model-net-mpi-replay.c 40.4 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
static void print_waiting_reqs(int32_t * reqs, int count)
238
239
240
241
{
    printf("\n Waiting reqs: ");
    int i;
    for(i = 0; i < count; i++ )
242
        printf(" %ld ", reqs[i]);
243
244
245
246
247
248
249
250
251
252
253
254
}
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);
            printf(" %ld ", current->req_id);
       }
}
255
static int clear_completed_reqs(nw_state * s,
256
        tw_lp * lp,
257
        int32_t * reqs, int count)
258
{
259
    int i, matched = 0;
260
261
262
263
264
265
266
267
268
    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])
            {
269
                ++matched;
270
271
272
273
274
                qlist_del(&current->ql);
                rc_stack_push(lp, current, free, s->matched_reqs);
            }
       }
    }
275
    return matched;
276
277
}
static void add_completed_reqs(nw_state * s, 
278
279
        tw_lp * lp,
        int count)
280
281
282
283
284
285
286
287
{
    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); 
    }
}
288
289
290
291
292
293
/* 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);
}

294
295
296
static int notify_posted_wait(nw_state* s,
        tw_bf * bf, nw_message * m, tw_lp * lp, 
        dumpi_req_id completed_req)
297
{
298
299
    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;
300

301
302
    m->fwd.wait_completed = 0;
    
303
304
    if(!wait_elem)
        return 0;
305
    
306
    int op_type = wait_elem->op_type;
307

308
309
310
311
312
313
314
315
316
317
318
319
320
    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)
321
            {
322
                wait_elem->num_completed++;
323
324
325
326
327
                if(wait_elem->num_completed > wait_elem->count)
                    printf("\n Num completed %ld count %ld LP %llu ",
                            wait_elem->num_completed,
                            wait_elem->count,
                            lp->gid);
328
329
330
331
                assert(wait_elem->num_completed <= wait_elem->count);
                if(wait_elem->num_completed == wait_elem->count)
                    wait_completed = 1;
           
332
                m->fwd.wait_completed = 1;
333
            }
334
        }
335
    }
336
    return wait_completed;
337
}
338

339
/* reverse handler of MPI wait operation */
340
static void codes_exec_mpi_wait_rc(nw_state* s, tw_lp* lp)
341
{
342
    if(s->wait_op)
343
     {
344
345
346
         struct pending_waits * wait_op = s->wait_op;
         free(wait_op);
         s->wait_op = NULL;
347
348
349
350
     }
   else
    {
        codes_issue_next_event_rc(lp);
351
        completed_requests * qi = rc_stack_pop(s->processed_ops);
352
        qlist_add(&qi->ql, &s->completed_reqs);
353
    }
354
    return;
355
}
356

357
/* execute MPI wait operation */
358
static void codes_exec_mpi_wait(nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op)
359
{
360
361
    /* check in the completed receives queue if the request ID has already been completed.*/
    assert(!s->wait_op);
362
    dumpi_req_id req_id = mpi_op->u.wait.req_id;
363
    struct completed_requests* current = NULL;
364

365
366
367
368
369
370
371
    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);
372
            rc_stack_push(lp, current, free, s->processed_ops);
373
374
375
376
            codes_issue_next_event(lp);
            return;
        }
    }
377
378
379
380
381
    /* 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;
382
383
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
384
    s->wait_op = wait_op;
385

386
    return;
387
388
}

389
390
391
392
static void codes_exec_mpi_wait_all_rc(
        nw_state* s, 
        tw_bf * bf,
        nw_message * m,
393
        tw_lp* lp)
394
{
395
396
397
398
399
400
401
402
  if(s->wait_op)
  {
      struct pending_waits * wait_op = s->wait_op;
      free(wait_op);
      s->wait_op = NULL;
  }
  else
  {
403
      add_completed_reqs(s, lp, m->fwd.num_matched);
404
405
406
      codes_issue_next_event_rc(lp);
  }
  return;
407
408
}
static void codes_exec_mpi_wait_all(
409
410
411
412
413
        nw_state* s, 
        tw_bf * bf,
        nw_message * m,
        tw_lp* lp, 
        struct codes_workload_op * mpi_op)
414
415
{
  int count = mpi_op->u.waits.count;
416
417
  /* If the count is not less than max wait reqs then stop */
  assert(count < MAX_WAIT_REQS);
418

