model-net-mpi-wrklds.c 46.3 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
#include "codes/codes.h"
#include "codes/configuration.h"
#include "codes/codes_mapping.h"
#include "codes/model-net.h"
14
#include "codes/rc-stack.h"
15

16
#define TRACE 0
17
#define TRACK 0
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;

25 26 27 28 29 30
/* 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;

31 32
typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
33
typedef int16_t dumpi_req_id;
34 35 36 37 38 39

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;
40 41
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;
42 43 44 45 46

/* 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;

47 48 49
/* runtime option for disabling computation time simulation */
static int disable_delay = 0;

50 51 52
/* 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). */
53 54 55 56
enum MPI_NW_EVENTS
{
	MPI_OP_GET_NEXT=1,
	MPI_SEND_ARRIVED,
57
    MPI_SEND_ARRIVED_CB, // for tracking message times on sender
58 59 60
	MPI_SEND_POSTED,
};

61
/* stores pointers of pending MPI operations to be matched with their respective sends/receives. */
62 63
struct mpi_msgs_queue
{
64 65
	struct codes_workload_op * mpi_op;
	struct mpi_msgs_queue * next;
66 67
};

68
/* stores request IDs of completed MPI operations (Isends or Irecvs) */
69 70 71
struct completed_requests
{
	dumpi_req_id req_id;
72
	struct completed_requests * next;
73 74
};

75
/* for wait operations, store the pending operation and number of completed waits so far. */
76 77
struct pending_waits
{
78
	struct codes_workload_op * mpi_op;
79 80 81 82
	int num_completed;
	tw_stime start_time;
};

83
/* 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). */
84 85 86 87 88 89 90 91 92 93 94 95 96 97
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;

98
    struct rc_stack * st;
99 100 101 102

    uint64_t num_completed;

    /* count of sends, receives, collectives and delays */
103 104 105 106
	unsigned long num_sends;
	unsigned long num_recvs;
	unsigned long num_cols;
	unsigned long num_delays;
107 108 109
	unsigned long num_wait;
	unsigned long num_waitall;
	unsigned long num_waitsome;
110 111

	/* time spent by the LP in executing the app trace*/
112
	double start_time;
113
	double elapsed_time;
114
	/* time spent in compute operations */
115 116 117 118 119 120 121
	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;
122
	/* FIFO for isend messages arrived on destination */
123
	struct mpi_queue_ptrs * arrival_queue;
124
	/* FIFO for irecv messages posted but not yet matched with send operations */
125
	struct mpi_queue_ptrs * pending_recvs_queue;
126
	/* list of pending waits (and saved pending wait for reverse computation) */
127
	struct pending_waits * pending_waits;
128
	/* List of completed send/receive requests */
129
	struct completed_requests * completed_reqs;
130 131
};

132 133 134 135
/* 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). */
136 137
struct nw_message
{
138
   int msg_type;
139

140 141 142 143 144 145 146 147 148 149 150
   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;
151 152
        // for callbacks - time message was received
        double msg_send_time;
153 154 155 156 157 158 159
        int16_t req_id;   
        int tag;
     } msg_info;

     /* required for reverse computation*/
     struct 
      {
160 161 162 163
        int found_match;
        short matched_op;
        dumpi_req_id saved_matched_req;
        struct codes_workload_op* ptr_match_op;
164
        struct codes_workload_op* saved_op;
165 166 167 168 169
        struct pending_waits* saved_pending_wait;

        double saved_send_time;
        double saved_recv_time;
        double saved_wait_time;
170 171
      } rc;
  } u;
172 173
};

