model-net-mpi-replay.c 48.5 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/*
 * 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"
16
#include "codes/codes-jobmap.h"
17

18 19
/* turning on track lp will generate a lot of output messages */
#define TRACK_LP -1
20
#define TRACE -1
21
#define MAX_WAIT_REQS 512
22 23 24 25 26
#define CS_LP_DBG 0
#define lprintf(_fmt, ...) \
        do {if (CS_LP_DBG) printf(_fmt, __VA_ARGS__);} while (0)
#define MAX_STATS 65536
//#define WORKLOAD_LOG 1
27 28 29 30 31 32

char workload_type[128];
char workload_file[8192];
char offset_file[8192];
static int wrkld_id;
static int num_net_traces = 0;
33
static int alloc_spec = 0;
34 35 36 37 38 39 40

/* 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;

41 42 43 44 45 46 47 48 49 50 51
/* variables for loading multiple applications */
/* Xu's additions start */
char workloads_conf_file[8192];
char alloc_file[8192];
int num_traces_of_job[5];
char file_name_of_job[5][8192];

struct codes_jobmap_ctx *jobmap_ctx;
struct codes_jobmap_params_list jobmap_p;
/* Xu's additions end */

52 53
typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
54
typedef int32_t dumpi_req_id;
55 56 57 58 59

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

60 61 62 63 64
FILE * workload_log = NULL;
FILE * workload_agg_log = NULL;
FILE * workload_meta_log = NULL;

static uint64_t sample_bytes_written = 0;
65

66 67 68 69 70 71 72 73 74 75 76 77
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;
78 79 80
static int enable_sampling = 0;
static double sampling_interval = 5000000;
static double sampling_end_time = 3000000000;
81 82 83 84 85 86 87 88 89 90 91 92

/* 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,
};

93 94 95 96 97 98 99 100 101
struct mpi_workload_sample
{
    /* Sampling data */
    int nw_id;
    unsigned long num_sends_sample;
    unsigned long num_bytes_sample;
    unsigned long num_waits_sample;
    double sample_end_time;
};
102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125
/* 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;
126
    int32_t req_ids[MAX_WAIT_REQS];
127
	int num_completed;
128 129
	int count;
    tw_stime start_time;
130 131 132 133 134 135 136 137 138 139 140 141 142
    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;
143 144
    int app_id;
    int local_rank;
145 146

    struct rc_stack * processed_ops;
147
    struct rc_stack * matched_reqs;
148 149 150 151 152 153 154 155 156 157

    /* 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;

158

159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175
	/* 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;
176

177 178
    tw_stime cur_interval_end;

179 180
    /* Pending wait operation */
    struct pending_waits * wait_op;
181 182 183 184

    unsigned long num_bytes_sent;
    unsigned long num_bytes_recvd;

185 186 187 188
    /* For sampling data */
    int sampling_indx;
    int max_arr_size;
    struct mpi_workload_sample * mpi_wkld_samples;
189
    char output_buf[512];
190 191 192 193 194 195 196 197
};

/* 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
{
198
   // forward message handler 
199
   int msg_type;
200 201
   int op_type;
   
202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223
   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;
224
       struct codes_workload_op * saved_op;
225
   } rc;
226 227 228 229
};

/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
230
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
231 232
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
233
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
234 235
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
236
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp);
237 238
/* execute the computational delay */
static void codes_exec_comp_delay(
239
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
240 241 242 243 244 245 246 247 248 249 250 251 252 253 254
/* 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 */
255 256 257 258 259 260 261 262
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);
263 264 265 266 267 268 269 270 271 272 273 274 275 276 277
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);