419
  int i = 0, num_matched = 0;
420
  m->fwd.num_matched = 0;
421

422
  /*if(lp->gid == TRACK)
423
  {
424
      printf("\n MPI Wait all posted ");
425
426
      print_waiting_reqs(mpi_op->u.waits.req_ids, count);
      print_completed_queue(&s->completed_reqs);
427
  }*/
428
429
430
431
432
433
434
435
436
437
438
439
440
      /* 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++;
       }
  }
441

442
  m->fwd.found_match = num_matched;
443
444
445
446
  if(num_matched == count)
  {
    /* No need to post a MPI Wait all then, issue next event */
      /* Remove all completed requests from the list */
447
448
449
      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);
450
451
      s->wait_op = NULL;
      codes_issue_next_event(lp);
452
453
  }
  else
454
455
456
457
458
459
460
461
462
463
464
  {
      /* 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;
465
	  wait_op->start_time = tw_now(lp);
466
      s->wait_op = wait_op;
467
  }
468
469
  return;
}
470
471
472
473

/* 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. */
474
475
476
477
478
static int rm_matching_rcv(nw_state * ns, 
        tw_bf * bf,
        nw_message * m, 
        tw_lp * lp, 
        mpi_msgs_queue * qitem)
479
480
{
    int matched = 0;
481
    int index = 0;
482
483
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;
484

485
486
    qlist_for_each(ent, &ns->pending_recvs_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
487
        if((qi->num_bytes == qitem->num_bytes)
488
489
                && ((qi->tag == qitem->tag) || qi->tag == -1)
                && ((qi->source_rank == qitem->source_rank) || qi->source_rank == -1))
490
491
492
493
        {
            matched = 1;
            break;
        }
494
        ++index;
495
496
497
498
    }
    
    if(matched)
    {
499
        m->rc.saved_recv_time = ns->recv_time;
500
        ns->recv_time += (tw_now(lp) - qi->req_init_time);
501
502
503
        
        if(qi->op_type == CODES_WK_IRECV)
            update_completed_queue(ns, bf, m, lp, qi->req_id);
504
        
505
        qlist_del(&qi->ql);
506
507
        
        rc_stack_push(lp, qi, free, ns->processed_ops);
508
        return index;
509
510
511
512
    }
    return -1;
}

513
514
515
516
static int rm_matching_send(nw_state * ns, 
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp, mpi_msgs_queue * qitem)
517
518
519
520
521
{
    int matched = 0;
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;

522
    int index = 0;
523
524
    qlist_for_each(ent, &ns->arrival_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
525
        if((qi->num_bytes == qitem->num_bytes) 
526
527
528
529
530
531
                && (qi->tag == qitem->tag || qitem->tag == -1)
                && ((qi->source_rank == qitem->source_rank) || qitem->source_rank == -1))
        {
            matched = 1;
            break;
        }
532
        ++index;
533
534
535
536
    }

    if(matched)
    {
537
        m->rc.saved_recv_time = ns->recv_time;
538
539
540
541
542
        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);

543
        qlist_del(&qi->ql);
544

545
        return index;
546
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
    }
    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(
572
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
573
574
575
576
577
{
	tw_event* e;
	tw_stime ts;
	nw_message* msg;

578
    m->rc.saved_delay = s->compute_time;
579
580
    s->compute_time += s_to_ns(mpi_op->u.delay.seconds);
    ts = s_to_ns(mpi_op->u.delay.seconds);
581
582
583
584
585
586
587
588
589
590
591