174
/* executes MPI wait operation */
175 176
static void codes_exec_mpi_wait(
        nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op);
177
/* reverse of mpi wait function. */
178
static void codes_exec_mpi_wait_rc(
179
        nw_state* s, nw_message* m, tw_lp* lp, struct codes_workload_op * mpi_op);
180
/* executes MPI isend and send operations */
181 182
static void codes_exec_mpi_send(
        nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op);
183
/* execute MPI irecv operation */
184 185
static void codes_exec_mpi_recv(
        nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op);
186
/* reverse of mpi recv function. */
187
static void codes_exec_mpi_recv_rc(
188
        nw_state* s, nw_message* m, tw_lp* lp, struct codes_workload_op * mpi_op);
189
/* execute the computational delay */
190 191
static void codes_exec_comp_delay(
        nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op);
192
/* execute collective operation, currently only skips these operations. */
193 194
static void codes_exec_mpi_col(
        nw_state* s, tw_lp* lp);
195
/* gets the next MPI operation from the network-workloads API. */
196 197
static void get_next_mpi_operation(
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
198
/* reverse handler of get next mpi operation. */
199 200
static void get_next_mpi_operation_rc(
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
201 202 203 204
/* Makes a call to get_next_mpi_operation. */
static void codes_issue_next_event(tw_lp* lp);

///////////////////// HELPER FUNCTIONS FOR MPI MESSAGE QUEUE HANDLING ///////////////
205
/* upon arrival of local completion message, inserts operation in completed send queue */
206 207
static void update_send_completion_queue(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
208
/* reverse of the above function */
209 210
static void update_send_completion_queue_rc(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
211
/* upon arrival of an isend operation, updates the arrival queue of the network */
212 213
static void update_arrival_queue(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
214
/* reverse of the above function */
215 216
static void update_arrival_queue_rc(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
217
/* callback to a message sender for computing message time */
218 219 220 221 222
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);
223

224
/* insert MPI operation in the waiting queue*/
225 226
static void mpi_pending_queue_insert_op(
        struct mpi_queue_ptrs* mpi_queue, struct codes_workload_op* mpi_op);
227
/* remove completed request IDs from the queue for reuse. Reverse of above function. */
228 229
static void remove_req_id(
        struct completed_requests** requests, int16_t req_id);
230
/* remove MPI operation from the waiting queue.*/
231 232
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);
233
/* remove the tail of the MPI operation from waiting queue */
234 235
static int mpi_queue_remove_tail(
        tw_lpid lpid, struct mpi_queue_ptrs* mpi_queue);
236
/* insert completed MPI requests in the queue. */
237 238
static void mpi_completed_queue_insert_op(
        struct completed_requests** mpi_completed_queue, dumpi_req_id req_id);
239
/* notifies the wait operations (if any) about the completed receives and sends requests. */
240 241
static int notify_waits(
        nw_state* s, tw_bf* bf, tw_lp* lp, nw_message* m, dumpi_req_id req_id);
242
/* reverse of notify waits function. */
243 244
static void notify_waits_rc(
        nw_state* s, tw_bf* bf, tw_lp* lp, nw_message* m, dumpi_req_id completed_req);
245 246
/* conversion from seconds to eanaoseconds */
static tw_stime s_to_ns(tw_stime ns);
247

248 249 250 251 252
/* 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);
}
253

254 255 256 257 258 259 260 261 262 263 264 265
/* 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;
}

266
/* helper function: counts number of elements in the queue */
267 268
static int numQueue(struct mpi_queue_ptrs* mpi_queue)
{
269
	struct mpi_msgs_queue* tmp = mpi_queue->queue_head;
270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290
	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)
	{
291
		if(tmp->mpi_op->op_type == CODES_WK_SEND || tmp->mpi_op->op_type == CODES_WK_ISEND)
292 293 294
			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);
295
		else if(tmp->mpi_op->op_type == CODES_WK_IRECV || tmp->mpi_op->op_type == CODES_WK_RECV)
296 297 298 299 300 301 302 303 304 305 306
			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 */
307
static void mpi_queue_update(struct mpi_queue_ptrs* mpi_queue, struct codes_workload_op* mpi_op, int pos)
308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343
{
	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;
}

344
/* prints the elements of a queue (for debugging purposes). */
345 346
static void printCompletedQueue(nw_state* s, tw_lp* lp)
{
347
	   if(TRACE == s->nw_id)
348
	   {
349
	   	printf("\n %lf contents of completed operations queue ", tw_now(lp));
350 351 352 353 354 355 356 357 358
	   	struct completed_requests* current = s->completed_reqs;
	   	while(current)
	    	{
			printf(" %d ",current->req_id);
			current = current->next;
	   	}
	   }
}

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

392 393
        for( i = 0; i < count; i++ )
            mpi_completed_queue_insert_op(&s->completed_reqs, wait_elem->mpi_op->u.waits.req_ids[i]);
394

