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
#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 -1
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
{
138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158
    int msg_type;
    int op_type;
    tw_lpid src_rank;
    tw_lpid dest_rank;
    int num_bytes;
    int data_type;
    double sim_start_time;
    // for callbacks - time message was received
    double msg_send_time;
    int16_t req_id;   
    int tag;
    dumpi_req_id saved_matched_req;
    struct codes_workload_op* ptr_match_op;
    int found_match;
    short matched_op;
    struct codes_workload_op* saved_op;
    struct pending_waits* saved_pending_wait;

    double saved_send_time;
    double saved_recv_time;
    double saved_wait_time;
159 160
};

161
/* executes MPI wait operation */
162 163
static void codes_exec_mpi_wait(
        nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op);
164
/* reverse of mpi wait function. */
165
static void codes_exec_mpi_wait_rc(
166
        nw_state* s, nw_message* m, tw_lp* lp, struct codes_workload_op * mpi_op);
167
/* executes MPI isend and send operations */
168 169
static void codes_exec_mpi_send(
        nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op);
170
/* execute MPI irecv operation */
171 172
static void codes_exec_mpi_recv(
        nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op);
173
/* reverse of mpi recv function. */
174
static void codes_exec_mpi_recv_rc(
175
        nw_state* s, nw_message* m, tw_lp* lp, struct codes_workload_op * mpi_op);
176
/* execute the computational delay */
177 178
static void codes_exec_comp_delay(
        nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op);
179
/* execute collective operation, currently only skips these operations. */
180 181
static void codes_exec_mpi_col(
        nw_state* s, tw_lp* lp);
182
/* gets the next MPI operation from the network-workloads API. */
183 184
static void get_next_mpi_operation(
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
185
/* reverse handler of get next mpi operation. */
186 187
static void get_next_mpi_operation_rc(
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
188 189 190 191
/* Makes a call to get_next_mpi_operation. */
static void codes_issue_next_event(tw_lp* lp);

///////////////////// HELPER FUNCTIONS FOR MPI MESSAGE QUEUE HANDLING ///////////////
192
/* upon arrival of local completion message, inserts operation in completed send queue */
193 194
static void update_send_completion_queue(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
195
/* reverse of the above function */
196 197
static void update_send_completion_queue_rc(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
198
/* upon arrival of an isend operation, updates the arrival queue of the network */
199 200
static void update_arrival_queue(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
201
/* reverse of the above function */
202 203
static void update_arrival_queue_rc(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
204
/* callback to a message sender for computing message time */
205 206 207 208 209
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);
210

211
/* insert MPI operation in the waiting queue*/
212 213
static void mpi_pending_queue_insert_op(
        struct mpi_queue_ptrs* mpi_queue, struct codes_workload_op* mpi_op);
214
/* remove completed request IDs from the queue for reuse. Reverse of above function. */
215 216
static void remove_req_id(
        struct completed_requests** requests, int16_t req_id);
217
/* remove MPI operation from the waiting queue.*/
218 219
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);
220
/* remove the tail of the MPI operation from waiting queue */
221 222
static int mpi_queue_remove_tail(
        tw_lpid lpid, struct mpi_queue_ptrs* mpi_queue);
223
/* insert completed MPI requests in the queue. */
224 225
static void mpi_completed_queue_insert_op(
        struct completed_requests** mpi_completed_queue, dumpi_req_id req_id);
226
/* notifies the wait operations (if any) about the completed receives and sends requests. */
227 228
static int notify_waits(
        nw_state* s, tw_bf* bf, tw_lp* lp, nw_message* m, dumpi_req_id req_id);
229
/* reverse of notify waits function. */
230 231
static void notify_waits_rc(
        nw_state* s, tw_bf* bf, tw_lp* lp, nw_message* m, dumpi_req_id completed_req);
232 233
/* conversion from seconds to eanaoseconds */
static tw_stime s_to_ns(tw_stime ns);
234

235 236 237 238 239
/* 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);
}
240

241 242 243 244 245 246 247 248 249 250 251 252
/* 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;
}

