model-net-mpi-replay.c 40.5 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
/*
 * Copyright (C) 2014 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
 */
#include <ross.h>
#include <inttypes.h>

#include "codes/codes-workload.h"
#include "codes/codes.h"
#include "codes/configuration.h"
#include "codes/codes_mapping.h"
#include "codes/model-net.h"
#include "codes/rc-stack.h"
#include "codes/quicklist.h"

17
18
/* turning on track lp will generate a lot of output messages */
#define TRACK_LP -1
19
#define TRACE -1
20
#define MAX_WAIT_REQS 512
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35

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

/* 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;

typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
36
typedef int32_t dumpi_req_id;
37
38
39
40
41

static int net_id = 0;
static float noise = 5.0;
static int num_net_lps, num_nw_lps;

42
43
44
45
#define CS_LP_DBG 0
#define lprintf(_fmt, ...) \
        do {if (CS_LP_DBG) printf(_fmt, __VA_ARGS__);} while (0)

46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
long long num_bytes_sent=0;
long long num_bytes_recvd=0;

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;

/* 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;

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

/* 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). */
enum MPI_NW_EVENTS
{
	MPI_OP_GET_NEXT=1,
	MPI_SEND_ARRIVED,
    MPI_SEND_ARRIVED_CB, // for tracking message times on sender
	MPI_SEND_POSTED,
};

/* stores pointers of pending MPI operations to be matched with their respective sends/receives. */
struct mpi_msgs_queue
{
    int op_type;
    int tag;
    int source_rank;
    int dest_rank;
    int num_bytes;
    tw_stime req_init_time;
	dumpi_req_id req_id;
    struct qlist_head ql;
};

/* stores request IDs of completed MPI operations (Isends or Irecvs) */
struct completed_requests
{
	dumpi_req_id req_id;
    struct qlist_head ql;
};

/* for wait operations, store the pending operation and number of completed waits so far. */
struct pending_waits
{
    int op_type;
94
    int32_t req_ids[MAX_WAIT_REQS];
95
	int num_completed;
96
97
	int count;
    tw_stime start_time;
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
    struct qlist_head ql;
};

typedef struct mpi_msgs_queue mpi_msgs_queue;
typedef struct completed_requests completed_requests;
typedef struct pending_waits pending_waits;

/* 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;

    struct rc_stack * processed_ops;
113
    struct rc_stack * matched_reqs;
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140

    /* count of sends, receives, collectives and delays */
	unsigned long num_sends;
	unsigned long num_recvs;
	unsigned long num_cols;
	unsigned long num_delays;
	unsigned long num_wait;
	unsigned long num_waitall;
	unsigned long num_waitsome;

	/* time spent by the LP in executing the app trace*/
	double start_time;
	double elapsed_time;
	/* time spent in compute operations */
	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;
	/* FIFO for isend messages arrived on destination */
	struct qlist_head arrival_queue;
	/* FIFO for irecv messages posted but not yet matched with send operations */
	struct qlist_head pending_recvs_queue;
	/* List of completed send/receive requests */
	struct qlist_head completed_reqs;
141
142
143

    /* Pending wait operation */
    struct pending_waits * wait_op;
144
145
146
147
148

    unsigned long num_bytes_sent;
    unsigned long num_bytes_recvd;

    char output_buf[512];
149
150
151
152
153
154
155
156
};

/* 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). */
struct nw_message
{
157
   // forward message handler 
158
   int msg_type;
159
160
   int op_type;
   
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
   struct
   {
       tw_lpid src_rank;
       tw_lpid dest_rank;
       int num_bytes;
       int num_matched;
       int data_type;
       double sim_start_time;
       // for callbacks - time message was received
       double msg_send_time;
       int16_t req_id;   
       int tag;
       int found_match;
       short wait_completed;
   } fwd;
   struct
   {
       double saved_send_time;
       double saved_recv_time;
       double saved_wait_time;
       double saved_delay;
       int saved_num_bytes;
183
       struct codes_workload_op * saved_op;
184
   } rc;
185
186
187
188
};

/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
189
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
190
191
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
192
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
193
194
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
195
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp);
196
197
/* execute the computational delay */
static void codes_exec_comp_delay(
198
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
/* gets the next MPI operation from the network-workloads API. */
static void get_next_mpi_operation(
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
/* reverse handler of get next mpi operation. */
static void get_next_mpi_operation_rc(
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
/* Makes a call to get_next_mpi_operation. */
static void codes_issue_next_event(tw_lp* lp);
/* reverse handler of next operation */
static void codes_issue_next_event_rc(tw_lp* lp);