395 396 397
        wait_elem->num_completed--;	
        s->pending_waits = wait_elem;
        tw_rand_reverse_unif(lp->rng);
398 399 400 401 402
   }
}

/* 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)
403 404 405 406 407 408 409
{
	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;
410
	m->u.rc.matched_op = 0;
411
	
412
	if(s->nw_id == TRACE)
413
		printf("\n %lf notify waits req id %d ", tw_now(lp), completed_req);
414 415 416

	if(!wait_elem)
		return 0;
417

418 419
	int op_type = wait_elem->mpi_op->op_type;

420
	if(op_type == CODES_WK_WAIT)
421 422 423
	{
		if(wait_elem->mpi_op->u.wait.req_id == completed_req)	
		  {
424
			m->u.rc.saved_wait_time = s->wait_time;
425
			s->wait_time += (tw_now(lp) - wait_elem->start_time);
426
                        remove_req_id(&s->completed_reqs, completed_req);
427
	
428
			m->u.rc.saved_pending_wait = wait_elem;			
429
            s->pending_waits = NULL;
430 431 432 433
			codes_issue_next_event(lp);	
			return 0;
		 }
	}
434
	else if(op_type == CODES_WK_WAITALL)
435
	{
436 437
	   int required_count = wait_elem->mpi_op->u.waits.count;
	  for(i = 0; i < required_count; i++)
438 439
	   {
	    if(wait_elem->mpi_op->u.waits.req_ids[i] == completed_req)
440
		{
441
			if(s->nw_id == TRACE)
442
				printCompletedQueue(s, lp);
443
			m->u.rc.matched_op = 1;
444
			wait_elem->num_completed++;	
445
		}
446 447
	   }
	   
448 449
	  if(wait_elem->num_completed == required_count)
	   {
450
            if(s->nw_id == TRACE)
451 452 453 454 455 456 457 458 459 460 461 462 463 464
            {
                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
465
       }
466
    }
467 468 469
	return 0;
}

470
/* reverse handler of MPI wait operation */
471
static void codes_exec_mpi_wait_rc(nw_state* s, nw_message* m, tw_lp* lp, struct codes_workload_op * mpi_op)
472 473 474 475
{
    if(s->pending_waits)
     {
    	s->pending_waits = NULL;
476
	    return;
477 478 479
     }
   else
    {
480 481
 	    mpi_completed_queue_insert_op(&s->completed_reqs, mpi_op->u.wait.req_id);	
	    tw_rand_reverse_unif(lp->rng);	
482
        //rc_stack_pop(s->st);
483 484
    }
}
485 486

/* execute MPI wait operation */
487
static void codes_exec_mpi_wait(nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op)
488
{
489 490
    /* check in the completed receives queue if the request ID has already been completed.*/
    assert(!s->pending_waits);
491
    dumpi_req_id req_id = mpi_op->u.wait.req_id;
492 493 494 495 496 497 498 499 500 501

    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;
502 503
    }

504
    /* If not, add the wait operation in the pending 'waits' list. */
505
    struct pending_waits* wait_op = malloc(sizeof(struct pending_waits));
506
    wait_op->mpi_op = mpi_op;
507 508
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
509
    s->pending_waits = wait_op;
510

511
    //rc_stack_push(lp, wait_op, free, s->st);
512 513
}

514
static void codes_exec_mpi_wait_all_rc(nw_state* s, nw_message* m, tw_lp* lp, struct codes_workload_op * mpi_op)
515
{
516
  if(s->nw_id == TRACE)
517 518 519 520
   {
       printf("\n %lf codes exec mpi waitall reverse %d ", tw_now(lp), m->u.rc.found_match);
       printCompletedQueue(s, lp); 
   } 
521
  if(m->u.rc.found_match)
522
    {
523
        int i;
524
        int count = mpi_op->u.waits.count;
525 526 527 528
        dumpi_req_id req_id[count];