278
/* Debugging functions, may generate unused function warning */
279
static void print_waiting_reqs(int32_t * reqs, int count)
280 281 282 283
{
    printf("\n Waiting reqs: ");
    int i;
    for(i = 0; i < count; i++ )
284
        printf(" %d ", reqs[i]);
285 286 287 288 289 290 291 292 293
}
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);
294
            printf(" %d ", current->req_id);
295 296
       }
}
297
static int clear_completed_reqs(nw_state * s,
298
        tw_lp * lp,
299
        int32_t * reqs, int count)
300
{
301
    int i, matched = 0;
302 303 304 305 306 307 308 309 310
    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])
            {
311
                ++matched;
312 313 314 315 316
                qlist_del(&current->ql);
                rc_stack_push(lp, current, free, s->matched_reqs);
            }
       }
    }
317
    return matched;
318 319
}
static void add_completed_reqs(nw_state * s, 
320 321
        tw_lp * lp,
        int count)
322 323 324 325 326 327 328 329
{
    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); 
    }
}
330

331 332 333 334 335 336
/* 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);
}

337 338 339
static int notify_posted_wait(nw_state* s,
        tw_bf * bf, nw_message * m, tw_lp * lp, 
        dumpi_req_id completed_req)
340
{
341 342
    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;
343

344 345
    m->fwd.wait_completed = 0;
    
346 347
    if(!wait_elem)
        return 0;
348
    
349
    int op_type = wait_elem->op_type;
350

351 352 353 354 355 356 357 358 359 360 361 362 363
    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)
364
            {
365
                wait_elem->num_completed++;
366
                if(wait_elem->num_completed > wait_elem->count)
367
                    printf("\n Num completed %d count %d LP %llu ",
368 369 370
                            wait_elem->num_completed,
                            wait_elem->count,
                            lp->gid);
371 372
                assert(wait_elem->num_completed <= wait_elem->count);
                if(wait_elem->num_completed == wait_elem->count)
373 374 375 376
                {
                 #if WORKLOAD_LOG == 1
                    fprintf(workload_log, "\n(%lf) MPI WAITALL COMPLETED AT %ld ", tw_now(lp), s->nw_id);
                #endif
377
                    wait_completed = 1;
378
                }
379
           
380
                m->fwd.wait_completed = 1;
381
            }
382
        }
383
    }
384
    return wait_completed;
385
}
386

387
/* reverse handler of MPI wait operation */
388
static void codes_exec_mpi_wait_rc(nw_state* s, tw_lp* lp)
389
{
390
    if(s->wait_op)
391
     {
392 393 394
         struct pending_waits * wait_op = s->wait_op;
         free(wait_op);
         s->wait_op = NULL;
395 396 397 398
     }
   else
    {
        codes_issue_next_event_rc(lp);
399
        completed_requests * qi = rc_stack_pop(s->processed_ops);
400
        qlist_add(&qi->ql, &s->completed_reqs);
401
    }
402
    return;
403
}
404

405
/* execute MPI wait operation */
406
static void codes_exec_mpi_wait(nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op)
407
{
408 409
    /* check in the completed receives queue if the request ID has already been completed.*/
    assert(!s->wait_op);
410
    dumpi_req_id req_id = mpi_op->u.wait.req_id;
411
    struct completed_requests* current = NULL;
412

413 414 415 416 417 418 419
    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);
420
            rc_stack_push(lp, current, free, s->processed_ops);
421 422 423 424
            codes_issue_next_event(lp);
            return;
        }
    }
425 426 427 428 429
    /* 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;
430 431
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
432
    s->wait_op = wait_op;
433

434
    return;
435 436
}

437 438 439 440
static void codes_exec_mpi_wait_all_rc(
        nw_state* s, 
        tw_bf * bf,
        nw_message * m,
441
        tw_lp* lp)
442
{
443 444 445 446 447 448 449 450 451 452 453
  if(bf->c1)
  {
    int sampling_indx = s->sampling_indx;
    s->mpi_wkld_samples[sampling_indx].num_waits_sample--;

