model-net-mpi-replay.c 78.2 KB
Newer Older
1 2 3 4 5 6 7
/*
 * Copyright (C) 2014 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
 */
#include <ross.h>
#include <inttypes.h>
8
#include <sys/stat.h>
9 10 11 12 13 14 15 16

#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
#include "codes/quickhash.h"
18
#include "codes/codes-jobmap.h"
19

20
/* turning on track lp will generate a lot of output messages */
21
#define MN_LP_NM "modelnet_dragonfly_custom"
22
#define CONTROL_MSG_SZ 64
23
#define TRACE -1
24
#define MAX_WAIT_REQS 512
25
#define CS_LP_DBG 1
26 27
#define RANK_HASH_TABLE_SZ 2000
#define NW_LP_NM "nw-lp"
28 29 30
#define lprintf(_fmt, ...) \
        do {if (CS_LP_DBG) printf(_fmt, __VA_ARGS__);} while (0)
#define MAX_STATS 65536
31
#define PAYLOAD_SZ 1024
32

33 34 35
static int msg_size_hash_compare(
            void *key, struct qhash_head *link);

36
/* NOTE: Message tracking works in sequential mode only! */
37
static int debug_cols = 0;
38
int enable_msg_tracking = 0;
39
int is_synthetic = 0;
40
tw_lpid TRACK_LP = -1;
41

42
static double total_syn_data = 0;
43
int unmatched = 0;
44 45 46 47 48
char workload_type[128];
char workload_file[8192];
char offset_file[8192];
static int wrkld_id;
static int num_net_traces = 0;
49
static int num_dumpi_traces = 0;
50
static uint64_t EAGER_THRESHOLD = 8192;
51

52
static int alloc_spec = 0;
53 54
static tw_stime self_overhead = 10.0;
static tw_stime mean_interval = 100000;
55 56 57

/* Doing LP IO*/
static char lp_io_dir[256] = {'\0'};
58
static char sampling_dir[32] = {'\0'};
59 60 61 62
static lp_io_handle io_handle;
static unsigned int lp_io_use_suffix = 0;
static int do_lp_io = 0;

63 64 65 66
/* variables for loading multiple applications */
char workloads_conf_file[8192];
char alloc_file[8192];
int num_traces_of_job[5];
67 68 69
tw_stime soft_delay_mpi = 2500;
tw_stime nic_delay = 1000;
tw_stime copy_per_byte_eager = 0.55;
70 71 72 73 74
char file_name_of_job[5][8192];

struct codes_jobmap_ctx *jobmap_ctx;
struct codes_jobmap_params_list jobmap_p;

75 76
/* Variables for Cortex Support */
/* Matthieu's additions start */
77
#ifdef ENABLE_CORTEX_PYTHON
78 79 80
static char cortex_file[512] = "\0";
static char cortex_class[512] = "\0";
static char cortex_gen[512] = "\0";
81
#endif
82 83
/* Matthieu's additions end */

84 85
typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
86
typedef int32_t dumpi_req_id;
87 88

static int net_id = 0;
89
static float noise = 2.0;
90 91 92
static int num_nw_lps = 0, num_mpi_lps = 0;

static int num_syn_clients;
93

94
FILE * workload_log = NULL;
95
FILE * msg_size_log = NULL;
96 97 98 99
FILE * workload_agg_log = NULL;
FILE * workload_meta_log = NULL;

static uint64_t sample_bytes_written = 0;
100

101 102 103
long long num_bytes_sent=0;
long long num_bytes_recvd=0;

104 105 106
long long num_syn_bytes_sent = 0;
long long num_syn_bytes_recvd = 0;

107 108 109 110 111 112
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;


/* runtime option for disabling computation time simulation */
static int disable_delay = 0;
113 114 115
static int enable_sampling = 0;
static double sampling_interval = 5000000;
static double sampling_end_time = 3000000000;
116
static int enable_debug = 0;
117

118 119 120
/* set group context */
struct codes_mctx group_ratio;