///////////////////// HELPER FUNCTIONS FOR MPI MESSAGE QUEUE HANDLING ///////////////
/* upon arrival of local completion message, inserts operation in completed send queue */
/* upon arrival of an isend operation, updates the arrival queue of the network */
214
215
216
217
218
219
220
221
static void update_completed_queue(
        nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp, dumpi_req_id req_id);
/* reverse of the above function */
static void update_completed_queue_rc(
        nw_state*s,
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp);
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
static void update_arrival_queue(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
/* reverse of the above function */
static void update_arrival_queue_rc(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
/* callback to a message sender for computing message time */
static void update_message_time(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
/* reverse for computing message time */
static void update_message_time_rc(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);

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

237
/* Debugging functions, may generate unused function warning */
238
static void print_waiting_reqs(int32_t * reqs, int count)
239
240
241
242
{
    printf("\n Waiting reqs: ");
    int i;
    for(i = 0; i < count; i++ )
243
        printf(" %d ", reqs[i]);
244
245
246
247
248
249
250
251
252
}
static void print_completed_queue(struct qlist_head * head)
{
    printf("\n Completed queue: ");
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
      qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, completed_requests, ql);
253
            printf(" %d ", current->req_id);
254
255
       }
}
256
static int clear_completed_reqs(nw_state * s,
257
        tw_lp * lp,
258
        int32_t * reqs, int count)
259
{
260
    int i, matched = 0;
261
262
263
264
265
266
267
268
269
    for( i = 0; i < count; i++)
    {
      struct qlist_head * ent = NULL;
      qlist_for_each(ent, &s->completed_reqs)
       {
            struct completed_requests* current = 
                qlist_entry(ent, completed_requests, ql);
            if(current->req_id == reqs[i])
            {
270
                ++matched;
271
272
273
274
275
                qlist_del(&current->ql);
                rc_stack_push(lp, current, free, s->matched_reqs);
            }
       }
    }
276
    return matched;
277
278
}
static void add_completed_reqs(nw_state * s, 
279
280
        tw_lp * lp,
        int count)
281
282
283
284
285
286
287
288
{
    int i;
    for( i = 0; i < count; i++)
    {
       struct completed_requests * req = rc_stack_pop(s->matched_reqs); 
       qlist_add(&req->ql, &s->completed_reqs); 
    }
}
289
290
291
292
293
294
/* 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);
}

295
296
297
static int notify_posted_wait(nw_state* s,
        tw_bf * bf, nw_message * m, tw_lp * lp, 
        dumpi_req_id completed_req)
298
{
299
300
    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;
301

302
303
    m->fwd.wait_completed = 0;
    
304
305
    if(!wait_elem)
        return 0;
306
    
307
    int op_type = wait_elem->op_type;
308

309
310
311
312
313
314
315
316
317
318
319
320
321
    if(op_type == CODES_WK_WAIT &&
            (wait_elem->req_ids[0] == completed_req))
    {
            wait_completed = 1;
    }
    else if(op_type == CODES_WK_WAITALL 
            || op_type == CODES_WK_WAITANY 
            || op_type == CODES_WK_WAITSOME)
    {
        int i;
        for(i = 0; i < wait_elem->count; i++)
        {
            if(wait_elem->req_ids[i] == completed_req)
322
            {
323
                wait_elem->num_completed++;
324
                if(wait_elem->num_completed > wait_elem->count)
325
                    printf("\n Num completed %d count %d LP %llu ",
326
327
328
                            wait_elem->num_completed,
                            wait_elem->count,
                            lp->gid);
329
330
331
332
                assert(wait_elem->num_completed <= wait_elem->count);
                if(wait_elem->num_completed == wait_elem->count)
                    wait_completed = 1;
           
333
                m->fwd.wait_completed = 1;
334
            }
335
        }
336
    }
337
    return wait_completed;
338
}
339

340
/* reverse handler of MPI wait operation */
341
static void codes_exec_mpi_wait_rc(nw_state* s, tw_lp* lp)
342
{
343
    if(s->wait_op)
344
     {
345
346
347
         struct pending_waits * wait_op = s->wait_op;
         free(wait_op);
         s->wait_op = NULL;
348
349
350
351
     }
   else
    {
        codes_issue_next_event_rc(lp);
352
        completed_requests * qi = rc_stack_pop(s->processed_ops);
353
        qlist_add(&qi->ql, &s->completed_reqs);
354
    }
355
    return;
356
}
357

358
/* execute MPI wait operation */
359
static void codes_exec_mpi_wait(nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op)
360
{
361
362
    /* check in the completed receives queue if the request ID has already been completed.*/
    assert(!s->wait_op);
363
    dumpi_req_id req_id = mpi_op->u.wait.req_id;
364
    struct completed_requests* current = NULL;
365

