model-net-mpi-wrklds.c 46.1 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
    
        unsigned long num_bytes_sent;
        unsigned long num_bytes_recvd;

        char output_buf[512];
135 136
};

137 138 139 140
/* 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). */
141 142
struct nw_message
{
143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163
    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;
164 165
};

166
/* executes MPI wait operation */
167 168
static void codes_exec_mpi_wait(
        nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op);
169
/* reverse of mpi wait function. */
170
static void codes_exec_mpi_wait_rc(
171
        nw_state* s, nw_message* m, tw_lp* lp, struct codes_workload_op * mpi_op);
172
/* executes MPI isend and send operations */
173 174
static void codes_exec_mpi_send(
        nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op);
175
/* execute MPI irecv operation */
176 177
static void codes_exec_mpi_recv(
        nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op);
178
/* reverse of mpi recv function. */
179
static void codes_exec_mpi_recv_rc(
180
        nw_state* s, nw_message* m, tw_lp* lp, struct codes_workload_op * mpi_op);
181
/* execute the computational delay */
182 183
static void codes_exec_comp_delay(
        nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op);
184
/* execute collective operation, currently only skips these operations. */
185 186
static void codes_exec_mpi_col(
        nw_state* s, tw_lp* lp);
187
/* gets the next MPI operation from the network-workloads API. */
188 189
static void get_next_mpi_operation(
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
190
/* reverse handler of get next mpi operation. */
191 192
static void get_next_mpi_operation_rc(
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
193 194 195 196
/* Makes a call to get_next_mpi_operation. */
static void codes_issue_next_event(tw_lp* lp);

///////////////////// HELPER FUNCTIONS FOR MPI MESSAGE QUEUE HANDLING ///////////////
197
/* upon arrival of local completion message, inserts operation in completed send queue */
198 199
static void update_send_completion_queue(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
200
/* reverse of the above function */
201 202
static void update_send_completion_queue_rc(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
203
/* upon arrival of an isend operation, updates the arrival queue of the network */
204 205
static void update_arrival_queue(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
206
/* reverse of the above function */
207 208
static void update_arrival_queue_rc(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
209
/* callback to a message sender for computing message time */
210 211 212 213 214
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);
215

216
/* insert MPI operation in the waiting queue*/
217 218
static void mpi_pending_queue_insert_op(
        struct mpi_queue_ptrs* mpi_queue, struct codes_workload_op* mpi_op);
219
/* remove completed request IDs from the queue for reuse. Reverse of above function. */
220 221
static void remove_req_id(
        struct completed_requests** requests, int16_t req_id);
222
/* remove MPI operation from the waiting queue.*/
223 224
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);
225
/* remove the tail of the MPI operation from waiting queue */
226 227
static int mpi_queue_remove_tail(
        tw_lpid lpid, struct mpi_queue_ptrs* mpi_queue);
228
/* insert completed MPI requests in the queue. */
229 230
static void mpi_completed_queue_insert_op(
        struct completed_requests** mpi_completed_queue, dumpi_req_id req_id);
231
/* notifies the wait operations (if any) about the completed receives and sends requests. */
232 233
static int notify_waits(
        nw_state* s, tw_bf* bf, tw_lp* lp, nw_message* m, dumpi_req_id req_id);
234
/* reverse of notify waits function. */
235 236
static void notify_waits_rc(
        nw_state* s, tw_bf* bf, tw_lp* lp, nw_message* m, dumpi_req_id completed_req);
237 238
/* conversion from seconds to eanaoseconds */
static tw_stime s_to_ns(tw_stime ns);
239

240 241 242 243 244
/* 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);
}
245

246 247 248 249 250 251 252 253 254 255 256 257
/* 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;
}

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

336
/* prints the elements of a queue (for debugging purposes). */
337 338
static void printCompletedQueue(nw_state* s, tw_lp* lp)
{
339
	   if(TRACE == s->nw_id)
340
	   {
341
	   	printf("\n %lf contents of completed operations queue ", tw_now(lp));
342 343 344 345 346 347 348 349 350
	   	struct completed_requests* current = s->completed_reqs;
	   	while(current)
	    	{
			printf(" %d ",current->req_id);
			current = current->next;
	   	}
	   }
}