121
/* MPI_OP_GET_NEXT is for getting next MPI operation when the previous operation completes.
122
* 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.
123 124 125 126 127 128 129
* 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,
130 131 132 133 134 135
    MPI_REND_ARRIVED,
    MPI_REND_ACK_ARRIVED,
    CLI_BCKGND_FIN,
    CLI_BCKGND_ARRIVE,
    CLI_BCKGND_GEN,
    CLI_NBR_FINISH,
136 137
};

138 139 140 141
struct mpi_workload_sample
{
    /* Sampling data */
    int nw_id;
142
    int app_id;
143 144 145 146 147
    unsigned long num_sends_sample;
    unsigned long num_bytes_sample;
    unsigned long num_waits_sample;
    double sample_end_time;
};
148 149 150 151 152 153 154
/* 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;
155
    uint64_t num_bytes;
156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171
    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;
172
    int32_t req_ids[MAX_WAIT_REQS];
173
	int num_completed;
174 175
	int count;
    tw_stime start_time;
176 177 178
    struct qlist_head ql;
};

179 180 181 182 183 184 185 186 187
struct msg_size_info
{
    int64_t msg_size;
    int num_msgs;
    tw_stime agg_latency;
    tw_stime avg_latency;
    struct qhash_head * hash_link;
    struct qlist_head ql; 
};
188 189 190 191 192 193 194 195 196 197
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;
198 199
    int app_id;
    int local_rank;
200

201 202 203
    int is_finished;
    int neighbor_completed;

204
    struct rc_stack * processed_ops;
205
    struct rc_stack * processed_wait_op;
206
    struct rc_stack * matched_reqs;
207 208 209 210 211 212 213 214 215 216

    /* 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;

217

218 219
	/* time spent by the LP in executing the app trace*/
	double start_time;
220 221 222 223 224 225 226 227 228

    double col_time;

    double reduce_time;
    int num_reduce;

    double all_reduce_time;
    int num_all_reduce;

229 230 231 232 233 234 235 236 237 238 239 240 241 242 243
	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;
244

245
    tw_stime cur_interval_end;
246
    
247 248
    /* Pending wait operation */
    struct pending_waits * wait_op;
249

250 251 252 253 254 255
    /* Message size latency information */
    struct qhash_table * msg_sz_table;
    struct qlist_head msg_sz_list;

    /* quick hash for maintaining message latencies */

256 257 258
    unsigned long num_bytes_sent;
    unsigned long num_bytes_recvd;

259 260 261
    unsigned long syn_data;
    unsigned long gen_data;
    
262 263 264 265
    /* For sampling data */
    int sampling_indx;
    int max_arr_size;
    struct mpi_workload_sample * mpi_wkld_samples;
266
    char output_buf[512];
267
    char col_stats[64];
268 269 270 271
};

/* data for handling reverse computation.
* saved_matched_req holds the request ID of matched receives/sends for wait operations.
272
* 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.
273 274 275
* 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
{
276
   // forward message handler
277
   int msg_type;
278
   int op_type;
279
   model_net_event_return event_rc;
280

281 282 283
   struct
   {
       tw_lpid src_rank;
284
       int dest_rank;
285
       uint64_t num_bytes;
286 287 288 289 290
       int num_matched;
       int data_type;
       double sim_start_time;
       // for callbacks - time message was received
       double msg_send_time;
291
       int32_t req_id;
292
       int tag;
293
       int app_id;
294 295 296 297 298 299 300 301 302
       int found_match;
       short wait_completed;
   } fwd;
   struct
   {
       double saved_send_time;
       double saved_recv_time;
       double saved_wait_time;
       double saved_delay;
303
       int32_t saved_num_bytes;
304
   } rc;
305 306
};

307
static void send_ack_back(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, mpi_msgs_queue * mpi_op);
308 309

static void send_ack_back_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
310 311
/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
312
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op, int is_rend);
313 314
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
315
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
316 317
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
318
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp);
319 320
/* execute the computational delay */
static void codes_exec_comp_delay(
321
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
322 323 324 325 326 327 328 329 330 331 332 333 334 335 336
/* 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 */
337 338 339 340 341 342 343 344
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);
345 346 347 348 349 350 351 352 353 354 355 356 357 358 359
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);