366
367
368
369
370
371
372
    struct qlist_head * ent = NULL;
    qlist_for_each(ent, &s->completed_reqs)
    {
        current = qlist_entry(ent, completed_requests, ql);
        if(current->req_id == req_id)
        {
            qlist_del(&current->ql);
373
            rc_stack_push(lp, current, free, s->processed_ops);
374
375
376
377
            codes_issue_next_event(lp);
            return;
        }
    }
378
379
380
381
382
    /* If not, add the wait operation in the pending 'waits' list. */
    struct pending_waits* wait_op = malloc(sizeof(struct pending_waits));
    wait_op->op_type = mpi_op->op_type;
    wait_op->req_ids[0] = req_id;
    wait_op->count = 1;
383
384
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
385
    s->wait_op = wait_op;
386

387
    return;
388
389
}

390
391
392
393
static void codes_exec_mpi_wait_all_rc(
        nw_state* s, 
        tw_bf * bf,
        nw_message * m,
394
        tw_lp* lp)
395
{
396
397
398
399
400
401
402
403
  if(s->wait_op)
  {
      struct pending_waits * wait_op = s->wait_op;
      free(wait_op);
      s->wait_op = NULL;
  }
  else
  {
404
      add_completed_reqs(s, lp, m->fwd.num_matched);
405
406
407
      codes_issue_next_event_rc(lp);
  }
  return;
408
409
}
static void codes_exec_mpi_wait_all(
410
411
412
413
414
        nw_state* s, 
        tw_bf * bf,
        nw_message * m,
        tw_lp* lp, 
        struct codes_workload_op * mpi_op)
415
416
{
  int count = mpi_op->u.waits.count;
417
418
  /* If the count is not less than max wait reqs then stop */
  assert(count < MAX_WAIT_REQS);
419

420
  int i = 0, num_matched = 0;
421
  m->fwd.num_matched = 0;
422

423
  /*if(lp->gid == TRACK)
424
  {
425
      printf("\n MPI Wait all posted ");
426
427
      print_waiting_reqs(mpi_op->u.waits.req_ids, count);
      print_completed_queue(&s->completed_reqs);
428
  }*/
429
430
431
432
433
434
435
436
437
438
439
440
441
      /* check number of completed irecvs in the completion queue */ 
  for(i = 0; i < count; i++)
  {
      dumpi_req_id req_id = mpi_op->u.waits.req_ids[i];
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
      qlist_for_each(ent, &s->completed_reqs)
       {
            current = qlist_entry(ent, completed_requests, ql);
            if(current->req_id == req_id)
                num_matched++;
       }
  }
442

443
  m->fwd.found_match = num_matched;
444
445
446
447
  if(num_matched == count)
  {
    /* No need to post a MPI Wait all then, issue next event */
      /* Remove all completed requests from the list */
448
449
450
      m->fwd.num_matched = clear_completed_reqs(s, lp, mpi_op->u.waits.req_ids, count);
      struct pending_waits* wait_op = s->wait_op;
      free(wait_op);
451
452
      s->wait_op = NULL;
      codes_issue_next_event(lp);
453
454
  }
  else
455
456
457
458
459
460
461
462
463
464
465
  {
      /* If not, add the wait operation in the pending 'waits' list. */
	  struct pending_waits* wait_op = malloc(sizeof(struct pending_waits));
	  wait_op->count = count;
      wait_op->op_type = mpi_op->op_type;
      assert(count < MAX_WAIT_REQS);

      for(i = 0; i < count; i++)
          wait_op->req_ids[i] =  mpi_op->u.waits.req_ids[i];

	  wait_op->num_completed = num_matched;
466
	  wait_op->start_time = tw_now(lp);
467
      s->wait_op = wait_op;
468
  }
469
470
  return;
}
471
472
473
474

/* 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. */
475
476
477
478
479
static int rm_matching_rcv(nw_state * ns, 
        tw_bf * bf,
        nw_message * m, 
        tw_lp * lp, 
        mpi_msgs_queue * qitem)
480
481
{
    int matched = 0;
482
    int index = 0;
483
484
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;
485

486
487
    qlist_for_each(ent, &ns->pending_recvs_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
488
        if((qi->num_bytes == qitem->num_bytes)
489
490
                && ((qi->tag == qitem->tag) || qi->tag == -1)
                && ((qi->source_rank == qitem->source_rank) || qi->source_rank == -1))
491
492
493
494
        {
            matched = 1;
            break;
        }
495
        ++index;
496
497
498
499
    }
    
    if(matched)
    {
500
        m->rc.saved_recv_time = ns->recv_time;
501
        ns->recv_time += (tw_now(lp) - qi->req_init_time);
502
503
504
        
        if(qi->op_type == CODES_WK_IRECV)
            update_completed_queue(ns, bf, m, lp, qi->req_id);
505
        
506
        qlist_del(&qi->ql);
507
508
        
        rc_stack_push(lp, qi, free, ns->processed_ops);
509
        return index;
510
511
512
513
    }
    return -1;
}