        for( i = 0; i < count; i++)
        {
529
            req_id[i] = mpi_op->u.waits.req_ids[i];
530 531 532 533
            mpi_completed_queue_insert_op(&s->completed_reqs, req_id[i]);
        }
        tw_rand_reverse_unif(lp->rng);
   }
534 535
    else
    {
536
        struct pending_waits* wait_op = s->pending_waits;
537
        //rc_stack_pop(s->st);
538 539 540 541 542
        s->pending_waits = NULL;
        assert(!s->pending_waits);
        if(lp->gid == TRACE)
            printf("\n %lf Nullifying codes waitall ", tw_now(lp));
   }
543
}
544 545
static void codes_exec_mpi_wait_all(
        nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op)
546 547
{
  //assert(!s->pending_waits);
548
  int count = mpi_op->u.waits.count;
549 550 551 552 553 554 555
  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)
    {
556
  	printf(" \n (%lf) MPI waitall posted %d count", tw_now(lp), mpi_op->u.waits.count);
557
	for(i = 0; i < count; i++)
558
		printf(" %d ", (int)mpi_op->u.waits.req_ids[i]);
559 560 561 562 563 564
   	printCompletedQueue(s, lp);	 
   }
  while(current) 
   {
	  for(i = 0; i < count; i++)
	   {
565
	     req_id[i] = mpi_op->u.waits.req_ids[i];
566
	     if(req_id[i] == current->req_id)
567
 		    num_completed++;
568 569 570 571 572
   	  }
	 current = current->next;
   }

  if(TRACE== lp->gid)
573
	  printf("\n %lf Num completed %d count %d ", tw_now(lp), num_completed, count);
574

575
  m->u.rc.found_match = 0;
576
  if(count == num_completed)
577
  {
578
	m->u.rc.found_match = 1;
579
	for( i = 0; i < count; i++)	
580
		remove_req_id(&s->completed_reqs, req_id[i]);
581

582 583 584 585 586 587
	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));
588
	  wait_op->mpi_op = mpi_op;  
589
	  wait_op->num_completed = num_completed;
590
	  wait_op->start_time = tw_now(lp);
591
      //rc_stack_push(lp, wait_op, free, s->st);
592
      s->pending_waits = wait_op;
593 594 595 596
  }
}

/* request ID is being reused so delete it from the list once the matching is done */
597 598
static void remove_req_id(
        struct completed_requests** mpi_completed_queue, dumpi_req_id req_id)
599 600 601 602
{
	struct completed_requests* current = *mpi_completed_queue;

	if(!current)
603 604
		tw_error(TW_LOC, "\n REQ ID DOES NOT EXIST");
	
605
    if(current->req_id == req_id)
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)	
	     {
618 619 620 621
            current->next = elem->next;
            free(elem);
            return;
         }
622 623 624 625 626 627
	   current = current->next;	
	}
	return;
}

/* inserts mpi operation in the completed requests queue */
628 629
static void mpi_completed_queue_insert_op(
        struct completed_requests** mpi_completed_queue, dumpi_req_id req_id)
630 631 632 633 634 635 636 637
{
	struct completed_requests* reqs = malloc(sizeof(struct completed_requests));
	assert(reqs);

	reqs->req_id = req_id;

	if(!(*mpi_completed_queue))	
	{
638 639 640
        reqs->next = NULL;
        *mpi_completed_queue = reqs;
        return;
641 642 643
	}
	reqs->next = *mpi_completed_queue;
	*mpi_completed_queue = reqs;
644
	return;
645 646
}

647
/* 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. */
648 649
static void mpi_pending_queue_insert_op(
        struct mpi_queue_ptrs* mpi_queue, struct codes_workload_op* mpi_op)
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;
656
    elem->next = NULL;
657 658 659 660 661 662 663

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

	if(mpi_queue->queue_tail)
	    mpi_queue->queue_tail->next = elem;
	
664
    mpi_queue->queue_tail = elem;
665 666 667 668 669 670
	mpi_queue->num_elems++;

	return;
}

/* match the send/recv operations */
671 672
static int match_receive(
        nw_state* s, tw_lp* lp, tw_lpid lpid, struct codes_workload_op* op1, struct codes_workload_op* op2)
673
{
674 675 676 677 678 679 680 681 682
        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);
683
                        s->recv_time += tw_now(lp) - op1->sim_start_time;
684 685 686 687
                        mpi_completed_queue_insert_op(&s->completed_reqs, op1->u.recv.req_id);
                        return 1;
                   }
        return -1;
688 689 690
}