253
/* helper function: counts number of elements in the queue */
254 255
static int numQueue(struct mpi_queue_ptrs* mpi_queue)
{
256
	struct mpi_msgs_queue* tmp = mpi_queue->queue_head;
257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277
	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)
	{
278
		if(tmp->mpi_op->op_type == CODES_WK_SEND || tmp->mpi_op->op_type == CODES_WK_ISEND)
279
			printf("\n lpid %llu send operation data type %d count %d tag %d source %d", 
280 281
				    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);
282
		else if(tmp->mpi_op->op_type == CODES_WK_IRECV || tmp->mpi_op->op_type == CODES_WK_RECV)
283
			printf("\n lpid %llu recv operation data type %d count %d tag %d source %d", 
284 285 286 287 288 289 290 291 292 293
				   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 */
294
static void mpi_queue_update(struct mpi_queue_ptrs* mpi_queue, struct codes_workload_op* mpi_op, int pos)
295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330
{
	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;
}

331
/* prints the elements of a queue (for debugging purposes). */
332 333
static void printCompletedQueue(nw_state* s, tw_lp* lp)
{
334
	   if(TRACE == s->nw_id)
335
	   {
336
	   	printf("\n %lf contents of completed operations queue ", tw_now(lp));
337 338 339 340 341 342 343 344 345
	   	struct completed_requests* current = s->completed_reqs;
	   	while(current)
	    	{
			printf(" %d ",current->req_id);
			current = current->next;
	   	}
	   }
}

346 347 348 349 350 351 352 353 354 355 356 357 358 359
/* 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;
    }
*/
360
  if(s->nw_id == TRACE)
361
	  printf("\n %lf reverse -- notify waits req id %d ", tw_now(lp), completed_req);
362
  
363
  printCompletedQueue(s, lp);
364
  
365
  if(m->matched_op == 1)
366 367
	s->pending_waits->num_completed--;
   /* if a wait-elem exists, it means the request ID has been matched*/
368
   if(m->matched_op == 2) 
369
    {
370
        if(s->nw_id == TRACE)
371 372 373 374
        {
            printf("\n %lf matched req id %d ", tw_now(lp), completed_req);
            printCompletedQueue(s, lp);
        }
375 376
        struct pending_waits* wait_elem = m->saved_pending_wait;
        s->wait_time = m->saved_wait_time;
377
        int count = wait_elem->mpi_op->u.waits.count; 
378

379 380
        for( i = 0; i < count; i++ )
            mpi_completed_queue_insert_op(&s->completed_reqs, wait_elem->mpi_op->u.waits.req_ids[i]);
381

382 383 384
        wait_elem->num_completed--;	
        s->pending_waits = wait_elem;
        tw_rand_reverse_unif(lp->rng);
385 386 387 388 389
   }
}

/* 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)
390 391 392 393 394 395 396
{
	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;
397
	m->matched_op = 0;
398
	
399
	if(s->nw_id == TRACE)
400
		printf("\n %lf notify waits req id %d ", tw_now(lp), completed_req);
401 402 403

	if(!wait_elem)
		return 0;
404

405 406
	int op_type = wait_elem->mpi_op->op_type;

407
	if(op_type == CODES_WK_WAIT)
408 409 410
	{
		if(wait_elem->mpi_op->u.wait.req_id == completed_req)	
		  {
411
			m->saved_wait_time = s->wait_time;
412
			s->wait_time += (tw_now(lp) - wait_elem->start_time);
413
                        remove_req_id(&s->completed_reqs, completed_req);
414
	
415
			m->saved_pending_wait = wait_elem;			
416
            s->pending_waits = NULL;
417 418 419 420
			codes_issue_next_event(lp);	
			return 0;
		 }
	}
421
	else if(op_type == CODES_WK_WAITALL)
422
	{
423 424
	   int required_count = wait_elem->mpi_op->u.waits.count;
	  for(i = 0; i < required_count; i++)
425 426
	   {
	    if(wait_elem->mpi_op->u.waits.req_ids[i] == completed_req)
427
		{
428
			if(s->nw_id == TRACE)
429
				printCompletedQueue(s, lp);
430
			m->matched_op = 1;
431
			wait_elem->num_completed++;	
432
		}
433 434
	   }
	   