351 352 353 354 355 356 357 358 359 360 361 362 363 364
/* 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;
    }
*/
365
  if(s->nw_id == TRACE)
366
	  printf("\n %lf reverse -- notify waits req id %d ", tw_now(lp), completed_req);
367
  
368
  printCompletedQueue(s, lp);
369
  
370
  if(m->matched_op == 1)
371 372
	s->pending_waits->num_completed--;
   /* if a wait-elem exists, it means the request ID has been matched*/
373
   if(m->matched_op == 2) 
374
    {
375
        if(s->nw_id == TRACE)
376 377 378 379
        {
            printf("\n %lf matched req id %d ", tw_now(lp), completed_req);
            printCompletedQueue(s, lp);
        }
380 381
        struct pending_waits* wait_elem = m->saved_pending_wait;
        s->wait_time = m->saved_wait_time;
382
        int count = wait_elem->mpi_op->u.waits.count; 
383

384 385
        for( i = 0; i < count; i++ )
            mpi_completed_queue_insert_op(&s->completed_reqs, wait_elem->mpi_op->u.waits.req_ids[i]);
386

387 388 389
        wait_elem->num_completed--;	
        s->pending_waits = wait_elem;
        tw_rand_reverse_unif(lp->rng);
390 391 392 393 394
   }
}

/* 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)
395 396 397 398 399 400 401
{
	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;
402
	m->matched_op = 0;
403
	
404
	if(s->nw_id == TRACE)
405
		printf("\n %lf notify waits req id %d ", tw_now(lp), completed_req);
406 407 408

	if(!wait_elem)
		return 0;
409

410 411
	int op_type = wait_elem->mpi_op->op_type;

412
	if(op_type == CODES_WK_WAIT)
413 414 415
	{
		if(wait_elem->mpi_op->u.wait.req_id == completed_req)	
		  {
416
			m->saved_wait_time = s->wait_time;
417
			s->wait_time += (tw_now(lp) - wait_elem->start_time);
418
                        remove_req_id(&s->completed_reqs, completed_req);
419
	
420
			m->saved_pending_wait = wait_elem;			
421
            s->pending_waits = NULL;
422 423 424 425
			codes_issue_next_event(lp);	
			return 0;
		 }
	}
426
	else if(op_type == CODES_WK_WAITALL)
427
	{
428 429
	   int required_count = wait_elem->mpi_op->u.waits.count;
	  for(i = 0; i < required_count; i++)
430 431
	   {
	    if(wait_elem->mpi_op->u.waits.req_ids[i] == completed_req)
432
		{
433
			if(s->nw_id == TRACE)
434
				printCompletedQueue(s, lp);
435
			m->matched_op = 1;
436
			wait_elem->num_completed++;	
437
		}
438 439
	   }
	   
440 441
	  if(wait_elem->num_completed == required_count)
	   {
442
            if(s->nw_id == TRACE)
443 444 445 446
            {
                printf("\n %lf req %d completed %d", tw_now(lp), completed_req, wait_elem->num_completed);
                printCompletedQueue(s, lp);
            }
447 448
            m->matched_op = 2;
            m->saved_wait_time = s->wait_time;
449
            s->wait_time += (tw_now(lp) - wait_elem->start_time);
450
            m->saved_pending_wait = wait_elem;
451 452 453 454 455 456
            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
457
       }
458
    }
459 460 461
	return 0;
}

462
/* reverse handler of MPI wait operation */
463
static void codes_exec_mpi_wait_rc(nw_state* s, nw_message* m, tw_lp* lp, struct codes_workload_op * mpi_op)
464 465 466
{
    if(s->pending_waits)
     {
Misbah Mubarak's avatar
Misbah Mubarak committed
467 468
        struct pending_waits * wait_op = s->pending_waits;
        free(wait_op);
469
    	s->pending_waits = NULL;
470
	    return;
471 472 473
     }
   else
    {
474 475
 	    mpi_completed_queue_insert_op(&s->completed_reqs, mpi_op->u.wait.req_id);	
	    tw_rand_reverse_unif(lp->rng);	
476
        //rc_stack_pop(s->st);
477 478
    }
}
479 480