514
515
516
517
static int rm_matching_send(nw_state * ns, 
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp, mpi_msgs_queue * qitem)
518
519
520
521
522
{
    int matched = 0;
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;

523
    int index = 0;
524
525
    qlist_for_each(ent, &ns->arrival_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
526
        if((qi->num_bytes == qitem->num_bytes) 
527
528
529
530
531
532
                && (qi->tag == qitem->tag || qitem->tag == -1)
                && ((qi->source_rank == qitem->source_rank) || qitem->source_rank == -1))
        {
            matched = 1;
            break;
        }
533
        ++index;
534
535
536
537
    }

    if(matched)
    {
538
        m->rc.saved_recv_time = ns->recv_time;
539
540
541
542
543
        ns->recv_time += (tw_now(lp) - qitem->req_init_time);

        if(qitem->op_type == CODES_WK_IRECV)
            update_completed_queue(ns, bf, m, lp, qitem->req_id);

544
        qlist_del(&qi->ql);
545

546
        return index;
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
    }
    return -1;
}
static void codes_issue_next_event_rc(tw_lp * lp)
{
	    tw_rand_reverse_unif(lp->rng);	
}

/* Trigger getting next event at LP */
static void codes_issue_next_event(tw_lp* lp)
{
   tw_event *e;
   nw_message* msg;

   tw_stime ts;

   ts = g_tw_lookahead + 0.1 + tw_rand_exponential(lp->rng, noise);
   e = tw_event_new( lp->gid, ts, lp );
   msg = tw_event_data(e);

   msg->msg_type = MPI_OP_GET_NEXT;
   tw_event_send(e);
}

/* Simulate delays between MPI operations */
static void codes_exec_comp_delay(
573
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
574
575
576
577
578
{
	tw_event* e;
	tw_stime ts;
	nw_message* msg;

579
    m->rc.saved_delay = s->compute_time;
580
581
    s->compute_time += s_to_ns(mpi_op->u.delay.seconds);
    ts = s_to_ns(mpi_op->u.delay.seconds);
582
583
584
585
586
587
588
589
590
591
592

	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); 
                
}

/* reverse computation operation for MPI irecv */
593
594
595
596
static void codes_exec_mpi_recv_rc(
        nw_state* ns, 
        tw_bf * bf, 
        nw_message* m, 
597
        tw_lp* lp)
598
{
599
	num_bytes_recvd -= m->rc.saved_num_bytes;
600
601
	ns->recv_time = m->rc.saved_recv_time;
	if(m->fwd.found_match >= 0)
602
	  {
603
		ns->recv_time = m->rc.saved_recv_time;
604
        int queue_count = qlist_count(&ns->arrival_queue); 
605
        
606
        mpi_msgs_queue * qi = rc_stack_pop(ns->processed_ops);	
607
       
608
        if(!m->fwd.found_match)
609
610
611
        {
            qlist_add(&qi->ql, &ns->arrival_queue);
        }
612
        else if(m->fwd.found_match >= queue_count)
613
614
615
        {
            qlist_add_tail(&qi->ql, &ns->arrival_queue);
        }
616
        else if(m->fwd.found_match > 0 && m->fwd.found_match < queue_count) 
617
        {
618
619
620
621
            int index = 1;
            struct qlist_head * ent = NULL;
            qlist_for_each(ent, &ns->arrival_queue)
            {
622
               if(index == m->fwd.found_match)
623
624
625
626
627
628
               {
                 qlist_add(&qi->ql, ent);
                 break;
               }
               index++; 
            }
629
        }
630
        if(qi->op_type == CODES_WK_IRECV)
631
        {
632
            update_completed_queue_rc(ns, bf, m, lp);
633
        }
634
635
        codes_issue_next_event_rc(lp);
      }
636
	else if(m->fwd.found_match < 0)
637
	    {
638
639
640
641
            struct qlist_head * ent = qlist_pop_back(&ns->pending_recvs_queue); 
            mpi_msgs_queue * qi = qlist_entry(ent, mpi_msgs_queue, ql);
            free(qi);
            
642
            if(m->op_type == CODES_WK_IRECV)
643
                codes_issue_next_event_rc(lp);
644
645
646
647
	    }
}

/* Execute MPI Irecv operation (non-blocking receive) */ 
648
649
650
651
652
653
static void codes_exec_mpi_recv(
        nw_state* s, 
        tw_bf * bf,
        nw_message * m, 
        tw_lp* lp, 
        struct codes_workload_op * mpi_op)
654
655
656
657
658
{
/* 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. */

659
	m->rc.saved_recv_time = s->recv_time;
660
661
    m->rc.saved_num_bytes = mpi_op->u.recv.num_bytes;