435 436
	  if(wait_elem->num_completed == required_count)
	   {
437
            if(s->nw_id == TRACE)
438 439 440 441
            {
                printf("\n %lf req %d completed %d", tw_now(lp), completed_req, wait_elem->num_completed);
                printCompletedQueue(s, lp);
            }
442 443
            m->matched_op = 2;
            m->saved_wait_time = s->wait_time;
444
            s->wait_time += (tw_now(lp) - wait_elem->start_time);
445
            m->saved_pending_wait = wait_elem;
446 447 448 449 450 451
            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
452
       }
453
    }
454 455 456
	return 0;
}

457
/* reverse handler of MPI wait operation */
458
static void codes_exec_mpi_wait_rc(nw_state* s, nw_message* m, tw_lp* lp, struct codes_workload_op * mpi_op)
459 460 461 462
{
    if(s->pending_waits)
     {
    	s->pending_waits = NULL;
463
	    return;
464 465 466
     }
   else
    {
467 468
 	    mpi_completed_queue_insert_op(&s->completed_reqs, mpi_op->u.wait.req_id);	
	    tw_rand_reverse_unif(lp->rng);	
469
        //rc_stack_pop(s->st);
470 471
    }
}
472 473

/* execute MPI wait operation */
474
static void codes_exec_mpi_wait(nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op)
475
{
Jonathan Jenkins's avatar
Jonathan Jenkins committed
476 477
    /* check in the completed receives queue if the request ID has already been completed.*/
    assert(!s->pending_waits);
478
    dumpi_req_id req_id = mpi_op->u.wait.req_id;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
479 480 481 482 483

    struct completed_requests* current = s->completed_reqs;
    while(current) {
        if(current->req_id == req_id) {
            remove_req_id(&s->completed_reqs, req_id);
484
            m->saved_wait_time = s->wait_time;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
485 486 487 488
            codes_issue_next_event(lp);
            return;
        }
        current = current->next;
489 490
    }

Jonathan Jenkins's avatar
Jonathan Jenkins committed
491
    /* If not, add the wait operation in the pending 'waits' list. */
492
    struct pending_waits* wait_op = malloc(sizeof(struct pending_waits));
493
    wait_op->mpi_op = mpi_op;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
494 495
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
496
    s->pending_waits = wait_op;
497

498
    //rc_stack_push(lp, wait_op, free, s->st);
499 500
}

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

        for( i = 0; i < count; i++)
        {
516
            req_id[i] = mpi_op->u.waits.req_ids[i];
517 518 519 520
            mpi_completed_queue_insert_op(&s->completed_reqs, req_id[i]);
        }
        tw_rand_reverse_unif(lp->rng);
   }
521 522
    else
    {
523 524 525 526 527
        s->pending_waits = NULL;
        assert(!s->pending_waits);
        if(lp->gid == TRACE)
            printf("\n %lf Nullifying codes waitall ", tw_now(lp));
   }
528
}
529 530
static void codes_exec_mpi_wait_all(
        nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op)
531 532
{
  //assert(!s->pending_waits);
533
  int count = mpi_op->u.waits.count;
534 535 536 537 538 539 540
  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)
    {
541
  	printf(" \n (%lf) MPI waitall posted %d count", tw_now(lp), mpi_op->u.waits.count);
542
	for(i = 0; i < count; i++)
543
		printf(" %d ", (int)mpi_op->u.waits.req_ids[i]);
544 545 546 547 548 549
   	printCompletedQueue(s, lp);	 
   }
  while(current) 
   {
	  for(i = 0; i < count; i++)
	   {
550
	     req_id[i] = mpi_op->u.waits.req_ids[i];
551
	     if(req_id[i] == current->req_id)
552
 		    num_completed++;
553 554 555 556 557
   	  }
	 current = current->next;
   }

  if(TRACE== lp->gid)
558
	  printf("\n %lf Num completed %d count %d ", tw_now(lp), num_completed, count);
559

560
  m->found_match = 0;
561
  if(count == num_completed)
562
  {
563
	m->found_match = 1;
564
	for( i = 0; i < count; i++)	
565
		remove_req_id(&s->completed_reqs, req_id[i]);
566