/* execute MPI wait operation */
481
static void codes_exec_mpi_wait(nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op)
482
{
Jonathan Jenkins's avatar
Jonathan Jenkins committed
483 484
    /* check in the completed receives queue if the request ID has already been completed.*/
    assert(!s->pending_waits);
485
    dumpi_req_id req_id = mpi_op->u.wait.req_id;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
486 487 488 489 490

    struct completed_requests* current = s->completed_reqs;
    while(current) {
        if(current->req_id == req_id) {
            remove_req_id(&s->completed_reqs, req_id);
491
            m->saved_wait_time = s->wait_time;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
492 493 494 495
            codes_issue_next_event(lp);
            return;
        }
        current = current->next;
496 497
    }

Jonathan Jenkins's avatar
Jonathan Jenkins committed
498
    /* If not, add the wait operation in the pending 'waits' list. */
499
    struct pending_waits* wait_op = malloc(sizeof(struct pending_waits));
500
    wait_op->mpi_op = mpi_op;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
501 502
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
503
    s->pending_waits = wait_op;
504

505
    //rc_stack_push(lp, wait_op, free, s->st);
506 507
}

508
static void codes_exec_mpi_wait_all_rc(nw_state* s, nw_message* m, tw_lp* lp, struct codes_workload_op * mpi_op)
509
{
510
  if(s->nw_id == TRACE)
511
   {
512
       printf("\n %lf codes exec mpi waitall reverse %d ", tw_now(lp), m->found_match);
513 514
       printCompletedQueue(s, lp); 
   } 
515
  if(m->found_match)
516
    {
517
        int i;
518
        int count = mpi_op->u.waits.count;
519 520 521 522
        dumpi_req_id req_id[count];

        for( i = 0; i < count; i++)
        {
523
            req_id[i] = mpi_op->u.waits.req_ids[i];
524 525 526 527
            mpi_completed_queue_insert_op(&s->completed_reqs, req_id[i]);
        }
        tw_rand_reverse_unif(lp->rng);
   }
528 529
    else
    {
Misbah Mubarak's avatar
Misbah Mubarak committed
530 531
        struct pending_waits* wait_op = s->pending_waits;
        free(wait_op);
532 533 534 535 536
        s->pending_waits = NULL;
        assert(!s->pending_waits);
        if(lp->gid == TRACE)
            printf("\n %lf Nullifying codes waitall ", tw_now(lp));
   }
537
}
538 539
static void codes_exec_mpi_wait_all(
        nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op)
540 541
{
  //assert(!s->pending_waits);
542
  int count = mpi_op->u.waits.count;
543 544 545 546 547 548 549
  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)
    {
550
  	printf(" \n (%lf) MPI waitall posted %d count", tw_now(lp), mpi_op->u.waits.count);
551
	for(i = 0; i < count; i++)
552
		printf(" %d ", (int)mpi_op->u.waits.req_ids[i]);
553 554 555 556 557 558
   	printCompletedQueue(s, lp);	 
   }
  while(current) 
   {
	  for(i = 0; i < count; i++)
	   {
559
	     req_id[i] = mpi_op->u.waits.req_ids[i];
560
	     if(req_id[i] == current->req_id)
561
 		    num_completed++;
562 563 564 565 566
   	  }
	 current = current->next;
   }

  if(TRACE== lp->gid)
567
	  printf("\n %lf Num completed %d count %d ", tw_now(lp), num_completed, count);
568

569
  m->found_match = 0;
570
  if(count == num_completed)
571
  {
572
	m->found_match = 1;
573
	for( i = 0; i < count; i++)	
574
		remove_req_id(&s->completed_reqs, req_id[i]);
575