	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 */
592
593
594
595
static void codes_exec_mpi_recv_rc(
        nw_state* ns, 
        tw_bf * bf, 
        nw_message* m, 
596
        tw_lp* lp)
597
{
598
	num_bytes_recvd -= m->rc.saved_num_bytes;
599
600
	ns->recv_time = m->rc.saved_recv_time;
	if(m->fwd.found_match >= 0)
601
	  {
602
		ns->recv_time = m->rc.saved_recv_time;
603
        int queue_count = qlist_count(&ns->arrival_queue); 
604
        
605
        mpi_msgs_queue * qi = rc_stack_pop(ns->processed_ops);	
606
       
607
        if(!m->fwd.found_match)
608
609
610
        {
            qlist_add(&qi->ql, &ns->arrival_queue);
        }
611
        else if(m->fwd.found_match >= queue_count)
612
613
614
        {
            qlist_add_tail(&qi->ql, &ns->arrival_queue);
        }
615
        else if(m->fwd.found_match > 0 && m->fwd.found_match < queue_count) 
616
        {
617
618
619
620
            int index = 1;
            struct qlist_head * ent = NULL;
            qlist_for_each(ent, &ns->arrival_queue)
            {
621
               if(index == m->fwd.found_match)
622
623
624
625
626
627
               {
                 qlist_add(&qi->ql, ent);
                 break;
               }
               index++; 
            }
628
        }
629
        if(qi->op_type == CODES_WK_IRECV)
630
        {
631
            update_completed_queue_rc(ns, bf, m, lp);
632
        }
633
634
        codes_issue_next_event_rc(lp);
      }
635
	else if(m->fwd.found_match < 0)
636
	    {
637
638
639
640
            struct qlist_head * ent = qlist_pop_back(&ns->pending_recvs_queue); 
            mpi_msgs_queue * qi = qlist_entry(ent, mpi_msgs_queue, ql);
            free(qi);
            
641
            if(m->op_type == CODES_WK_IRECV)
642
                codes_issue_next_event_rc(lp);
643
644
645
646
	    }
}

/* Execute MPI Irecv operation (non-blocking receive) */ 
647
648
649
650
651
652
static void codes_exec_mpi_recv(
        nw_state* s, 
        tw_bf * bf,
        nw_message * m, 
        tw_lp* lp, 
        struct codes_workload_op * mpi_op)
653
654
655
656
657
{
/* 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. */

658
	m->rc.saved_recv_time = s->recv_time;
659
660
    m->rc.saved_num_bytes = mpi_op->u.recv.num_bytes;

661
662
663
664
665
666
667
668
669
670
671
	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;

672
673
674
675
    if(s->nw_id == TRACK_LP)
        printf("\n Receive op posted num bytes %d source %d ", recv_op->num_bytes,
                recv_op->source_rank);

676
	int found_matching_sends = rm_matching_send(s, bf, m, lp, recv_op);
677
678
679
680

	/* save the req id inserted in the completed queue for reverse computation. */
	if(found_matching_sends < 0)
	  {
681
	   	  m->fwd.found_match = -1;
682
          qlist_add_tail(&recv_op->ql, &s->pending_recvs_queue);
683
684
685
686
687
688
689
	
	       /* 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;
		   }
690
      }
691
692
	else
	  {
693
        m->fwd.found_match = found_matching_sends;
694
        codes_issue_next_event(lp); 
695
696
	    rc_stack_push(lp, recv_op, free, s->processed_ops);
      }
697
698
699
}

/* executes MPI send and isend operations */
700
701
702
703
704
static void codes_exec_mpi_send(nw_state* s, 
        tw_bf * bf,
        nw_message * m,
        tw_lp* lp, 
        struct codes_workload_op * mpi_op)
705
{
706
    m->rc.saved_num_bytes = mpi_op->u.send.num_bytes;
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
	/* 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;
729
    s->num_bytes_sent += mpi_op->u.send.num_bytes;
730
731
732
733

	nw_message local_m;
	nw_message remote_m;