    if(bf->c2)
    {
        s->cur_interval_end -= sampling_interval;
        s->sampling_indx--;
    }
  }
454 455 456 457 458 459 460 461
  if(s->wait_op)
  {
      struct pending_waits * wait_op = s->wait_op;
      free(wait_op);
      s->wait_op = NULL;
  }
  else
  {
462
      add_completed_reqs(s, lp, m->fwd.num_matched);
463 464 465
      codes_issue_next_event_rc(lp);
  }
  return;
466 467
}
static void codes_exec_mpi_wait_all(
468 469 470 471 472
        nw_state* s, 
        tw_bf * bf,
        nw_message * m,
        tw_lp* lp, 
        struct codes_workload_op * mpi_op)
473
{
474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499
#if WORKLOAD_LOG == 1
  fprintf(workload_log, "\n MPI WAITALL POSTED AT %ld ", s->nw_id);
#endif
  if(enable_sampling)
  {
    bf->c1 = 1;
    if(tw_now(lp) >= s->cur_interval_end)
    {
        bf->c2 = 1;
        int indx = s->sampling_indx;
        s->mpi_wkld_samples[indx].nw_id = s->nw_id;
        s->mpi_wkld_samples[indx].sample_end_time = s->cur_interval_end;
        s->cur_interval_end += sampling_interval;
        s->sampling_indx++;
    }
    if(s->sampling_indx >= MAX_STATS)
    {
        struct mpi_workload_sample * tmp = malloc((MAX_STATS + s->max_arr_size) * sizeof(struct mpi_workload_sample));
        memcpy(tmp, s->mpi_wkld_samples, s->sampling_indx);
        free(s->mpi_wkld_samples);
        s->mpi_wkld_samples = tmp;
        s->max_arr_size += MAX_STATS;
    }
    int indx = s->sampling_indx;
    s->mpi_wkld_samples[indx].num_waits_sample++;
  }
500
  int count = mpi_op->u.waits.count;
501 502
  /* If the count is not less than max wait reqs then stop */
  assert(count < MAX_WAIT_REQS);
503

504
  int i = 0, num_matched = 0;
505
  m->fwd.num_matched = 0;
506

507
  /*if(lp->gid == TRACK)
508
  {
509
      printf("\n MPI Wait all posted ");
510 511
      print_waiting_reqs(mpi_op->u.waits.req_ids, count);
      print_completed_queue(&s->completed_reqs);
512
  }*/
513 514 515 516 517 518 519 520 521 522 523 524 525
      /* 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++;
       }
  }
526

527
  m->fwd.found_match = num_matched;
528 529 530 531
  if(num_matched == count)
  {
    /* No need to post a MPI Wait all then, issue next event */
      /* Remove all completed requests from the list */
532 533 534
      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);
535 536
      s->wait_op = NULL;
      codes_issue_next_event(lp);
537 538
  }
  else
539 540 541 542 543 544 545 546 547 548 549
  {
      /* 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;
550
	  wait_op->start_time = tw_now(lp);
551
      s->wait_op = wait_op;
552
  }
553 554
  return;
}
555 556 557 558

/* 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. */
559 560 561 562 563
static int rm_matching_rcv(nw_state * ns, 
        tw_bf * bf,
        nw_message * m, 
        tw_lp * lp, 
        mpi_msgs_queue * qitem)
564 565
{
    int matched = 0;
566
    int index = 0;
567 568
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;
569

570 571
    qlist_for_each(ent, &ns->pending_recvs_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
572
        if((qi->num_bytes == qitem->num_bytes)
573 574
                && ((qi->tag == qitem->tag) || qi->tag == -1)
                && ((qi->source_rank == qitem->source_rank) || qi->source_rank == -1))
575 576 577 578
        {
            matched = 1;
            break;
        }
579
        ++index;
580 581 582 583
    }
    
    if(matched)
    {
584
        m->rc.saved_recv_time = ns->recv_time;
585
        ns->recv_time += (tw_now(lp) - qi->req_init_time);
586 587 588
        
        if(qi->op_type == CODES_WK_IRECV)
            update_completed_queue(ns, bf, m, lp, qi->req_id);
589
        
590
        qlist_del(&qi->ql);
591 592
        
        rc_stack_push(lp, qi, free, ns->processed_ops);
593
        return index;
594 595 596 597
    }
    return -1;
}