567 568 569 570 571 572
	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));
573
	  wait_op->mpi_op = mpi_op;  
574
	  wait_op->num_completed = num_completed;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
575
	  wait_op->start_time = tw_now(lp);
576
      //rc_stack_push(lp, wait_op, free, s->st);
577
      s->pending_waits = wait_op;
578 579 580 581
  }
}

/* request ID is being reused so delete it from the list once the matching is done */
582 583
static void remove_req_id(
        struct completed_requests** mpi_completed_queue, dumpi_req_id req_id)
584 585 586 587
{
	struct completed_requests* current = *mpi_completed_queue;

	if(!current)
588 589
		tw_error(TW_LOC, "\n REQ ID DOES NOT EXIST");
	
590
    if(current->req_id == req_id)
591 592 593 594 595 596 597 598 599 600 601 602
	{
		*mpi_completed_queue = current->next;
		free(current);
		return;
	}
	
	struct completed_requests* elem;
	while(current->next)
	{
	   elem = current->next;
	   if(elem->req_id == req_id)	
	     {
603 604 605 606
            current->next = elem->next;
            free(elem);
            return;
         }
607 608 609 610 611 612
	   current = current->next;	
	}
	return;
}

/* inserts mpi operation in the completed requests queue */
613 614
static void mpi_completed_queue_insert_op(
        struct completed_requests** mpi_completed_queue, dumpi_req_id req_id)
615 616 617 618 619 620 621 622
{
	struct completed_requests* reqs = malloc(sizeof(struct completed_requests));
	assert(reqs);

	reqs->req_id = req_id;

	if(!(*mpi_completed_queue))	
	{
623 624 625
        reqs->next = NULL;
        *mpi_completed_queue = reqs;
        return;
626 627 628
	}
	reqs->next = *mpi_completed_queue;
	*mpi_completed_queue = reqs;
629
	return;
630 631
}

632
/* 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. */
633 634
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
{
	/* insert mpi operation */
	struct mpi_msgs_queue* elem = malloc(sizeof(struct mpi_msgs_queue));
	assert(elem);

	elem->mpi_op = mpi_op;
641
    elem->next = NULL;
642 643 644 645 646 647 648

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

	if(mpi_queue->queue_tail)
	    mpi_queue->queue_tail->next = elem;
	
649
    mpi_queue->queue_tail = elem;
650 651 652 653 654 655
	mpi_queue->num_elems++;

	return;
}

/* match the send/recv operations */
656 657
static int match_receive(
        nw_state* s, tw_lp* lp, tw_lpid lpid, struct codes_workload_op* op1, struct codes_workload_op* op2)
658
{
659 660 661 662 663 664 665 666 667
        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);
668
                        s->recv_time += tw_now(lp) - op1->sim_start_time;
669 670 671 672
                        mpi_completed_queue_insert_op(&s->completed_reqs, op1->u.recv.req_id);
                        return 1;
                   }
        return -1;
673 674 675
}

/* used for reverse computation. removes the tail of the queue */
676
static int mpi_queue_remove_tail(tw_lpid lpid, struct mpi_queue_ptrs* mpi_queue)
677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710
{
	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. */
711
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)
712 713 714 715 716 717 718 719 720
{
	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 */
721
	int rcv_val = 0;
722
	if(mpi_op->op_type == CODES_WK_SEND || mpi_op->op_type == CODES_WK_ISEND)
723
	  {
724
		rcv_val = match_receive(s, lp, lp->gid, tmp->mpi_op, mpi_op);
725
		m->saved_matched_req = tmp->mpi_op->u.recv.req_id;  
726
	 }
727
	else if(mpi_op->op_type == CODES_WK_RECV || mpi_op->op_type == CODES_WK_IRECV)
728
	  {
729
		rcv_val = match_receive(s, lp, lp->gid, mpi_op, tmp->mpi_op);
730
	  	m->saved_matched_req = mpi_op->u.recv.req_id;
731 732
	  }
	if(rcv_val >= 0)
733
	{
734
		/* TODO: fix RC */
735
		m->ptr_match_op = tmp->mpi_op;
736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757
		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;
758
	   