662
663
664
665
666
667
668
669
670
671
672
	num_bytes_recvd += mpi_op->u.recv.num_bytes;

    mpi_msgs_queue * recv_op = (mpi_msgs_queue*) malloc(sizeof(mpi_msgs_queue));
    recv_op->req_init_time = tw_now(lp);
    recv_op->op_type = mpi_op->op_type;
    recv_op->source_rank = mpi_op->u.recv.source_rank;
    recv_op->dest_rank = mpi_op->u.recv.dest_rank;
    recv_op->num_bytes = mpi_op->u.recv.num_bytes;
    recv_op->tag = mpi_op->u.recv.tag;
    recv_op->req_id = mpi_op->u.recv.req_id;

673
674
675
676
    if(s->nw_id == TRACK_LP)
        printf("\n Receive op posted num bytes %d source %d ", recv_op->num_bytes,
                recv_op->source_rank);

677
	int found_matching_sends = rm_matching_send(s, bf, m, lp, recv_op);
678
679
680
681

	/* save the req id inserted in the completed queue for reverse computation. */
	if(found_matching_sends < 0)
	  {
682
	   	  m->fwd.found_match = -1;
683
          qlist_add_tail(&recv_op->ql, &s->pending_recvs_queue);
684
685
686
687
688
689
690
	
	       /* for mpi irecvs, this is a non-blocking receive so just post it and move on with the trace read. */
		if(mpi_op->op_type == CODES_WK_IRECV)
		   {
			codes_issue_next_event(lp);	
			return;
		   }
691
      }
692
693
	else
	  {
694
        m->fwd.found_match = found_matching_sends;
695
        codes_issue_next_event(lp); 
696
697
	    rc_stack_push(lp, recv_op, free, s->processed_ops);
      }
698
699
700
}

/* executes MPI send and isend operations */
701
702
703
704
705
static void codes_exec_mpi_send(nw_state* s, 
        tw_bf * bf,
        nw_message * m,
        tw_lp* lp, 
        struct codes_workload_op * mpi_op)
706
{
707
    m->rc.saved_num_bytes = mpi_op->u.send.num_bytes;
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
	/* 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;
730
    s->num_bytes_sent += mpi_op->u.send.num_bytes;
731
732
733
734

	nw_message local_m;
	nw_message remote_m;

735
736
737
    local_m.fwd.sim_start_time = tw_now(lp);
    local_m.fwd.dest_rank = mpi_op->u.send.dest_rank;
    local_m.fwd.src_rank = mpi_op->u.send.source_rank;
738
    local_m.op_type = mpi_op->op_type; 
739
    local_m.msg_type = MPI_SEND_POSTED;
740
741
742
    local_m.fwd.tag = mpi_op->u.send.tag;
    local_m.fwd.num_bytes = mpi_op->u.send.num_bytes;
    local_m.fwd.req_id = mpi_op->u.send.req_id;
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760

    remote_m = local_m;
	remote_m.msg_type = MPI_SEND_ARRIVED;

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

	/* isend executed, now get next MPI operation from the queue */ 
	if(mpi_op->op_type == CODES_WK_ISEND)
	   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));
}

761
762
static void update_completed_queue_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
763
   
764
765
766
    if(bf->c0)
    {
       struct qlist_head * ent = qlist_pop_back(&s->completed_reqs);
767
768
769
770
771
772
773

        completed_requests * req = qlist_entry(ent, completed_requests, ql);
      /*if(lp->gid == TRACK)
      {
          printf("\n After popping %ld ", req->req_id);
        print_completed_queue(&s->completed_reqs);
      }*/
774
775
776
777
       free(req);
    }
    else if(bf->c1)
    {
778
       struct pending_waits* wait_elem = rc_stack_pop(s->processed_ops); 
779
       s->wait_op = wait_elem;
780
781
       s->wait_time = m->rc.saved_wait_time;
       add_completed_reqs(s, lp, m->fwd.num_matched);
782
783
       codes_issue_next_event_rc(lp); 
    }
784
785
    if(m->fwd.wait_completed > 0)
           s->wait_op->num_completed--;
786
787
788
789
790
791
792
793
794
795
}

static void update_completed_queue(nw_state* s, 
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp,
        dumpi_req_id req_id)
{
    bf->c0 = 0;
    bf->c1 = 0;
796
    m->fwd.num_matched = 0; 
797

798
799
    int waiting = 0;
    waiting = notify_posted_wait(s, bf, m, lp, req_id);
800
801
802
803
804
805
806
 
    if(!waiting)
    {
        bf->c0 = 1;
        completed_requests * req = malloc(sizeof(completed_requests));
        req->req_id = req_id;
        qlist_add_tail(&req->ql, &s->completed_reqs);
807
808
809
810
811
812
       