360
/*static void update_message_size_rc(
361 362 363 364
        struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
365
{*/
366
/*TODO: Complete reverse handler */
367
/*    (void)ns;
368 369 370
    (void)lp;
    (void)bf;
    (void)m;
371
}*/
372 373 374 375 376 377 378 379 380 381
/* update the message size */
static void update_message_size(
        struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m,
        mpi_msgs_queue * qitem,
        int is_eager,
        int is_send)
{
382 383 384
            (void)bf;
            (void)is_eager;

385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401
            struct qhash_head * hash_link = NULL;
            tw_stime msg_init_time = qitem->req_init_time;
        
            if(!ns->msg_sz_table)
                ns->msg_sz_table = qhash_init(msg_size_hash_compare, quickhash_64bit_hash, RANK_HASH_TABLE_SZ); 
            
            hash_link = qhash_search(ns->msg_sz_table, &(qitem->num_bytes));

            if(is_send)
                msg_init_time = m->fwd.sim_start_time;
            
            /* update hash table */
            if(!hash_link)
            {
                struct msg_size_info * msg_info = malloc(sizeof(struct msg_size_info));
                msg_info->msg_size = qitem->num_bytes;
                msg_info->num_msgs = 1;
402
                msg_info->agg_latency = tw_now(lp) - msg_init_time;
403
                msg_info->avg_latency = msg_info->agg_latency;
404
                qhash_add(ns->msg_sz_table, &(msg_info->msg_size), msg_info->hash_link);
405 406 407 408 409 410 411
                qlist_add(&msg_info->ql, &ns->msg_sz_list);
                //printf("\n Msg size %d aggregate latency %f num messages %d ", m->fwd.num_bytes, msg_info->agg_latency, msg_info->num_msgs);
            }
            else
            {
                struct msg_size_info * tmp = qhash_entry(hash_link, struct msg_size_info, hash_link);
                tmp->num_msgs++;
412
                tmp->agg_latency += tw_now(lp) - msg_init_time;  
413 414 415 416 417 418 419 420 421 422
                tmp->avg_latency = (tmp->agg_latency / tmp->num_msgs);
//                printf("\n Msg size %d aggregate latency %f num messages %d ", qitem->num_bytes, tmp->agg_latency, tmp->num_msgs);
            }
}
static void notify_background_traffic_rc(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
423 424 425
    (void)ns;
    (void)bf;
    (void)m;
426 427 428 429 430
        
    int num_jobs = codes_jobmap_get_num_jobs(jobmap_ctx); 
    
    for(int i = 0; i < num_jobs - 1; i++)
        tw_rand_reverse_unif(lp->rng); 
431 432 433 434 435 436 437 438
}

static void notify_background_traffic(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
439 440 441
        (void)bf;
        (void)m;

442 443 444 445 446 447 448 449 450 451 452 453 454 455 456
        struct codes_jobmap_id jid; 
        jid = codes_jobmap_to_local_id(ns->nw_id, jobmap_ctx);
        
        int num_jobs = codes_jobmap_get_num_jobs(jobmap_ctx); 
        
        for(int other_id = 0; other_id < num_jobs; other_id++)
        {
            if(other_id == jid.job)
                continue;

            struct codes_jobmap_id other_jid;
            other_jid.job = other_id;

            int num_other_ranks = codes_jobmap_get_num_ranks(other_id, jobmap_ctx);

457
            lprintf("\n Other ranks %d ", num_other_ranks);
458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503
            tw_stime ts = (1.1 * g_tw_lookahead) + tw_rand_exponential(lp->rng, mean_interval/10000);
            tw_lpid global_dest_id;
     
            for(int k = 0; k < num_other_ranks; k++)    
            {
                other_jid.rank = k;
                int intm_dest_id = codes_jobmap_to_global_id(other_jid, jobmap_ctx); 
                global_dest_id = codes_mapping_get_lpid_from_relative(intm_dest_id, NULL, NW_LP_NM, NULL, 0);

                tw_event * e;
                struct nw_message * m_new;  
                e = tw_event_new(global_dest_id, ts, lp);
                m_new = tw_event_data(e);
                m_new->msg_type = CLI_BCKGND_FIN;
                tw_event_send(e);   
            }
        }
        return;
}
static void notify_neighbor_rc(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
       if(bf->c0)
       {
            notify_background_traffic_rc(ns, lp, bf, m);
            return;
       }
   