/* used for reverse computation. removes the tail of the queue */
691
static int mpi_queue_remove_tail(tw_lpid lpid, struct mpi_queue_ptrs* mpi_queue)
692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725
{
	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. */
726
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)
727 728 729 730 731 732 733 734 735
{
	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 */
736
	int rcv_val = 0;
737
	if(mpi_op->op_type == CODES_WK_SEND || mpi_op->op_type == CODES_WK_ISEND)
738
	  {
739
		rcv_val = match_receive(s, lp, lp->gid, tmp->mpi_op, mpi_op);
740
		m->u.rc.saved_matched_req = tmp->mpi_op->u.recv.req_id;  
741
	 }
742
	else if(mpi_op->op_type == CODES_WK_RECV || mpi_op->op_type == CODES_WK_IRECV)
743
	  {
744
		rcv_val = match_receive(s, lp, lp->gid, mpi_op, tmp->mpi_op);
745
	  	m->u.rc.saved_matched_req = mpi_op->u.recv.req_id;
746 747
	  }
	if(rcv_val >= 0)
748
	{
749
		/* TODO: fix RC */
750
		m->u.rc.ptr_match_op = tmp->mpi_op;
751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772
		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;
773
	   
774
	    if(mpi_op->op_type == CODES_WK_SEND || mpi_op->op_type == CODES_WK_ISEND)
775
	     {
776
		    rcv_val = match_receive(s, lp, lp->gid, elem->mpi_op, mpi_op);
777
	     	m->u.rc.saved_matched_req = elem->mpi_op->u.recv.req_id; 
778
	     }
779
	    else if(mpi_op->op_type == CODES_WK_RECV || mpi_op->op_type == CODES_WK_IRECV)
780
	     {
781 782 783
            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;
         }
784
   	     if(rcv_val >= 0)
785 786
		 {
		    m->u.rc.ptr_match_op = elem->mpi_op;
787
		    if(elem == mpi_queue->queue_tail)
788
			    mpi_queue->queue_tail = tmp;
789
		    
790 791 792 793
		    tmp->next = elem->next;

		    free(elem);
		    mpi_queue->num_elems--;
794
		
795
		    return indx;
796
		 }
797
	   tmp = tmp->next;
798
    }
799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817
	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 */
818 819
static void codes_exec_comp_delay(
        nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op)
820 821 822 823 824
{
	tw_event* e;
	tw_stime ts;
	nw_message* msg;

825 826 827 828 829 830 831 832 833
        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);
834 835 836 837 838 839 840 841 842
	
	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 */
843
static void codes_exec_mpi_recv_rc(nw_state* s, nw_message* m, tw_lp* lp, struct codes_workload_op * mpi_op)
844
{
845
	num_bytes_recvd -= mpi_op->u.recv.num_bytes;
846 847
	s->recv_time = m->u.rc.saved_recv_time;
	if(m->u.rc.found_match >= 0)
848
	  {
849 850
		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);
851
		remove_req_id(&s->completed_reqs, mpi_op->u.recv.req_id);
852
		tw_rand_reverse_unif(lp->rng);
853
	  }
854
	else if(m->u.rc.found_match < 0)
855
	    {
856
		mpi_queue_remove_tail(lp->gid, s->pending_recvs_queue);
857
		if(mpi_op->op_type == CODES_WK_IRECV)
858
			tw_rand_reverse_unif(lp->rng);
859 860 861 862
	    }
}

/* Execute MPI Irecv operation (non-blocking receive) */ 
863
static void codes_exec_mpi_recv(nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op)
864 865 866 867 868
{
/* 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. */

869
	m->u.rc.saved_recv_time = s->recv_time;
870
	mpi_op->sim_start_time = tw_now(lp);
871
	num_bytes_recvd += mpi_op->u.recv.num_bytes;
872 873

	if(lp->gid == TRACE)
874
		printf("\n %lf codes exec mpi recv req id %d", tw_now(lp), (int)mpi_op->u.recv.req_id);
875 876
	
	dumpi_req_id req_id;
877
	int found_matching_sends = mpi_queue_remove_matching_op(s, lp, m, s->arrival_queue, mpi_op);
878 879 880
	