        /*if(lp->gid == TRACK)
        {
            printf("\n Forward mode adding %ld ", req_id);
            print_completed_queue(&s->completed_reqs);
        }*/
813
814
815
816
    }
    else 
     {
            bf->c1 = 1;
817
818
            m->fwd.num_matched = clear_completed_reqs(s, lp, s->wait_op->req_ids, s->wait_op->count);
            m->rc.saved_wait_time = s->wait_time;
819
            s->wait_time += (tw_now(lp) - s->wait_op->start_time);  
820
821
822

            struct pending_waits* wait_elem = s->wait_op;
            rc_stack_push(lp, wait_elem, free, s->processed_ops);
823
824
825
826
827
            s->wait_op = NULL;
            codes_issue_next_event(lp); 
     }
}

828
/* reverse handler for updating arrival queue function */
829
830
831
static void update_arrival_queue_rc(nw_state* s, 
        tw_bf * bf, 
        nw_message * m, tw_lp * lp)
832
{
833
	s->recv_time = m->rc.saved_recv_time;
834
835
    s->num_bytes_recvd -= m->fwd.num_bytes;

836
    codes_local_latency_reverse(lp);
837
  
838
    if(m->fwd.found_match >= 0)
839
	{
840
        mpi_msgs_queue * qi = rc_stack_pop(s->processed_ops);
841
842
        int queue_count = qlist_count(&s->pending_recvs_queue); 

843
        if(!m->fwd.found_match)
844
845
846
        {
            qlist_add(&qi->ql, &s->pending_recvs_queue);
        }
847
        else if(m->fwd.found_match >= queue_count)
848
        {
849
850
            qlist_add_tail(&qi->ql, &s->pending_recvs_queue);
        }
851
        else if(m->fwd.found_match > 0 && m->fwd.found_match < queue_count)
852
853
854
855
856
        {
            int index = 1;
            struct qlist_head * ent = NULL;
            qlist_for_each(ent, &s->pending_recvs_queue)
            {
857
               if(index == m->fwd.found_match)
858
859
860
861
862
863
               {
                 qlist_add(&qi->ql, ent);
                 break;
               }
               index++; 
            }
864
        }
865
866
        if(qi->op_type == CODES_WK_IRECV)
            update_completed_queue_rc(s, bf, m, lp);
867
    }
868
	else if(m->fwd.found_match < 0)
869
870
871
872
873
874
875
876
877
878
	{
	    struct qlist_head * ent = qlist_pop_back(&s->arrival_queue); 
        mpi_msgs_queue * qi = qlist_entry(ent, mpi_msgs_queue, ql);
        free(qi);
    }
}

/* 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)
{
879
	m->rc.saved_recv_time = s->recv_time;
880
    s->num_bytes_recvd += m->fwd.num_bytes;
881
882
883

    // send a callback to the sender to increment times
    tw_event *e_callback =
884
        tw_event_new(rank_to_lpid(m->fwd.src_rank),
885
886
887
                codes_local_latency(lp), lp);
    nw_message *m_callback = tw_event_data(e_callback);
    m_callback->msg_type = MPI_SEND_ARRIVED_CB;
888
    m_callback->fwd.msg_send_time = tw_now(lp) - m->fwd.sim_start_time;
889
890
891
892
    tw_event_send(e_callback);

    /* Now reconstruct the queue item */
    mpi_msgs_queue * arrived_op = (mpi_msgs_queue *) malloc(sizeof(mpi_msgs_queue));
893
    arrived_op->req_init_time = m->fwd.sim_start_time;
894
    arrived_op->op_type = m->op_type;
895
896
897
898
    arrived_op->source_rank = m->fwd.src_rank;
    arrived_op->dest_rank = m->fwd.dest_rank;
    arrived_op->num_bytes = m->fwd.num_bytes;
    arrived_op->tag = m->fwd.tag;
899

900
    if(s->nw_id == TRACK_LP)
901
        printf("\n Send op arrived source rank %d num bytes %d ", arrived_op->source_rank,
902
903
                arrived_op->num_bytes);

904
    int found_matching_recv = rm_matching_rcv(s, bf, m, lp, arrived_op);
905
906
907

    if(found_matching_recv < 0)
    {
908
        m->fwd.found_match = -1;
909
910
911
912
        qlist_add_tail(&arrived_op->ql, &s->arrival_queue);
    }
    else
    {
913
        m->fwd.found_match = found_matching_recv;
914
        free(arrived_op);
915
916
917
918
919
920
921
922
    }
}
static void update_message_time(
        nw_state * s,
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp)
{
923
924
    m->rc.saved_send_time = s->send_time;
    s->send_time += m->fwd.msg_send_time;
925
926
927
928
929
930
931
932
}

static void update_message_time_rc(
        nw_state * s,
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp)
{
933
    s->send_time = m->rc.saved_send_time;
934
935
936
937
938
939
940
941
942
943
}