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

/* request ID is being reused so delete it from the list once the matching is done */
590 591
static void remove_req_id(
        struct completed_requests** mpi_completed_queue, dumpi_req_id req_id)
592 593 594 595
{
	struct completed_requests* current = *mpi_completed_queue;

	if(!current)
596 597
		tw_error(TW_LOC, "\n REQ ID DOES NOT EXIST");
	
598
    if(current->req_id == req_id)
599 600 601 602 603 604 605 606 607 608 609 610
	{
		*mpi_completed_queue = current->next;
		free(current);
		return;
	}
	
	struct completed_requests* elem;
	while(current->next)
	{
	   elem = current->next;
	   if(elem->req_id == req_id)	
	     {
611 612 613 614
            current->next = elem->next;
            free(elem);
            return;
         }
615 616 617 618 619 620
	   current = current->next;	
	}
	return;
}

/* inserts mpi operation in the completed requests queue */
621 622
static void mpi_completed_queue_insert_op(
        struct completed_requests** mpi_completed_queue, dumpi_req_id req_id)
623 624 625 626 627 628 629 630
{
	struct completed_requests* reqs = malloc(sizeof(struct completed_requests));
	assert(reqs);

	reqs->req_id = req_id;

	if(!(*mpi_completed_queue))	
	{
631 632 633
        reqs->next = NULL;
        *mpi_completed_queue = reqs;
        return;
634 635 636
	}
	reqs->next = *mpi_completed_queue;
	*mpi_completed_queue = reqs;
637
	return;
638 639
}

640
/* 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. */
641 642
static void mpi_pending_queue_insert_op(
        struct mpi_queue_ptrs* mpi_queue, struct codes_workload_op* mpi_op)
643 644 645 646 647 648
{
	/* insert mpi operation */
	struct mpi_msgs_queue* elem = malloc(sizeof(struct mpi_msgs_queue));
	assert(elem);

	elem->mpi_op = mpi_op;
649
    elem->next = NULL;
650 651 652 653 654 655 656

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

	if(mpi_queue->queue_tail)
	    mpi_queue->queue_tail->next = elem;
	
657
    mpi_queue->queue_tail = elem;
658 659 660 661 662 663
	mpi_queue->num_elems++;

	return;
}

/* match the send/recv operations */
664 665
static int match_receive(
        nw_state* s, tw_lp* lp, tw_lpid lpid, struct codes_workload_op* op1, struct codes_workload_op* op2)
666
{
667 668 669 670 671 672 673 674 675
        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);
676
                        s->recv_time += tw_now(lp) - op1->sim_start_time;
677 678 679 680
                        mpi_completed_queue_insert_op(&s->completed_reqs, op1->u.recv.req_id);
                        return 1;
                   }
        return -1;
681 682 683
}

/* used for reverse computation. removes the tail of the queue */
684
static int mpi_queue_remove_tail(tw_lpid lpid, struct mpi_queue_ptrs* mpi_queue)
685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718
{
	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. */
719
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)
720 721 722 723 724 725 726 727 728
{
	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 */
729
	int rcv_val = 0;
730
	if(mpi_op->op_type == CODES_WK_SEND || mpi_op->op_type == CODES_WK_ISEND)
731
	  {
732
		rcv_val = match_receive(s, lp, lp->gid, tmp->mpi_op, mpi_op);
733
		m->saved_matched_req = tmp->mpi_op->u.recv.req_id;  
734
	 }
735
	else if(mpi_op->op_type == CODES_WK_RECV || mpi_op->op_type == CODES_WK_IRECV)
736
	  {
737
		rcv_val = match_receive(s, lp, lp->gid, mpi_op, tmp->mpi_op);
738
	  	m->saved_matched_req = mpi_op->u.recv.req_id;
739 740
	  }
	if(rcv_val >= 0)
741
	{
742
		/* TODO: fix RC */
743
		m->ptr_match_op = tmp->mpi_op;
744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765
		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;
766
	   