759
	    if(mpi_op->op_type == CODES_WK_SEND || mpi_op->op_type == CODES_WK_ISEND)
760
	     {
761
		    rcv_val = match_receive(s, lp, lp->gid, elem->mpi_op, mpi_op);
762
	     	m->saved_matched_req = elem->mpi_op->u.recv.req_id; 
763
	     }
764
	    else if(mpi_op->op_type == CODES_WK_RECV || mpi_op->op_type == CODES_WK_IRECV)
765
	     {
766
            rcv_val = match_receive(s, lp, lp->gid, mpi_op, elem->mpi_op);
767
            m->saved_matched_req = mpi_op->u.recv.req_id;
768
         }
769
   	     if(rcv_val >= 0)
770
		 {
771
		    m->ptr_match_op = elem->mpi_op;
772
		    if(elem == mpi_queue->queue_tail)
773
			    mpi_queue->queue_tail = tmp;
774
		    
775 776 777 778
		    tmp->next = elem->next;

		    free(elem);
		    mpi_queue->num_elems--;
779
		
780
		    return indx;
781
		 }
782
	   tmp = tmp->next;
783
    }
784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802
	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 */
803 804
static void codes_exec_comp_delay(
        nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op)
805 806 807 808 809
{
	tw_event* e;
	tw_stime ts;
	nw_message* msg;

810 811
    s->compute_time += s_to_ns(mpi_op->u.delay.seconds);
    ts = s_to_ns(mpi_op->u.delay.seconds);
812
	ts += g_tw_lookahead + 0.1 + tw_rand_exponential(lp->rng, noise);
813 814 815 816 817 818 819 820 821
	
	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 */
822
static void codes_exec_mpi_recv_rc(nw_state* s, nw_message* m, tw_lp* lp, struct codes_workload_op * mpi_op)
823
{
824
	num_bytes_recvd -= mpi_op->u.recv.num_bytes;
825 826
	s->recv_time = m->saved_recv_time;
	if(m->found_match >= 0)
827
	  {
828 829
		s->recv_time = m->saved_recv_time;
		mpi_queue_update(s->arrival_queue, m->ptr_match_op, m->found_match);
830
		remove_req_id(&s->completed_reqs, mpi_op->u.recv.req_id);
831
		tw_rand_reverse_unif(lp->rng);
832
	  }
833
	else if(m->found_match < 0)
834
	    {
835
		mpi_queue_remove_tail(lp->gid, s->pending_recvs_queue);
836
		if(mpi_op->op_type == CODES_WK_IRECV)
837
			tw_rand_reverse_unif(lp->rng);
838 839 840 841
	    }
}

/* Execute MPI Irecv operation (non-blocking receive) */ 
842
static void codes_exec_mpi_recv(nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op)
843 844 845 846 847
{
/* 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. */

848
	m->saved_recv_time = s->recv_time;
849
	mpi_op->sim_start_time = tw_now(lp);
850
	num_bytes_recvd += mpi_op->u.recv.num_bytes;
851 852

	if(lp->gid == TRACE)
853
		printf("\n %lf codes exec mpi recv req id %d", tw_now(lp), (int)mpi_op->u.recv.req_id);
854
	
855
	int found_matching_sends = mpi_queue_remove_matching_op(s, lp, m, s->arrival_queue, mpi_op);
856 857 858
	
	/* save the req id inserted in the completed queue for reverse computation. */
	//m->matched_recv = req_id;
859 860 861

	if(found_matching_sends < 0)
	  {
862
		m->found_match = -1;
863 864 865
		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. */
866
		if(mpi_op->op_type == CODES_WK_IRECV)
867 868 869 870 871 872
		   {
			codes_issue_next_event(lp);	
			return;
		   }
		else
			printf("\n CODES MPI RECV OPERATION!!! ");
873
	  }
874
	else
875
	  {
876
	   	m->found_match = found_matching_sends;
877
        codes_issue_next_event(lp); 
878 879 880 881
	 }
}

/* executes MPI send and isend operations */
882
static void codes_exec_mpi_send(nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op)
883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907
{