       if(bf->c1)
       {
          tw_rand_reverse_unif(lp->rng); 
       }
} 
static void notify_neighbor(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
    if(ns->local_rank == num_dumpi_traces - 1 
            && ns->is_finished == 1
            && ns->neighbor_completed == 1)
    {
504
//        printf("\n All workloads completed, notifying background traffic ");
505 506 507 508 509 510 511 512 513 514 515 516 517
        bf->c0 = 1;
        notify_background_traffic(ns, lp, bf, m);
        return;
    }
    
    struct codes_jobmap_id nbr_jid;
    nbr_jid.job = ns->app_id;
    tw_lpid global_dest_id;

    if(ns->is_finished == 1 && (ns->neighbor_completed == 1 || ns->local_rank == 0))
    {
        bf->c1 = 1;

518
//        printf("\n Local rank %d notifying neighbor %d ", ns->local_rank, ns->local_rank+1);
519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539
        tw_stime ts = (1.1 * g_tw_lookahead) + tw_rand_exponential(lp->rng, mean_interval/10000);
        nbr_jid.rank = ns->local_rank + 1;
        
        /* Send a notification to the neighbor about completion */
        int intm_dest_id = codes_jobmap_to_global_id(nbr_jid, jobmap_ctx); 
        global_dest_id = codes_mapping_get_lpid_from_relative(intm_dest_id, NULL, NW_LP_NM, NULL, 0);
       
        tw_event * e;
        struct nw_message * m_new;  
        e = tw_event_new(global_dest_id, ts, lp);
        m_new = tw_event_data(e); 
        m_new->msg_type = CLI_NBR_FINISH;
        tw_event_send(e);   
    }
}
void finish_bckgnd_traffic_rc(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
540 541 542 543
        (void)b;
        (void)msg;
        (void)lp;

544 545 546 547 548 549 550 551 552
        ns->is_finished = 0;
        return;
}
void finish_bckgnd_traffic(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
553 554
        (void)b;
        (void)msg;
555
        ns->is_finished = 1;
556
        lprintf("\n LP %llu completed sending data %lu completed at time %lf ", lp->gid, ns->gen_data, tw_now(lp));
557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638
        return;
}

void finish_nbr_wkld_rc(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
    ns->neighbor_completed = 0;
    
    notify_neighbor_rc(ns, lp, b, msg);
}

void finish_nbr_wkld(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
    ns->neighbor_completed = 1;

    notify_neighbor(ns, lp, b, msg);
}
static void gen_synthetic_tr_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
    if(bf->c0)
        return;

    model_net_event_rc2(lp, &m->event_rc);
    s->gen_data -= PAYLOAD_SZ;

    num_syn_bytes_sent -= PAYLOAD_SZ;
    tw_rand_reverse_unif(lp->rng);
    tw_rand_reverse_unif(lp->rng);

}

/* generate synthetic traffic */
static void gen_synthetic_tr(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
    if(s->is_finished == 1)
    {
        bf->c0 = 1;
        return;
    }

    /* Get job information */
    tw_lpid global_dest_id;

    struct codes_jobmap_id jid;
    jid = codes_jobmap_to_local_id(s->nw_id, jobmap_ctx); 

    int num_clients = codes_jobmap_get_num_ranks(jid.job, jobmap_ctx);
    int dest_svr = tw_rand_integer(lp->rng, 0, num_clients - 1);

    if(dest_svr == s->local_rank)
    {
       dest_svr = (s->local_rank + 1) % num_clients;
    }
   
    jid.rank = dest_svr;

    int intm_dest_id = codes_jobmap_to_global_id(jid, jobmap_ctx); 
    global_dest_id = codes_mapping_get_lpid_from_relative(intm_dest_id, NULL, NW_LP_NM, NULL, 0);

    nw_message remote_m;
    remote_m.fwd.sim_start_time = tw_now(lp);
    remote_m.fwd.dest_rank = dest_svr;
    remote_m.msg_type = CLI_BCKGND_ARRIVE;
    remote_m.fwd.num_bytes = PAYLOAD_SZ;
    remote_m.fwd.app_id = s->app_id;
    remote_m.fwd.src_rank = s->local_rank;

    m->event_rc = model_net_event(net_id, "synthetic-tr", global_dest_id, PAYLOAD_SZ, 0.0, 
            sizeof(nw_message), (const void*)&remote_m, 
            0, NULL, lp);
    
    s->gen_data += PAYLOAD_SZ;
    num_syn_bytes_sent += PAYLOAD_SZ; 