598 599 600 601
static int rm_matching_send(nw_state * ns, 
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp, mpi_msgs_queue * qitem)
602 603 604 605 606
{
    int matched = 0;
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;

607
    int index = 0;
608 609
    qlist_for_each(ent, &ns->arrival_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
610
        if((qi->num_bytes == qitem->num_bytes) 
611 612 613 614 615 616
                && (qi->tag == qitem->tag || qitem->tag == -1)
                && ((qi->source_rank == qitem->source_rank) || qitem->source_rank == -1))
        {
            matched = 1;
            break;
        }
617
        ++index;
618 619 620 621
    }

    if(matched)
    {
622
        m->rc.saved_recv_time = ns->recv_time;
623 624 625 626 627
        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);

628
        qlist_del(&qi->ql);
629

630
        return index;
631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656
    }
    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(
657
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
658 659 660 661 662
{
	tw_event* e;
	tw_stime ts;
	nw_message* msg;

663
    m->rc.saved_delay = s->compute_time;
664 665
    s->compute_time += s_to_ns(mpi_op->u.delay.seconds);
    ts = s_to_ns(mpi_op->u.delay.seconds);
666 667 668 669 670 671 672 673 674 675 676

	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 */
677 678 679 680
static void codes_exec_mpi_recv_rc(
        nw_state* ns, 
        tw_bf * bf, 
        nw_message* m, 
681
        tw_lp* lp)
682
{
683
	num_bytes_recvd -= m->rc.saved_num_bytes;
684 685
	ns->recv_time = m->rc.saved_recv_time;
	if(m->fwd.found_match >= 0)
686
	  {
687
		ns->recv_time = m->rc.saved_recv_time;
688
        int queue_count = qlist_count(&ns->arrival_queue); 
689
        
690
        mpi_msgs_queue * qi = rc_stack_pop(ns->processed_ops);	
691
       
692
        if(!m->fwd.found_match)
693 694 695
        {
            qlist_add(&qi->ql, &ns->arrival_queue);
        }
696
        else if(m->fwd.found_match >= queue_count)
697 698 699
        {
            qlist_add_tail(&qi->ql, &ns->arrival_queue);
        }
700
        else if(m->fwd.found_match > 0 && m->fwd.found_match < queue_count) 
701
        {
702 703 704 705
            int index = 1;
            struct qlist_head * ent = NULL;
            qlist_for_each(ent, &ns->arrival_queue)
            {
706
               if(index == m->fwd.found_match)
707 708 709 710 711 712
               {
                 qlist_add(&qi->ql, ent);
                 break;
               }
               index++; 
            }
713
        }
714
        if(qi->op_type == CODES_WK_IRECV)
715
        {
716
            update_completed_queue_rc(ns, bf, m, lp);
717
        }
718 719
        codes_issue_next_event_rc(lp);
      }
720
	else if(m->fwd.found_match < 0)
721
	    {
722 723 724 725
            struct qlist_head * ent = qlist_pop_back(&ns->pending_recvs_queue); 
            mpi_msgs_queue * qi = qlist_entry(ent, mpi_msgs_queue, ql);
            free(qi);
            
726
            if(m->op_type == CODES_WK_IRECV)
727
                codes_issue_next_event_rc(lp);
728 729 730 731
	    }
}

/* Execute MPI Irecv operation (non-blocking receive) */ 
732 733 734 735 736 737
static void codes_exec_mpi_recv(
        nw_state* s, 
        tw_bf * bf,
        nw_message * m, 
        tw_lp* lp, 
        struct codes_workload_op * mpi_op)
738 739 740 741 742
{
/* 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. */

743
	m->rc.saved_recv_time = s->recv_time;
744 745
    m->rc.saved_num_bytes = mpi_op->u.recv.num_bytes;

746 747 748 749 750 751 752 753 754 755 756
	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;

