_id(&s->completed_reqs, completed_req);
423
	
424
			m->u.rc.saved_pending_wait = wait_elem;			
425 426 427 428 429
			s->pending_waits = NULL;
			codes_issue_next_event(lp);	
			return 0;
		 }
	}
430
	else if(op_type == CODES_WK_WAITALL)
431
	{
432 433
	   int required_count = wait_elem->mpi_op->u.waits.count;
	  for(i = 0; i < required_count; i++)
434 435
	   {
	    if(wait_elem->mpi_op->u.waits.req_ids[i] == completed_req)
436 437 438
		{
			if(lp->gid == TRACE)
				printCompletedQueue(s, lp);
439
			m->u.rc.matched_op = 1;
440
			wait_elem->num_completed++;	
441
		}
442 443
	   }
	   
444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462
	  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
          }
    }
463 464 465
	return 0;
}

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

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

    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));
501
    wait_op->mpi_op = mpi_op;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
502 503
    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
{
509 510 511 512 513
  if(lp->gid == TRACE)
   {
       printf("\n %lf codes exec mpi waitall reverse %d ", tw_now(lp), m->u.rc.found_match);
       printCompletedQueue(s, lp); 
   } 
514
  if(m->u.rc.found_match)
515
    {
516 517 518 519 520 521 522 523 524 525 526
        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);
   }
527 528
    else
    {
529 530 531 532 533 534 535
        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));
   }
536
}
537 538
static void codes_exec_mpi_wait_all(
        nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op)
539 540
{
  //assert(!s->pending_waits);
541
  int count = mpi_op->u.waits.count;
542 543 544 545 546 547 548
  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)
    {
549
  	printf(" \n (%lf) MPI waitall posted %d count", tw_now(lp), mpi_op->u.waits.count);
550
	for(i = 0; i < count; i++)
551
		printf(" %d ", (int)mpi_op->u.waits.req_ids[i]);
552 553 554 555 556 557
   	printCompletedQueue(s, lp);	 
   }
  while(current) 
   {
	  for(i = 0; i < count; i++)
	   {
558
	     req_id[i] = mpi_op->u.waits.req_ids[i];
559
	     if(req_id[i] == current->req_id)
560
 		    num_completed++;
561 562 563 564 565
   	  }
	 current = current->next;
   }

  if(TRACE== lp->gid)
566
	  printf("\n %lf Num completed %d count %d ", tw_now(lp), num_completed, count);
567

568
  m->u.rc.found_match = 0;
569
  if(count == num_completed)
570
  {
571
	m->u.rc.found_match = 1;
572
	for( i = 0; i < count; i++)	
573
		remove_req_id(&s->completed_reqs, req_id[i]);
574

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

/* request ID is being reused so delete it from the list once the matching is done */
589 590
static void remove_req_id(
        struct completed_requests** mpi_completed_queue, dumpi_req_id req_id)
591 592 593 594
{
	struct completed_requests* current = *mpi_completed_queue;

	if(!current)
595 596
		tw_error(TW_LOC, "\n REQ ID DOES NOT EXIST");
	
597
    if(current->req_id == req_id)
598 599 600 601 602 603 604 605 606 607 608 609
	{
		*mpi_completed_queue = current->next;
		free(current);
		return;
	}
	
	struct completed_requests* elem;
	while(current->next)
	{
	   elem = current->next;
	   if(elem->req_id == req_id)	
	     {
610 611 612 613
            current->next = elem->next;
            free(elem);
            return;
         }
614 615 616 617 618 619
	   current = current->next;	
	}
	return;
}

/* inserts mpi operation in the completed requests queue */
620 621
static void mpi_completed_queue_insert_op(
        struct completed_requests** mpi_completed_queue, dumpi_req_id req_id)
622 623 624 625 626 627 628 629
{
	struct completed_requests* reqs = malloc(sizeof(struct completed_requests));
	assert(reqs);

	reqs->req_id = req_id;

	if(!(*mpi_completed_queue))	
	{
630 631 632
        reqs->next = NULL;
        *mpi_completed_queue = reqs;
        return;
633 634 635
	}
	reqs->next = *mpi_completed_queue;
	*mpi_completed_queue = reqs;
636
	return;
637 638
}

639
/* 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. */
640 641
static void mpi_pending_queue_insert_op(
        struct mpi_queue_ptrs* mpi_queue, struct codes_workload_op* mpi_op)
642 643 644 645 646 647
{
	/* insert mpi operation */
	struct mpi_msgs_queue* elem = malloc(sizeof(struct mpi_msgs_queue));
	assert(elem);

	elem->mpi_op = mpi_op;
648
    elem->next = NULL;
649 650 651 652 653 654 655

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

	if(mpi_queue->queue_tail)
	    mpi_queue->queue_tail->next = elem;
	
656
    mpi_queue->queue_tail = elem;
657 658 659 660 661 662
	mpi_queue->num_elems++;

	return;
}

/* match the send/recv operations */
663 664
static int match_receive(
        nw_state* s, tw_lp* lp, tw_lpid lpid, struct codes_workload_op* op1, struct codes_workload_op* op2)
665
{
666 667 668 669 670 671 672 673 674
        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);
675
                        s->recv_time += tw_now(lp) - op1->sim_start_time;
676 677 678 679
                        mpi_completed_queue_insert_op(&s->completed_reqs, op1->u.recv.req_id);
                        return 1;
                   }
        return -1;
680 681 682
}