    /* New event after MEAN_INTERVAL */  
639
    tw_stime ts = mean_interval  + tw_rand_exponential(lp->rng, noise); 
640 641 642 643 644 645 646 647 648 649
    tw_event * e;
    nw_message * m_new;
    e = tw_event_new(lp->gid, ts, lp);
    m_new = tw_event_data(e);
    m_new->msg_type = CLI_BCKGND_GEN;
    tw_event_send(e);
}

void arrive_syn_tr_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
650 651 652
    (void)bf;
    (void)m;
    (void)lp;
653 654 655 656 657 658 659
//    printf("\n Data arrived %d total data %ld ", m->fwd.num_bytes, s->syn_data);
    int data = m->fwd.num_bytes;
    s->syn_data -= data;
    num_syn_bytes_recvd -= data;
}
void arrive_syn_tr(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
660 661 662
    (void)bf;
    (void)lp;

663 664 665 666 667
//    printf("\n Data arrived %d total data %ld ", m->fwd.num_bytes, s->syn_data);
    int data = m->fwd.num_bytes;
    s->syn_data += data;
    num_syn_bytes_recvd += data;
}
668
/* Debugging functions, may generate unused function warning */
669
static void print_waiting_reqs(int32_t * reqs, int count)
670
{
671
    lprintf("\n Waiting reqs: %d count", count);
672 673
    int i;
    for(i = 0; i < count; i++ )
674
        lprintf(" %d ", reqs[i]);
675
}
676 677 678 679 680 681 682 683 684 685 686 687
static void print_msgs_queue(struct qlist_head * head, int is_send)
{
    if(is_send)
        printf("\n Send msgs queue: ");
    else
        printf("\n Recv msgs queue: ");

    struct qlist_head * ent = NULL;
    mpi_msgs_queue * current = NULL;
    qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, mpi_msgs_queue, ql);
688
            printf(" \n Source %d Dest %d bytes %llu tag %d ", current->source_rank, current->dest_rank, current->num_bytes, current->tag);
689 690
       }
}
691 692 693 694 695 696 697 698
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);
699
            printf(" %d ", current->req_id);
700 701
       }
}
702
static int clear_completed_reqs(nw_state * s,
703
        tw_lp * lp,
704
        int32_t * reqs, int count)
705
{
706 707 708
    (void)s;
    (void)lp;

709
    int i, matched = 0;
710

711 712 713
    for( i = 0; i < count; i++)
    {
      struct qlist_head * ent = NULL;
714 715 716
      struct completed_requests * current = NULL;
      struct completed_requests * prev = NULL;

717 718
      qlist_for_each(ent, &s->completed_reqs)
       {
719 720 721 722 723
            current = qlist_entry(ent, completed_requests, ql);
            
            if(prev)
              rc_stack_push(lp, prev, free, s->matched_reqs);
            
724 725
            if(current->req_id == reqs[i])
            {
726
                ++matched;
727
                qlist_del(&current->ql);
728
                prev = current;
729
            }
730 731
            else
                prev = NULL;
732
       }
733 734 735

      if(prev)
          rc_stack_push(lp, prev, free, s->matched_reqs);
736
    }
737
    return matched;
738
}
739
static void add_completed_reqs(nw_state * s,
740 741
        tw_lp * lp,
        int count)
742
{
743
    (void)lp;
744 745 746
    int i;
    for( i = 0; i < count; i++)
    {
747 748
       struct completed_requests * req = rc_stack_pop(s->matched_reqs);
       qlist_add(&req->ql, &s->completed_reqs);
749 750
    }
}
751

752 753 754 755 756 757
/* 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);
}