/* 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 */
   char * params = NULL;
   dumpi_trace_params params_d;
  
   memset(s, 0, sizeof(*s));
944
   s->nw_id = codes_mapping_get_lp_relative_id(lp->gid, 0, 0);
945
946
947
948

   if(!num_net_traces) 
	num_net_traces = num_net_lps;

949
950
   assert(num_net_traces <= num_net_lps);

951
952
953
954
955
956
957
958
959
960
961
962
   if (strcmp(workload_type, "dumpi") == 0){
       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)
	    return;

   /* Initialize the RC stack */
   rc_stack_create(&s->processed_ops);
963
   rc_stack_create(&s->matched_reqs);
964
965

   assert(s->processed_ops != NULL);
966
   assert(s->matched_reqs != NULL);
967
968
969
970
971
972
973
974
975
976

   wrkld_id = codes_workload_load("dumpi-trace-workload", params, 0, (int)s->nw_id);

   INIT_QLIST_HEAD(&s->arrival_queue);
   INIT_QLIST_HEAD(&s->pending_recvs_queue);
   INIT_QLIST_HEAD(&s->completed_reqs);

   /* clock starts when the first event is processed */
   s->start_time = tw_now(lp);
   codes_issue_next_event(lp);
977
978
   s->num_bytes_sent = 0;
   s->num_bytes_recvd = 0;
Misbah Mubarak's avatar
Misbah Mubarak committed
979
980
   s->compute_time = 0;
   s->elapsed_time = 0;
981
982
983
984
985
986
987

   return;
}

void nw_test_event_handler(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
	*(int *)bf = (int)0;
988
989
    rc_stack_gc(lp, s->matched_reqs);
    rc_stack_gc(lp, s->processed_ops);
990
991
992
993
994
995
996
997
998
999
1000

    switch(m->msg_type)
	{
		case MPI_SEND_ARRIVED:
			update_arrival_queue(s, bf, m, lp);
		break;

		case MPI_SEND_ARRIVED_CB:
			update_message_time(s, bf, m, lp);
		break;

1001
1002
1003
1004
1005
1006
1007
        case MPI_SEND_POSTED:
        {
           if(m->op_type == CODES_WK_SEND)
               codes_issue_next_event(lp);
           else
            if(m->op_type == CODES_WK_ISEND)
            {
1008
              update_completed_queue(s, bf, m, lp, m->fwd.req_id);  
1009
1010
1011
            }
        }
        break;
1012
1013
1014
1015
1016
1017
1018
1019
		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)
{
1020
    codes_workload_get_next_rc2(wrkld_id, 0, (int)s->nw_id);
1021

1022
	if(m->op_type == CODES_WK_END)
1023
    {
1024
		return;
1025
    }
1026
	switch(m->op_type)
1027
1028
1029
1030
	{
		case CODES_WK_SEND:
		case CODES_WK_ISEND:
		{
1031
			model_net_event_rc(net_id, lp, m->rc.saved_num_bytes);
1032
			if(m->op_type == CODES_WK_ISEND)
1033
1034
				codes_issue_next_event_rc(lp);
			s->num_sends--;
1035
1036
            s->num_bytes_sent += m->rc.saved_num_bytes;
			num_bytes_sent -= m->rc.saved_num_bytes;
1037
1038
1039
1040
1041
1042
		}
		break;

		case CODES_WK_IRECV:
		case CODES_WK_RECV:
		{
1043
			codes_exec_mpi_recv_rc(s, bf, m, lp);
1044
1045
1046
			s->num_recvs--;
		}
		break;
1047
1048
1049
        
		
        case CODES_WK_DELAY:
1050
1051
		{
			s->num_delays--;
1052
            tw_rand_reverse_unif(lp->rng);
1053
1054
            if(!disable_delay)
                s->compute_time = m->rc.saved_delay;
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
		}
		break;
		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:
		{
			s->num_cols--;
		    codes_issue_next_event_rc(lp);
        }
		break;
	
1071
1072
1073
1074
1075
1076
1077
1078
		case CODES_WK_WAITSOME:
		case CODES_WK_WAITANY:
        {
           s->num_waitsome--;
           codes_issue_next_event_rc(lp); 
        }
        break;

1079
1080
1081
		case CODES_WK_WAIT:
		{
			s->num_wait--;
1082
			codes_exec_mpi_wait_rc(s, lp);
1083
1084
1085
1086
1087
		}
		break;
		case CODES_WK_WAITALL:
		{
			s->num_waitall--;
1088
            codes_exec_mpi_wait_all_rc(s, bf, m, lp);
1089
1090
1091
		}
		break;
		default:
1092
			printf("\n Invalid op type %d ", m->op_type);
1093
1094
1095
1096
1097
	}
}

static void get_next_mpi_operation(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
1098
1099
1100
		//struct codes_workload_op * mpi_op = malloc(sizeof(struct codes_workload_op));
        struct codes_workload_op mpi_op;
        codes_workload_get_next(wrkld_id, 0, (int)s->nw_id, &mpi_op);
1101
      