/* used for reverse computation. removes the tail of the queue */
683
static int mpi_queue_remove_tail(tw_lpid lpid, struct mpi_queue_ptrs* mpi_queue)
684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717
{
	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. */
718
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)
719 720 721 722 723 724 725 726 727
{
	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 */
728
	int rcv_val = 0;
729
	if(mpi_op->op_type == CODES_WK_SEND || mpi_op->op_type == CODES_WK_ISEND)
730
	  {
731
		rcv_val = match_receive(s, lp, lp->gid, tmp->mpi_op, mpi_op);
732
		m->u.rc.saved_matched_req = tmp->mpi_op->u.recv.req_id;  
733
	 }
734
	else if(mpi_op->op_type == CODES_WK_RECV || mpi_op->op_type == CODES_WK_IRECV)
735
	  {
736
		rcv_val = match_receive(s, lp, lp->gid, mpi_op, tmp->mpi_op);
737
	  	m->u.rc.saved_matched_req = mpi_op->u.recv.req_id;
738 739
	  }
	if(rcv_val >= 0)
740
	{
741
		/* TODO: fix RC */
742
		m->u.rc.ptr_match_op = tmp->mpi_op;
743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764
		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;
765
	   
766
	    if(mpi_op->op_type == CODES_WK_SEND || mpi_op->op_type == CODES_WK_ISEND)
767
	     {
768
		    rcv_val = match_receive(s, lp, lp->gid, elem->mpi_op, mpi_op);
769
	     	m->u.rc.saved_matched_req = elem->mpi_op->u.recv.req_id; 
770
	     }
771
	    else if(mpi_op->op_type == CODES_WK_RECV || mpi_op->op_type == CODES_WK_IRECV)
772
	     {
773 774 775
            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;
         }
776
   	     if(rcv_val >= 0)
777 778
		 {
		    m->u.rc.ptr_match_op = elem->mpi_op;
779
		    if(elem == mpi_queue->queue_tail)
780
			    mpi_queue->queue_tail = tmp;
781
		    
782 783 784 785
		    tmp->next = elem->next;

		    free(elem);
		    mpi_queue->num_elems--;
786
		
787
		    return indx;
788
		 }
789
	   tmp = tmp->next;
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 */
810 811
static void codes_exec_comp_delay(
        nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op)
812 813 814 815 816
{
	tw_event* e;
	tw_stime ts;
	nw_message* msg;

817 818 819 820 821 822 823 824 825
        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);
826 827 828 829 830 831 832 833 834
	
	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 */
835
static void codes_exec_mpi_recv_rc(nw_state* s, nw_message* m, tw_lp* lp)
836
{
837 838 839
	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)
840
	  {
841 842 843
		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);
844
		tw_rand_reverse_unif(lp->rng);
845
	  }
846
	else if(m->u.rc.found_match < 0)
847
	    {
848 849
		mpi_queue_remove_tail(lp->gid, s->pending_recvs_queue);
		if(m->op->op_type == CODES_WK_IRECV)
850
			tw_rand_reverse_unif(lp->rng);
851 852 853 854
	    }
}

/* Execute MPI Irecv operation (non-blocking receive) */ 
855
static void codes_exec_mpi_recv(nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op)
856 857 858 859 860
{
/* 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. */

861
	m->u.rc.saved_recv_time = s->recv_time;
862
	mpi_op->sim_start_time = tw_now(lp);
863
	num_bytes_recvd += mpi_op->u.recv.num_bytes;
864 865

	if(lp->gid == TRACE)
866
		printf("\n %lf codes exec mpi recv req id %d", tw_now(lp), (int)mpi_op->u.recv.req_id);
867 868
	
	dumpi_req_id req_id;
869
	int found_matching_sends = mpi_queue_remove_matching_op(s, lp, m, s->arrival_queue, mpi_op);
870 871 872
	
	/* save the req id inserted in the completed queue for reverse computation. */
	//m->matched_recv = req_id;
873 874 875

	if(found_matching_sends < 0)
	  {
876
		m->u.rc.found_match = -1;
877 878 879
		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. */
880
		if(mpi_op->op_type == CODES_WK_IRECV)
881 882 883 884 885 886
		   {
			codes_issue_next_event(lp);	
			return;
		   }
		else
			printf("\n CODES MPI RECV OPERATION!!! ");
887
	  }
888
	else
889
	  {
890
	   	m->u.rc.found_match = found_matching_sends;
891
		codes_issue_next_event(lp); 
892 893 894 895
	 }
}

/* executes MPI send and isend operations */
896
static void codes_exec_mpi_send(nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op)
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;

922 923
	nw_message local_m;
	nw_message remote_m;
924

925 926 927 928 929 930 931 932
    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;
933

934 935
    remote_m = local_m;
	remote_m.msg_type = MPI_SEND_ARRIVED;
936

937
	model_net_event(net_id, "test", dest_rank, mpi_op->u.send.num_bytes, 0.0, 
938
	    sizeof(nw_message), (const void*)&remote_m, sizeof(nw_message), (const void*)&local_m, lp);
939

940
	/* isend executed, now get next MPI operation from the queue */ 
941
	if(mpi_op->op_type == CODES_WK_ISEND)
942 943 944 945
	   codes_issue_next_event(lp);
}

/* MPI collective operations */
946
static void codes_exec_mpi_col(nw_state* s, tw_lp* lp)
947 948 949 950 951 952 953 954 955 956 957 958 959 960
{
	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);