758
static int notify_posted_wait(nw_state* s,
759
        tw_bf * bf, nw_message * m, tw_lp * lp,
760
        dumpi_req_id completed_req)
761
{
762 763
    (void)bf;

764 765
    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;
766

767
    m->fwd.wait_completed = 0;
768

769 770
    if(!wait_elem)
        return 0;
771

772
    int op_type = wait_elem->op_type;
773

774 775 776 777 778
    if(op_type == CODES_WK_WAIT &&
            (wait_elem->req_ids[0] == completed_req))
    {
            wait_completed = 1;
    }
779 780
    else if(op_type == CODES_WK_WAITALL
            || op_type == CODES_WK_WAITANY
781 782 783 784 785 786
            || op_type == CODES_WK_WAITSOME)
    {
        int i;
        for(i = 0; i < wait_elem->count; i++)
        {
            if(wait_elem->req_ids[i] == completed_req)
787
            {
788
                wait_elem->num_completed++;
789
                if(wait_elem->num_completed > wait_elem->count)
790
                    printf("\n Num completed %d count %d LP %llu ",
791 792 793
                            wait_elem->num_completed,
                            wait_elem->count,
                            lp->gid);
794 795
//                if(wait_elem->num_completed > wait_elem->count)
//                    tw_lp_suspend(lp, 1, 0);
796

797
                if(wait_elem->num_completed >= wait_elem->count)
798
                {
799
                    if(enable_debug)
800
                        fprintf(workload_log, "\n(%lf) APP ID %d MPI WAITALL COMPLETED AT %llu ", tw_now(lp), s->app_id, s->nw_id);
801
                    wait_completed = 1;
802
                }
803

804
                m->fwd.wait_completed = 1;
805
            }
806
        }
807
    }
808
    return wait_completed;
809
}
810

811
/* reverse handler of MPI wait operation */
812
static void codes_exec_mpi_wait_rc(nw_state* s, tw_bf * bf, tw_lp* lp)
813
{
814
    if(bf->c2)
815
     {
816 817 818
         struct pending_waits * wait_op = s->wait_op;
         free(wait_op);
         s->wait_op = NULL;
819
     }
820
   if(bf->c1)
821 822
    {
        codes_issue_next_event_rc(lp);
823
        completed_requests * qi = rc_stack_pop(s->processed_ops);
824
        qlist_add(&qi->ql, &s->completed_reqs);
825
    }
826
    return;
827
}
828

829
/* execute MPI wait operation */
830
static void codes_exec_mpi_wait(nw_state* s, tw_bf * bf, tw_lp* lp, struct codes_workload_op * mpi_op)
831
{
832 833
    /* check in the completed receives queue if the request ID has already been completed.*/
    assert(!s->wait_op);
834
    dumpi_req_id req_id = mpi_op->u.wait.req_id;
835

836
    struct completed_requests* current = NULL;
837

838 839 840 841 842 843
    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)
        {
844
            bf->c1=1;
845
            qlist_del(&current->ql);
846
            rc_stack_push(lp, current, free, s->processed_ops);
847 848 849 850
            codes_issue_next_event(lp);
            return;
        }
    }
851 852

    bf->c2 = 1;
853 854 855 856 857
    /* 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;
858 859
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
860
    s->wait_op = wait_op;
861

862
    return;
863 864
}

865
static void codes_exec_mpi_wait_all_rc(
866
        nw_state* s,
867 868
        tw_bf * bf,
        nw_message * m,
869
        tw_lp* lp)
870
{
871 872 873 874 875 876 877 878 879 880 881
  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--;
    }
  }
882 883 884 885 886 887 888 889
  if(s->wait_op)
  {
      struct pending_waits * wait_op = s->wait_op;
      free(wait_op);
      s->wait_op = NULL;
  }
  else
  {
890
      add_completed_reqs(s, lp, m->fwd.num_matched);
891 892 893
      codes_issue_next_event_rc(lp);
  }
  return;
894
}
895