767
	    if(mpi_op->op_type == CODES_WK_SEND || mpi_op->op_type == CODES_WK_ISEND)
768
	     {
769
		    rcv_val = match_receive(s, lp, lp->gid, elem->mpi_op, mpi_op);
770
	     	m->saved_matched_req = elem->mpi_op->u.recv.req_id; 
771
	     }
772
	    else if(mpi_op->op_type == CODES_WK_RECV || mpi_op->op_type == CODES_WK_IRECV)
773
	     {
774
            rcv_val = match_receive(s, lp, lp->gid, mpi_op, elem->mpi_op);
775
            m->saved_matched_req = mpi_op->u.recv.req_id;
776
         }
777
   	     if(rcv_val >= 0)
778
		 {
779
		    m->ptr_match_op = elem->mpi_op;
780
		    if(elem == mpi_queue->queue_tail)
781
			    mpi_queue->queue_tail = tmp;
782
		    
783 784 785 786
		    tmp->next = elem->next;

		    free(elem);
		    mpi_queue->num_elems--;
787
		
788
		    return indx;
789
		 }
790
	   tmp = tmp->next;
791
    }
792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810
	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 */
811 812
static void codes_exec_comp_delay(
        nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op)
813 814 815 816 817
{
	tw_event* e;
	tw_stime ts;
	nw_message* msg;

818 819
    s->compute_time += s_to_ns(mpi_op->u.delay.seconds);
    ts = s_to_ns(mpi_op->u.delay.seconds);
820
	ts += g_tw_lookahead + 0.1 + tw_rand_exponential(lp->rng, noise);
821 822 823 824 825 826 827 828 829
	
	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 */
830
static void codes_exec_mpi_recv_rc(nw_state* s, nw_message* m, tw_lp* lp, struct codes_workload_op * mpi_op)
831
{
832
	num_bytes_recvd -= mpi_op->u.recv.num_bytes;
833
        s->num_bytes_recvd -= mpi_op->u.recv.num_bytes; 
834 835
	s->recv_time = m->saved_recv_time;
	if(m->found_match >= 0)
836
	  {
837 838
		s->recv_time = m->saved_recv_time;
		mpi_queue_update(s->arrival_queue, m->ptr_match_op, m->found_match);
839
		remove_req_id(&s->completed_reqs, mpi_op->u.recv.req_id);
840
		tw_rand_reverse_unif(lp->rng);
841
	  }
842
	else if(m->found_match < 0)
843
	    {
844
		mpi_queue_remove_tail(lp->gid, s->pending_recvs_queue);
845
		if(mpi_op->op_type == CODES_WK_IRECV)
846
			tw_rand_reverse_unif(lp->rng);
847 848 849 850
	    }
}

/* Execute MPI Irecv operation (non-blocking receive) */ 
851
static void codes_exec_mpi_recv(nw_state* s, tw_lp* lp, nw_message * m, struct codes_workload_op * mpi_op)
852 853 854 855 856
{
/* 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. */

857
	m->saved_recv_time = s->recv_time;
858
	mpi_op->sim_start_time = tw_now(lp);
859
	num_bytes_recvd += mpi_op->u.recv.num_bytes;
860
        s->num_bytes_recvd += mpi_op->u.recv.num_bytes;
861 862

	if(lp->gid == TRACE)
863
		printf("\n %lf codes exec mpi recv req id %d", tw_now(lp), (int)mpi_op->u.recv.req_id);
864
	
865
	int found_matching_sends = mpi_queue_remove_matching_op(s, lp, m, s->arrival_queue, mpi_op);
866 867 868
	
	/* save the req id inserted in the completed queue for reverse computation. */
	//m->matched_recv = req_id;
869 870 871

	if(found_matching_sends < 0)
	  {
872
		m->found_match = -1;
873 874 875
		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. */
876
		if(mpi_op->op_type == CODES_WK_IRECV)
877 878 879 880 881 882
		   {
			codes_issue_next_event(lp);	
			return;
		   }
		else
			printf("\n CODES MPI RECV OPERATION!!! ");
883
	  }
884
	else
885
	  {
886
	   	m->found_match = found_matching_sends;
887
        codes_issue_next_event(lp);