757 758 759 760
    if(s->nw_id == TRACK_LP)
        printf("\n Receive op posted num bytes %d source %d ", recv_op->num_bytes,
                recv_op->source_rank);

761
	int found_matching_sends = rm_matching_send(s, bf, m, lp, recv_op);
762 763 764 765

	/* save the req id inserted in the completed queue for reverse computation. */
	if(found_matching_sends < 0)
	  {
766
	   	  m->fwd.found_match = -1;
767
          qlist_add_tail(&recv_op->ql, &s->pending_recvs_queue);
768 769 770 771 772 773 774
	
	       /* 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;
		   }
775
      }
776 777
	else
	  {
778
        m->fwd.found_match = found_matching_sends;
779
        codes_issue_next_event(lp); 
780 781
	    rc_stack_push(lp, recv_op, free, s->processed_ops);
      }
782 783
}

784 785 786 787 788 789 790 791
int get_global_id_of_job_rank(tw_lpid job_rank, int app_id)
{
    struct codes_jobmap_id lid;
    lid.job = app_id;
    lid.rank = job_rank;
    int global_rank = codes_jobmap_to_global_id(lid, jobmap_ctx);
    return global_rank;
}
792
/* executes MPI send and isend operations */
793 794 795 796 797
static void codes_exec_mpi_send(nw_state* s, 
        tw_bf * bf,
        nw_message * m,
        tw_lp* lp, 
        struct codes_workload_op * mpi_op)
798
{
799 800 801 802 803 804 805 806
	/* model-net event */
    int global_dest_rank = mpi_op->u.send.dest_rank;
    
    if(alloc_spec)
    {
        global_dest_rank = get_global_id_of_job_rank(mpi_op->u.send.dest_rank, s->app_id);
    }

807
    m->rc.saved_num_bytes = mpi_op->u.send.num_bytes;
808 809 810 811 812 813 814 815
	/* 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;
816 817
		num_routers = codes_mapping_get_lp_count(lp_group_name, 1,
                  "modelnet_dragonfly_router", NULL, 1);
818
	 	lps_per_rep = (2 * num_nw_lps) + num_routers;	
819 820
		factor = global_dest_rank / num_nw_lps;
		dest_rank = (lps_per_rep * factor) + (global_dest_rank % num_nw_lps);	
821 822 823 824 825
	}
	else
	{
		/* other cases like torus/simplenet/loggp etc. */
		codes_mapping_get_lp_id(lp_group_name, lp_type_name, NULL, 1,  
826
	    	  global_dest_rank, mapping_offset, &dest_rank);
827 828 829
	}

	num_bytes_sent += mpi_op->u.send.num_bytes;
830
    s->num_bytes_sent += mpi_op->u.send.num_bytes;
831

832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854
    if(enable_sampling)
    {
        if(tw_now(lp) >= s->cur_interval_end)
        {
            bf->c1 = 1;
            int indx = s->sampling_indx;
            s->mpi_wkld_samples[indx].nw_id = s->nw_id;
            s->mpi_wkld_samples[indx].sample_end_time = s->cur_interval_end;
            s->sampling_indx++;
            s->cur_interval_end += sampling_interval; 
        }
        if(s->sampling_indx >= MAX_STATS)
        {
            struct mpi_workload_sample * tmp = malloc((MAX_STATS + s->max_arr_size) * sizeof(struct mpi_workload_sample));
            memcpy(tmp, s->mpi_wkld_samples, s->sampling_indx);
            free(s->mpi_wkld_samples);
            s->mpi_wkld_samples = tmp;
            s->max_arr_size += MAX_STATS;
        }
        int indx = s->sampling_indx;
        s->mpi_wkld_samples[indx].num_sends_sample++;
        s->mpi_wkld_samples[indx].num_bytes_sample += mpi_op->u.send.num_bytes;
    }
855 856 857
	nw_message local_m;
	nw_message remote_m;

858 859 860
    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;
861
    local_m.op_type = mpi_op->op_type; 
862
    local_m.msg_type = MPI_SEND_POSTED;
863 864 865
    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;
866 867 868 869 870 871 872