	/* save the req id inserted in the completed queue for reverse computation. */
	//m->matched_recv = req_id;
881 882 883

	if(found_matching_sends < 0)
	  {
884
		m->u.rc.found_match = -1;
885 886 887
		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. */
888
		if(mpi_op->op_type == CODES_WK_IRECV)
889 890 891 892 893 894
		   {
			codes_issue_next_event(lp);	
			return;
		   }
		else
			printf("\n CODES MPI RECV OPERATION!!! ");
895
	  }
896
	else
897
	  {
898
	   	m->u.rc.found_match = found_matching_sends;
899
        codes_issue_next_event(lp); 
900 901 902 903
	 }
}

/* executes MPI send and isend operations */
904
static void codes_exec_mpi_send(nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op)
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
{
	/* 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;

930 931
	nw_message local_m;
	nw_message remote_m;
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;
941

942 943
    remote_m = local_m;
	remote_m.msg_type = MPI_SEND_ARRIVED;
944

945
	model_net_event(net_id, "test", dest_rank, mpi_op->u.send.num_bytes, 0.0, 
946
	    sizeof(nw_message), (const void*)&remote_m, sizeof(nw_message), (const void*)&local_m, lp);
947

948
	/* isend executed, now get next MPI operation from the queue */ 
949
	if(mpi_op->op_type == CODES_WK_ISEND)
950
	   codes_issue_next_event(lp);
951

952 953 954
}

/* MPI collective operations */
955
static void codes_exec_mpi_col(nw_state* s, tw_lp* lp)
956 957 958 959 960 961 962 963 964 965 966 967 968 969
{
	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
	s->recv_time = m->u.rc.saved_recv_time;
1002 1003
    codes_local_latency_reverse(lp);
    
1004
    //rc_stack_pop(s->st);
1005

1006
	if(m->u.rc.found_match >= 0)
1007
	{
1008 1009
		// TODO: Modify for recvs
		if(lp->gid == TRACE)
1010 1011 1012
			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);
1013
		//int count = numQueue(s->pending_recvs_queue);
1014 1015
		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);
1016
	
1017 1018 1019
		/*if(lp->gid == TRACE)
			printf("\n Reverse: after adding pending recvs queue %d ", s->pending_recvs_queue->num_elems);*/
	}
1020
	else if(m->u.rc.found_match < 0)
1021
	{
1022
		mpi_queue_remove_tail(lp->gid, s->arrival_queue);	
1023 1024 1025 1026 1027 1028 1029 1030
		/*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)
{
1031 1032
	//int count_before = numQueue(s->pending_recvs_queue);
	int is_blocking = 0; /* checks if the recv operation was blocking or not */
1033

1034
	m->u.rc.saved_recv_time = s->recv_time;
1035

1036 1037 1038 1039 1040 1041 1042 1043
    // send a callback to the sender to increment times
    tw_event *e_callback =
        tw_event_new(rank_to_lpid(m->u.msg_info.src_rank),
                codes_local_latency(lp), lp);
    nw_message *m_callback = tw_event_data(e_callback);
    m_callback->msg_type = MPI_SEND_ARRIVED_CB;
    m_callback->u.msg_info.msg_send_time = tw_now(lp) - m->u.msg_info.sim_start_time;
    tw_event_send(e_callback);
1044 1045 1046 1047

        /*NOTE: this computes send time with respect to the receiver, not the
         * sender
         * s->send_time += tw_now(lp) - m->u.msg_info.sim_start_time; */
1048 1049
	dumpi_req_id req_id = -1;