1102
        m->op_type = mpi_op.op_type;
1103

1104
        if(mpi_op.op_type == CODES_WK_END)
1105
1106
1107
1108
        {
            s->elapsed_time = tw_now(lp) - s->start_time;
            return;
        }
1109
		switch(mpi_op.op_type)
1110
1111
1112
1113
1114
		{
			case CODES_WK_SEND:
			case CODES_WK_ISEND:
			 {
				s->num_sends++;
1115
				codes_exec_mpi_send(s, bf, m, lp, &mpi_op);
1116
1117
1118
1119
1120
1121
1122
			 }
			break;
	
			case CODES_WK_RECV:
			case CODES_WK_IRECV:
			  {
				s->num_recvs++;
1123
				codes_exec_mpi_recv(s, bf, m, lp, &mpi_op);
1124
1125
1126
1127
1128
			  }
			break;

			case CODES_WK_DELAY:
			  {
1129
                
1130
				s->num_delays++;
1131
1132
1133
                if(disable_delay)
                    codes_issue_next_event(lp);
                else
1134
				    codes_exec_comp_delay(s, m, lp, &mpi_op);
1135
1136
1137
1138
1139
			  }
			break;

            case CODES_WK_WAITSOME:
            case CODES_WK_WAITANY:
1140
1141
1142
1143
1144
1145
            {
                s->num_waitsome++;
                codes_issue_next_event(lp);
            }
            break;

1146
			case CODES_WK_WAITALL:
1147
			  {
1148
				s->num_waitall++;
1149
			    codes_exec_mpi_wait_all(s, bf, m, lp, &mpi_op);
1150
              }
1151
1152
1153
1154
			break;
			case CODES_WK_WAIT:
			{
				s->num_wait++;
1155
                codes_exec_mpi_wait(s, lp, &mpi_op);
1156
1157
			}
			break;
1158
1159
1160
1161
1162
1163
1164
1165
			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:
1166
			{
1167
				s->num_cols++;
1168
1169
1170
1171
			    codes_issue_next_event(lp);
            }
			break;
			default:
1172
				printf("\n Invalid op type %d ", mpi_op.op_type);
1173
1174
1175
1176
1177
1178
		}
        return;
}

void nw_test_finalize(nw_state* s, tw_lp* lp)
{
1179
1180
1181
    int written = 0;
    if(!s->nw_id)
        written = sprintf(s->output_buf, "# Format <LP ID> <Terminal ID> <Total sends> <Total Recvs> <Bytes sent> <Bytes recvd> <Send time> <Comm. time> <Compute time>");
1182
1183
1184
1185
	if(s->nw_id < num_net_traces)
	{
		int count_irecv = qlist_count(&s->pending_recvs_queue);
        int count_isend = qlist_count(&s->arrival_queue);
1186
		printf("\n LP %llu unmatched irecvs %d unmatched sends %d Total sends %ld receives %ld collectives %ld delays %ld wait alls %ld waits %ld send time %lf wait %lf", 
1187
1188
			lp->gid, count_irecv, count_isend, s->num_sends, s->num_recvs, s->num_cols, s->num_delays, s->num_waitall, s->num_wait, s->send_time, s->wait_time);

1189
        written += sprintf(s->output_buf + written, "\n %llu %llu %ld %ld %ld %ld %lf %lf %lf", lp->gid, s->nw_id, s->num_sends, s->num_recvs, s->num_bytes_sent, 
1190
1191
1192
                s->num_bytes_recvd, s->send_time, s->elapsed_time - s->compute_time, s->compute_time);
        lp_io_write(lp->gid, "mpi-replay-stats", written, s->output_buf);

1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
		if(s->elapsed_time - s->compute_time > max_comm_time)
			max_comm_time = s->elapsed_time - s->compute_time;
		
		if(s->elapsed_time > max_time )
			max_time = s->elapsed_time;

		/*if(s->wait_time > max_wait_time)
			max_wait_time = s->wait_time;
        */
		if(s->send_time > max_send_time)
			max_send_time = s->send_time;

		if(s->recv_time > max_recv_time)
			max_recv_time = s->recv_time;

		avg_time += s->elapsed_time;
		avg_comm_time += (s->elapsed_time - s->compute_time);
		avg_wait_time += s->wait_time;
		avg_send_time += s->send_time;
		 avg_recv_time += s->recv_time;

		//printf("\n LP %ld Time spent in communication %llu ", lp->gid, total_time - s->compute_time);
1215
	    rc_stack_destroy(s->matched_reqs);