    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);

873 874 875 876 877 878 879 880 881
#if WORKLOAD_LOG == 1
    if(mpi_op->op_type == CODES_WK_ISEND)
        fprintf(workload_log, "\n (%lf) MPI ISEND SOURCE %ld DEST %ld BYTES %ld ", 
                tw_now(lp), s->nw_id, mpi_op->u.send.dest_rank, mpi_op->u.send.num_bytes);
    else
        fprintf(workload_log, "\n (%lf) MPI SEND SOURCE %ld DEST %ld BYTES %ld ", 
                tw_now(lp), s->nw_id, mpi_op->u.send.dest_rank, mpi_op->u.send.num_bytes);
#endif

882 883 884 885 886 887 888 889 890 891 892
	/* 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));
}

893 894
static void update_completed_queue_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
895
   
896 897 898
    if(bf->c0)
    {
       struct qlist_head * ent = qlist_pop_back(&s->completed_reqs);
899 900 901 902 903 904 905

        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);
      }*/
906 907 908 909
       free(req);
    }
    else if(bf->c1)
    {
910
       struct pending_waits* wait_elem = rc_stack_pop(s->processed_ops); 
911
       s->wait_op = wait_elem;
912 913
       s->wait_time = m->rc.saved_wait_time;
       add_completed_reqs(s, lp, m->fwd.num_matched);
914 915
       codes_issue_next_event_rc(lp); 
    }
916 917
    if(m->fwd.wait_completed > 0)
           s->wait_op->num_completed--;
918 919 920 921 922 923 924 925 926 927
}

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;
928
    m->fwd.num_matched = 0; 
929

930 931
    int waiting = 0;
    waiting = notify_posted_wait(s, bf, m, lp, req_id);
932 933 934 935 936 937 938
 
    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);
939 940 941 942 943 944
       
        /*if(lp->gid == TRACK)
        {
            printf("\n Forward mode adding %ld ", req_id);
            print_completed_queue(&s->completed_reqs);
        }*/
945 946 947 948
    }
    else 
     {
            bf->c1 = 1;
949 950
            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;
951
            s->wait_time += (tw_now(lp) - s->wait_op->start_time);  
952 953 954

            struct pending_waits* wait_elem = s->wait_op;
            rc_stack_push(lp, wait_elem, free, s->processed_ops);
955 956 957 958 959
            s->wait_op = NULL;
            codes_issue_next_event(lp); 
     }
}

960
/* reverse handler for updating arrival queue function */
961 962 963
static void update_arrival_queue_rc(nw_state* s, 
        tw_bf * bf, 
        nw_message * m, tw_lp * lp)
964
{
965
	s->recv_time = m->rc.saved_recv_time;
966 967
    s->num_bytes_recvd -= m->fwd.num_bytes;

968
    codes_local_latency_reverse(lp);
969
  
970
    if(m->fwd.found_match >= 0)
971
	{
972
        mpi_msgs_queue * qi = rc_stack_pop(s->processed_ops);
973 974
        int queue_count = qlist_count(&s->pending_recvs_queue); 

975
        if(!m->fwd.found_match)
976 977 978
        {
            qlist_add(&qi->ql, &s->pending_recvs_queue);
        }
979
        else if(m->fwd.found_match >= queue_count)
980
        {
981 982
            qlist_add_tail(&qi->ql, &s->pending_recvs_queue);
        }
983
        else if(m->fwd.found_match > 0 && m->fwd.found_match < queue_count)
984 985 986 987 988
        {
            int index = 1;
            struct qlist_head * ent = NULL;
            qlist_for_each(ent, &s->pending_recvs_queue)
            {
989
               if(index == m->fwd.found_match)
990 991 992 993 994 995
               {
                 qlist_add(&qi->ql, ent);
                 break;
               }
               index++; 
            }
996
        }
997 998
        if(qi->op_type == CODES_WK_IRECV)
            update_completed_queue_rc(s, bf, m, lp);
999
    }
1000
	else if(m->fwd.found_match < 0)
1001 1002 1003 1004 1005 1006 1007 1008 1009 1010
	{
	    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)
{
1011
	m->rc.saved_recv_time = s->recv_time;
1012
    s->num_bytes_recvd += m->fwd.num_bytes;
1013 1014 1015