734
735
736
    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;
737
    local_m.op_type = mpi_op->op_type; 
738
    local_m.msg_type = MPI_SEND_POSTED;
739
740
741
    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;
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759

    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));
}

760
761
static void update_completed_queue_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
762
   
763
764
765
    if(bf->c0)
    {
       struct qlist_head * ent = qlist_pop_back(&s->completed_reqs);
766
767
768
769
770
771
772

        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);
      }*/
773
774
775
776
       free(req);
    }
    else if(bf->c1)
    {
777
       struct pending_waits* wait_elem = rc_stack_pop(s->processed_ops); 
778
       s->wait_op = wait_elem;
779
780
       s->wait_time = m->rc.saved_wait_time;
       add_completed_reqs(s, lp, m->fwd.num_matched);
781
782
       codes_issue_next_event_rc(lp); 
    }
783
784
    if(m->fwd.wait_completed > 0)
           s->wait_op->num_completed--;
785
786
787
788
789
790
791
792
793
794
}

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;
795
    m->fwd.num_matched = 0; 
796

797
798
    int waiting = 0;
    waiting = notify_posted_wait(s, bf, m, lp, req_id);
799
800
801
802
803
804
805
 
    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);
806
807
808
809
810
811
       
        /*if(lp->gid == TRACK)
        {
            printf("\n Forward mode adding %ld ", req_id);
            print_completed_queue(&s->completed_reqs);
        }*/
812
813
814
815
    }
    else 
     {
            bf->c1 = 1;
816
817
            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;
818
            s->wait_time += (tw_now(lp) - s->wait_op->start_time);  
819
820
821

            struct pending_waits* wait_elem = s->wait_op;
            rc_stack_push(lp, wait_elem, free, s->processed_ops);
822
823
824
825
826
            s->wait_op = NULL;
            codes_issue_next_event(lp); 
     }
}

827
/* reverse handler for updating arrival queue function */
828
829
830
static void update_arrival_queue_rc(nw_state* s, 
        tw_bf * bf, 
        nw_message * m, tw_lp * lp)
831
{
832
	s->recv_time = m->rc.saved_recv_time;
833
834
    s->num_bytes_recvd -= m->fwd.num_bytes;

835
    codes_local_latency_reverse(lp);
836
  
837
    if(m->fwd.found_match >= 0)
838
	{
839
        mpi_msgs_queue * qi = rc_stack_pop(s->processed_ops);
840
841
        int queue_count = qlist_count(&s->pending_recvs_queue); 

842
        if(!m->fwd.found_match)
843
844
845
        {
            qlist_add(&qi->ql, &s->pending_recvs_queue);
        }
846
        else if(m->fwd.found_match >= queue_count)
847
        {
848
849
            qlist_add_tail(&qi->ql, &s->pending_recvs_queue);
        }
850
        else if(m->fwd.found_match > 0 && m->fwd.found_match < queue_count)
851
852
853
854
855
        {
            int index = 1;
            struct qlist_head * ent = NULL;
            qlist_for_each(ent, &s->pending_recvs_queue)
            {
856
               if(index == m->fwd.found_match)
857
858
859
860
861
862
               {
                 qlist_add(&qi->ql, ent);
                 break;
               }
               index++; 
            }
863
        }
864
865
        if(qi->op_type == CODES_WK_IRECV)
            update_completed_queue_rc(s, bf, m, lp);
866
    }
867
	else if(m->fwd.found_match < 0)
868
869
870
871
872
873
874
875
876
877
	{
	    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)
{
878
	m->rc.saved_recv_time = s->recv_time;
879
    s->num_bytes_recvd += m->fwd.num_bytes;
880
881
882