1050
        /* Now reconstruct the mpi op */
1051 1052 1053 1054 1055 1056 1057 1058
    struct codes_workload_op * arrived_op = (struct codes_workload_op *) malloc(sizeof(struct codes_workload_op));
    arrived_op->sim_start_time = m->u.msg_info.sim_start_time;
    arrived_op->op_type = m->u.msg_info.op_type;
    arrived_op->u.send.source_rank = m->u.msg_info.src_rank;
    arrived_op->u.send.dest_rank = m->u.msg_info.dest_rank;
    arrived_op->u.send.num_bytes = m->u.msg_info.num_bytes;
    arrived_op->u.send.tag = m->u.msg_info.tag;
    arrived_op->u.send.req_id = m->u.msg_info.req_id;
1059
    //rc_stack_push(lp, arrived_op, free, s->st);
1060 1061 1062 1063

    int found_matching_recv = mpi_queue_remove_matching_op(s, lp, m, s->pending_recvs_queue, arrived_op);

    if(TRACE == lp->gid)
1064
        printf("\n %lf update arrival queue req id %d %d", tw_now(lp), arrived_op->u.send.req_id, arrived_op->u.send.source_rank);
1065 1066 1067 1068 1069 1070 1071 1072 1073 1074
    if(found_matching_recv < 0)
     {
        m->u.rc.found_match = -1;
        mpi_pending_queue_insert_op(s->arrival_queue, arrived_op);
    }
    else
      {
        m->u.rc.found_match = found_matching_recv;
        notify_waits(s, bf, lp, m, m->u.rc.saved_matched_req);
      }
1075 1076
}

1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095
static void update_message_time(
        nw_state * s,
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp)
{
    m->u.rc.saved_send_time = s->send_time;
    s->send_time += m->u.msg_info.msg_send_time;
}

static void update_message_time_rc(
        nw_state * s,
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp)
{
    s->send_time = m->u.rc.saved_send_time;
}

1096 1097 1098 1099
/* initializes the network node LP, loads the trace file in the structs, calls the first MPI operation to be executed */
void nw_test_init(nw_state* s, tw_lp* lp)
{
   /* initialize the LP's and load the data */
1100
   char * params = NULL;
1101 1102 1103 1104 1105
   dumpi_trace_params params_d;
  
   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);
  
1106
   memset(s, 0, sizeof(*s));
1107
   s->nw_id = (mapping_rep_id * num_nw_lps) + mapping_offset;
1108 1109 1110
   s->completed_reqs = NULL;

   s->pending_waits = NULL;
1111 1112 1113
   if(!num_net_traces) 
	num_net_traces = num_net_lps;

1114
   if (strcmp(workload_type, "dumpi") == 0){
1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125
       strcpy(params_d.file_name, workload_file);
       params_d.num_net_traces = num_net_traces;

       params = (char*)&params_d;
   }
  /* In this case, the LP will not generate any workload related events*/
   if(s->nw_id >= params_d.num_net_traces)
     {
	//printf("\n network LP not generating events %d ", (int)s->nw_id);
	return;
     }
1126 1127 1128 1129 1130

   /* Initialize the RC stack */
   rc_stack_create(&s->st);
   assert(s->st != NULL);

1131
   wrkld_id = codes_workload_load("dumpi-trace-workload", params, 0, (int)s->nw_id);
1132 1133 1134 1135

   s->arrival_queue = queue_init(); 
   s->pending_recvs_queue = queue_init();

1136 1137
   /* clock starts when the first event is processed */
   s->start_time = tw_now(lp);
1138 1139 1140 1141 1142 1143 1144
   codes_issue_next_event(lp);

   return;
}

void nw_test_event_handler(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
1145
	*(int *)bf = (int)0;
1146 1147 1148
    rc_stack_gc(lp, s->st);

    switch(m->msg_type)
1149 1150 1151 1152 1153 1154 1155 1156 1157
	{
		case MPI_SEND_POSTED:
			update_send_completion_queue(s, bf, m, lp);
		break;

		case MPI_SEND_ARRIVED:
			update_arrival_queue(s, bf, m, lp);
		break;

1158 1159 1160 1161
		case MPI_SEND_ARRIVED_CB:
			update_message_time(s, bf, m, lp);
		break;