    // send a callback to the sender to increment times
    tw_event *e_callback =
1016
        tw_event_new(rank_to_lpid(m->fwd.src_rank),
1017 1018 1019
                codes_local_latency(lp), lp);
    nw_message *m_callback = tw_event_data(e_callback);
    m_callback->msg_type = MPI_SEND_ARRIVED_CB;
1020
    m_callback->fwd.msg_send_time = tw_now(lp) - m->fwd.sim_start_time;
1021 1022 1023 1024
    tw_event_send(e_callback);

    /* Now reconstruct the queue item */
    mpi_msgs_queue * arrived_op = (mpi_msgs_queue *) malloc(sizeof(mpi_msgs_queue));
1025
    arrived_op->req_init_time = m->fwd.sim_start_time;
1026
    arrived_op->op_type = m->op_type;
1027 1028 1029 1030
    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;
1031

1032
    if(s->nw_id == TRACK_LP)
1033
        printf("\n Send op arrived source rank %d num bytes %d ", arrived_op->source_rank,
1034 1035
                arrived_op->num_bytes);

1036
    int found_matching_recv = rm_matching_rcv(s, bf, m, lp, arrived_op);
1037 1038 1039

    if(found_matching_recv < 0)
    {
1040
        m->fwd.found_match = -1;
1041 1042 1043 1044
        qlist_add_tail(&arrived_op->ql, &s->arrival_queue);
    }
    else
    {
1045
        m->fwd.found_match = found_matching_recv;
1046
        free(arrived_op);
1047 1048 1049 1050 1051 1052 1053 1054
    }
}
static void update_message_time(
        nw_state * s,
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp)
{
1055 1056
    m->rc.saved_send_time = s->send_time;
    s->send_time += m->fwd.msg_send_time;
1057 1058 1059 1060 1061 1062 1063 1064
}

static void update_message_time_rc(
        nw_state * s,
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp)
{
1065
    s->send_time = m->rc.saved_send_time;
1066 1067 1068 1069 1070 1071 1072 1073
}

/* 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;
1074
 
1075
   memset(s, 0, sizeof(*s));
1076
   s->nw_id = codes_mapping_get_lp_relative_id(lp->gid, 0, 0);
1077 1078
   s->mpi_wkld_samples = malloc(MAX_STATS * sizeof(struct mpi_workload_sample)); 
   s->sampling_indx = 0;
1079 1080 1081

   if(!num_net_traces) 
	num_net_traces = num_net_lps;
1082 1083 1084
   /* In this case, the LP will not generate any workload related events*/
   if(s->nw_id >= num_net_traces)
	    return;
1085
   assert(num_net_traces <= num_net_lps);
1086 1087
   
   struct codes_jobmap_id lid; 
1088

1089 1090 1091
   if(alloc_spec)
   {
        lid = codes_jobmap_to_local_id(s->nw_id, jobmap_ctx); 
1092

1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110
        if(lid.job == -1)
        {
            printf("network LP nw id %d not generating events, lp gid is %ld \n", (int)s->nw_id, lp->gid); 
            s->app_id = -1;
            s->local_rank = -1;
            return;
        }
   }
   else
   {
       /* Only one job running */
       lid.job = 0;
       lid.rank = s->nw_id;
       s->app_id = 0;
   }
   if (strcmp(workload_type, "dumpi") == 0){
       strcpy(params_d.file_name, file_name_of_job[lid.job]);
       params_d.num_net_traces = num_traces_of_job[lid.job];
1111
       params = (char*)&params_d;
1112 1113 1114 1115
       s->app_id = lid.job;
       s->local_rank = lid.rank; 
       //printf("lp global id: %llu, file name: %s, num traces: %d, app id: %d, local id: %d\n", 
       //        s->nw_id, params_d.file_name, params_d.num_net_traces, s->app_id, s->local_rank);
1116
   }
1117 1118 1119

   wrkld_id = codes_workload_load("dumpi-trace-workload", params, s->app_id, s->local_rank);