    // send a callback to the sender to increment times
    tw_event *e_callback =
883
        tw_event_new(rank_to_lpid(m->fwd.src_rank),
884
885
886
                codes_local_latency(lp), lp);
    nw_message *m_callback = tw_event_data(e_callback);
    m_callback->msg_type = MPI_SEND_ARRIVED_CB;
887
    m_callback->fwd.msg_send_time = tw_now(lp) - m->fwd.sim_start_time;
888
889
890
891
    tw_event_send(e_callback);

    /* Now reconstruct the queue item */
    mpi_msgs_queue * arrived_op = (mpi_msgs_queue *) malloc(sizeof(mpi_msgs_queue));
892
    arrived_op->req_init_time = m->fwd.sim_start_time;
893
    arrived_op->op_type = m->op_type;
894
895
896
897
    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;
898

899
900
901
902
    if(s->nw_id == TRACK_LP)
        printf("\n Send op arrived source rank %ld num bytes %d ", arrived_op->source_rank,
                arrived_op->num_bytes);

903
    int found_matching_recv = rm_matching_rcv(s, bf, m, lp, arrived_op);
904
905
906

    if(found_matching_recv < 0)
    {
907
        m->fwd.found_match = -1;
908
909
910
911
        qlist_add_tail(&arrived_op->ql, &s->arrival_queue);
    }
    else
    {
912
        m->fwd.found_match = found_matching_recv;
913
        free(arrived_op);
914
915
916
917
918
919
920
921
    }
}
static void update_message_time(
        nw_state * s,
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp)
{
922
923
    m->rc.saved_send_time = s->send_time;
    s->send_time += m->fwd.msg_send_time;
924
925
926
927
928
929
930
931
}

static void update_message_time_rc(
        nw_state * s,
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp)
{
932
    s->send_time = m->rc.saved_send_time;
933
934
935
936
937
938
939
940
941
942
}

/* 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));
943
   s->nw_id = codes_mapping_get_lp_relative_id(lp->gid, 0, 0);
944
945
946
947

   if(!num_net_traces) 
	num_net_traces = num_net_lps;

948
949
   assert(num_net_traces <= num_net_lps);

950
951
952
953
954
955
956
957
958
959
960
961
   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);
962
   rc_stack_create(&s->matched_reqs);
963
964

   assert(s->processed_ops != NULL);
965
   assert(s->matched_reqs != NULL);
966
967
968
969
970
971
972
973
974
975

   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);
976
977
   s->num_bytes_sent = 0;
   s->num_bytes_recvd = 0;
Misbah Mubarak's avatar
Misbah Mubarak committed
978
979
   s->compute_time = 0;
   s->elapsed_time = 0;
980
981
982
983
984
985
986

   return;
}

void nw_test_event_handler(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
	*(int *)bf = (int)0;
987
988
    rc_stack_gc(lp, s->matched_reqs);
    rc_stack_gc(lp, s->processed_ops);
989
990
991
992
993
994
995
996
997
998
999

    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;

1000
1001
1002
1003
1004
1005
1006
        case MPI_SEND_POSTED:
        {
           if(m->op_type == CODES_WK_SEND)
               codes_issue_next_event(lp);
           else
            if(m->op_type == CODES_WK_ISEND)
            {
1007
              update_completed_queue(s, bf, m, lp, m->fwd.req_id);  
1008
1009
1010
            }
        }
        break;
1011
1012
1013
1014
1015
1016
1017
1018
		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)
{
1019
    codes_workload_get_next_rc2(wrkld_id, 0, (int)s->nw_id);
1020