1162 1163 1164 1165 1166 1167 1168 1169
		case MPI_OP_GET_NEXT:
			get_next_mpi_operation(s, bf, m, lp);	
		break; 
	}
}

static void get_next_mpi_operation_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
1170 1171
    struct codes_workload_op * mpi_op = m->u.rc.saved_op;
        //(struct codes_workload_op *)rc_stack_pop(s->st);
1172 1173 1174 1175
	
    codes_workload_get_next_rc(wrkld_id, 0, (int)s->nw_id, mpi_op);

	if(mpi_op->op_type == CODES_WK_END)
1176
		return;
1177

1178
	switch(mpi_op->op_type)
1179
	{
1180 1181
		case CODES_WK_SEND:
		case CODES_WK_ISEND:
1182
		{
1183
			if(lp->gid == TRACE)
1184 1185 1186
				printf("\n %lf reverse send req %d ", tw_now(lp), (int)mpi_op->u.send.req_id);
			model_net_event_rc(net_id, lp, mpi_op->u.send.num_bytes);
			if(mpi_op->op_type == CODES_WK_ISEND)
1187 1188
				tw_rand_reverse_unif(lp->rng);	
			s->num_sends--;
1189
			num_bytes_sent -= mpi_op->u.send.num_bytes;
1190 1191
		}
		break;
1192 1193

		case CODES_WK_IRECV:
1194
		case CODES_WK_RECV:
1195
		{
1196
			codes_exec_mpi_recv_rc(s, m, lp, mpi_op);
1197 1198 1199
			s->num_recvs--;
		}
		break;
1200
		case CODES_WK_DELAY:
1201 1202
		{
			s->num_delays--;
1203 1204
                        if (!disable_delay) {
                            tw_rand_reverse_unif(lp->rng);
1205
                            s->compute_time -= s_to_ns(mpi_op->u.delay.seconds);
1206
                        }
1207 1208
		}
		break;
1209 1210 1211 1212 1213 1214 1215 1216
		case CODES_WK_BCAST:
		case CODES_WK_ALLGATHER:
		case CODES_WK_ALLGATHERV:
		case CODES_WK_ALLTOALL:
		case CODES_WK_ALLTOALLV:
		case CODES_WK_REDUCE:
		case CODES_WK_ALLREDUCE:
		case CODES_WK_COL:
1217 1218 1219 1220 1221
		{
			s->num_cols--;
			tw_rand_reverse_unif(lp->rng);
		}
		break;
1222
	
1223
		case CODES_WK_WAIT:
1224
		{
1225
			s->num_wait--;
1226
			codes_exec_mpi_wait_rc(s, m, lp, mpi_op);
1227 1228
		}
		break;
1229
		case CODES_WK_WAITALL:
1230 1231
		{
			s->num_waitall--;
1232
			codes_exec_mpi_wait_all_rc(s, m, lp, mpi_op);
1233 1234
		}
		break;
1235 1236
		case CODES_WK_WAITSOME:
		case CODES_WK_WAITANY:
1237
		{
1238 1239
			s->num_waitsome--;
			tw_rand_reverse_unif(lp->rng);
1240 1241
		}
		break;
1242
		default:
1243
			printf("\n Invalid op type %d ", mpi_op->op_type);
1244 1245 1246 1247 1248
	}
}

static void get_next_mpi_operation(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
1249 1250 1251
		struct codes_workload_op * mpi_op = malloc(sizeof(struct codes_workload_op));
        codes_workload_get_next(wrkld_id, 0, (int)s->nw_id, mpi_op);

1252 1253 1254 1255 1256 1257
        s->num_completed++;

        if(s->nw_id == TRACK && s->num_completed % 10000 == 0)
            printf("\n Status: LP %ld completed %ld MPI operations ", 
                    s->nw_id, s->num_completed);

1258
        m->u.rc.saved_op = mpi_op;
1259
        if(mpi_op->op_type == CODES_WK_END && s->num_completed == 50000)
1260
        {
1261
            //rc_stack_push(lp, mpi_op, free, s->st);
1262 1263 1264 1265
            s->elapsed_time = tw_now(lp) - s->start_time;
            return;
        }
		switch(mpi_op->op_type)
1266
		{
1267 1268
			case CODES_WK_SEND:
			case CODES_WK_ISEND:
1269 1270
			 {
				s->num_sends++;
1271
				codes_exec_mpi_send(s, lp, mpi_op);
1272 1273
			 }
			break;
1274
	
1275 1276
			case CODES_WK_RECV:
			case CODES_WK_IRECV:
1277 1278
			  {
				s->num_recvs++;
1279
				codes_exec_mpi_recv(s, lp, m, mpi_op);
1280 1281 1282
			  }