1120 1121 1122

   /* Initialize the RC stack */
   rc_stack_create(&s->processed_ops);
1123
   rc_stack_create(&s->matched_reqs);
1124 1125

   assert(s->processed_ops != NULL);
1126
   assert(s->matched_reqs != NULL);
1127 1128 1129 1130 1131

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

1132
   /* clock starts ticking when the first event is processed */
1133 1134
   s->start_time = tw_now(lp);
   codes_issue_next_event(lp);
1135 1136
   s->num_bytes_sent = 0;
   s->num_bytes_recvd = 0;
1137 1138
   s->compute_time = 0;
   s->elapsed_time = 0;
1139

1140 1141 1142 1143 1144 1145 1146 1147 1148 1149
   if(enable_sampling && sampling_interval > 0)
   {
       s->max_arr_size = MAX_STATS;
       s->cur_interval_end = sampling_interval;
       if(!g_tw_mynode && !s->nw_id)
       {
           fprintf(workload_meta_log, "\n mpi_proc_id num_waits "
                   " num_sends num_bytes_sent sample_end_time");
       }
   }
1150 1151 1152 1153 1154 1155
   return;
}

void nw_test_event_handler(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
	*(int *)bf = (int)0;
1156 1157
    rc_stack_gc(lp, s->matched_reqs);
    rc_stack_gc(lp, s->processed_ops);
1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168

    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;

1169 1170 1171 1172 1173 1174 1175
        case MPI_SEND_POSTED:
        {
           if(m->op_type == CODES_WK_SEND)
               codes_issue_next_event(lp);
           else
            if(m->op_type == CODES_WK_ISEND)
            {
1176
              update_completed_queue(s, bf, m, lp, m->fwd.req_id);  
1177 1178 1179
            }
        }
        break;
1180 1181 1182 1183 1184 1185 1186 1187
		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)
{
1188
    codes_workload_get_next_rc2(wrkld_id, 0, (int)s->nw_id);
1189

1190
	if(m->op_type == CODES_WK_END)
1191
    {
1192
		return;
1193
    }
1194
	switch(m->op_type)
1195 1196 1197 1198
	{
		case CODES_WK_SEND:
		case CODES_WK_ISEND:
		{
1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211
            if(enable_sampling)
            {
               int indx = s->sampling_indx;
               
               s->mpi_wkld_samples[indx].num_sends_sample--;
               s->mpi_wkld_samples[indx].num_bytes_sample -= m->rc.saved_num_bytes;
               
               if(bf->c1)
               {
                   s->sampling_indx--;
                   s->cur_interval_end -= sampling_interval;
               }
            }
1212
			model_net_event_rc(net_id, lp, m->rc.saved_num_bytes);
1213
			if(m->op_type == CODES_WK_ISEND)
1214 1215
				codes_issue_next_event_rc(lp);
			s->num_sends--;
1216 1217
            s->num_bytes_sent += m->rc.saved_num_bytes;
			num_bytes_sent -= m->rc.saved_num_bytes;
1218 1219 1220 1221 1222 1223
		}
		break;

		case CODES_WK_IRECV:
		case CODES_WK_RECV:
		{
1224
			codes_exec_mpi_recv_rc(s, bf, m, lp);
1225 1226 1227
			s->num_recvs--;
		}
		break;
1228 1229 1230
        
		
        case CODES_WK_DELAY:
1231 1232
		{
			s->num_delays--;
1233
            tw_rand_reverse_unif(lp->rng);
1234 1235
            if(!disable_delay)
                s->compute_time = m->rc.saved_delay;
1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251
		}
		break;
		case CODES_WK_BCAST:
		case CODES_WK_ALLGA