1021
	if(m->op_type == CODES_WK_END)
1022
    {
1023
		return;
1024
    }
1025
	switch(m->op_type)
1026
1027
1028
1029
	{
		case CODES_WK_SEND:
		case CODES_WK_ISEND:
		{
1030
			model_net_event_rc(net_id, lp, m->rc.saved_num_bytes);
1031
			if(m->op_type == CODES_WK_ISEND)
1032
1033
				codes_issue_next_event_rc(lp);
			s->num_sends--;
1034
1035
            s->num_bytes_sent += m->rc.saved_num_bytes;
			num_bytes_sent -= m->rc.saved_num_bytes;
1036
1037
1038
1039
1040
1041
		}
		break;

		case CODES_WK_IRECV:
		case CODES_WK_RECV:
		{
1042
			codes_exec_mpi_recv_rc(s, bf, m, lp);
1043
1044
1045
			s->num_recvs--;
		}
		break;
1046
1047
1048
        
		
        case CODES_WK_DELAY:
1049
1050
		{
			s->num_delays--;
1051
            tw_rand_reverse_unif(lp->rng);
1052
1053
            if(!disable_delay)
                s->compute_time = m->rc.saved_delay;
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
		}
		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;
	
1070
1071
1072
1073
1074
1075
1076
1077
		case CODES_WK_WAITSOME:
		case CODES_WK_WAITANY:
        {
           s->num_waitsome--;
           codes_issue_next_event_rc(lp); 
        }
        break;

1078
1079
1080
		case CODES_WK_WAIT:
		{
			s->num_wait--;
1081
			codes_exec_mpi_wait_rc(s, lp);
1082
1083
1084
1085
1086
		}
		break;
		case CODES_WK_WAITALL:
		{
			s->num_waitall--;
1087
            codes_exec_mpi_wait_all_rc(s, bf, m, lp);
1088
1089
1090
		}
		break;
		default:
1091
			printf("\n Invalid op type %d ", m->op_type);
1092
1093
1094
1095
1096
	}
}

static void get_next_mpi_operation(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
1097
1098
1099
		//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);
1100
      
1101
        m->op_type = mpi_op.op_type;
1102

1103
        if(mpi_op.op_type == CODES_WK_END)
1104
1105
1106
1107
        {
            s->elapsed_time = tw_now(lp) - s->start_time;
            return;
        }
1108
		switch(mpi_op.op_type)
1109
1110
1111
1112
1113
		{
			case CODES_WK_SEND:
			case CODES_WK_ISEND:
			 {
				s->num_sends++;
1114
				codes_exec_mpi_send(s, bf, m, lp, &mpi_op);
1115
1116
1117
1118
1119
1120
1121
			 }
			break;
	
			case CODES_WK_RECV:
			case CODES_WK_IRECV:
			  {
				s->num_recvs++;
1122
				codes_exec_mpi_recv(s, bf, m, lp, &mpi_op);
1123
1124
1125
1126
1127
			  }
			break;

			case CODES_WK_DELAY:
			  {
1128
                
1129
				s->num_delays++;
1130
1131
1132
                if(disable_delay)
                    codes_issue_next_event(lp);
                else
1133
				    codes_exec_comp_delay(s, m, lp, &mpi_op);
1134
1135
1136
1137
1138
			  }
			break;

            case CODES_WK_WAITSOME:
            case CODES_WK_WAITANY:
1139
1140
1141
1142
1143
1144
            {
                s->num_waitsome++;
                codes_issue_next_event(lp);
            }
            break;

1145
			case CODES_WK_WAITALL:
1146
			  {
1147
				s->num_waitall++;
1148
			    codes_exec_mpi_wait_all(s, bf, m, lp, &mpi_op);
1149
              }
1150
1151
1152
1153
			break;
			case CODES_WK_WAIT:
			{
				s->num_wait++;
1154
                codes_exec_mpi_wait(s, lp, &mpi_op);
1155
1156
			}
			break;
1157
1158
1159
1160
1161
1162
1163
1164
			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:
1165
			{
1166
				s->num_cols++;
1167
1168
1169
1170
			    codes_issue_next_event(lp);
            }
			break;
			default:
1171
				printf("\n Invalid op type %d ", mpi_op.op_type);
1172
1173
1174
1175
1176
1177
		}
        return;
}

void nw_test_finalize(nw_state* s, tw_lp* lp)
{
Misbah Mubarak's avatar