896
static void codes_exec_mpi_wait_all(
897
        nw_state* s,
898 899
        tw_bf * bf,
        nw_message * m,
900
        tw_lp* lp,
901
        struct codes_workload_op * mpi_op)
902
{
903
  if(enable_debug)
904
    fprintf(workload_log, "\n MPI WAITALL POSTED AT %llu ", s->nw_id);
905

906 907 908 909 910 911 912 913
  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;
914
        s->mpi_wkld_samples[indx].app_id = s->app_id;
915 916 917 918 919 920
        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)
    {
921
        struct mpi_workload_sample * tmp = calloc((MAX_STATS + s->max_arr_size), sizeof(struct mpi_workload_sample));
922 923 924 925 926 927 928 929
        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++;
  }
930
  int count = mpi_op->u.waits.count;
931 932
  /* If the count is not less than max wait reqs then stop */
  assert(count < MAX_WAIT_REQS);
933

934
  int i = 0, num_matched = 0;
935
  m->fwd.num_matched = 0;
936

937
  if(lp->gid == TRACK_LP)
938
  {
939
      printf("\n MPI Wait all posted ");
940 941
      print_waiting_reqs(mpi_op->u.waits.req_ids, count);
      print_completed_queue(&s->completed_reqs);
942
  }
943
      /* check number of completed irecvs in the completion queue */
944 945 946 947 948 949 950
  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)
       {
951
            current = qlist_entry(ent, struct completed_requests, ql);
952 953 954 955
            if(current->req_id == req_id)
                num_matched++;
       }
  }
956

957
  m->fwd.found_match = num_matched;
958 959 960 961
  if(num_matched == count)
  {
    /* No need to post a MPI Wait all then, issue next event */
      /* Remove all completed requests from the list */
962 963 964
      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);
965 966
      s->wait_op = NULL;
      codes_issue_next_event(lp);
967 968
  }
  else
969 970 971 972 973 974 975 976 977 978 979
  {
      /* 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;
980
	  wait_op->start_time = tw_now(lp);
981
      s->wait_op = wait_op;
982
  }
983 984
  return;
}
985

986 987
/* search for a matching mpi operation and remove it from the list.
 * Record the index in the list from where the element got deleted.
988
 * Index is used for inserting the element once again in the queue for reverse computation. */
989
static int rm_matching_rcv(nw_state * ns,
990
        tw_bf * bf,
991 992
        nw_message * m,
        tw_lp * lp,
993
        mpi_msgs_queue * qitem)
994 995
{
    int matched = 0;
996
    int index = 0;
997 998
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;
999

1000 1001
    qlist_for_each(ent, &ns->pending_recvs_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
1002 1003
        if(//(qi->num_bytes == qitem->num_bytes)
                //&& 
1004
               ((qi->tag == qitem->tag) || qi->tag == -1)
1005
                && ((qi->source_rank == qitem->source_rank) || qi->source_rank == -1))
1006 1007
        {
            matched = 1;
1008
            qi->num_bytes = qitem->num_bytes;
1009 1010
            break;
        }
1011
        ++index;
1012
    }
1013

1014 1015
    if(matched)
    {
1016 1017 1018 1019 1020 1021 1022
        if(enable_msg_tracking && qitem->num_bytes < EAGER_THRESHOLD)
        {
            update_message_size(ns, lp, bf, m, qitem, 1, 1);
        }
        if(qitem->num_bytes >= EAGER_THRESHOLD)
        {
            /* Matching receive found, need to notify the sender to transmit
1023 1024
             * the data * (only works in sequential mode)*/
            bf->c10 = 1;
1025 1026
            send_ack_back(ns, bf, m, lp, qitem);
        }
1027
        m->rc.saved_recv_time = ns->recv_time;
1028
        ns->recv_time += (tw_now(lp) - m->fwd.sim_start_time);
1029

1030 1031
        if(qi->op_type == CODES_WK_IRECV)
            update_completed_queue(ns, bf, m, lp, qi->req_id);
1032 1033
        else if(qi->op_type == CODES_WK_RECV)
            codes_issue_next_event(lp);
1034

1035
        qlist_del(&qi->ql);
1036

1037
        rc_stack_push(lp, qi, free, ns->processed_ops);
1038
        return index;
1039 1040 1041 1042
    }
    return -1;
}

1043
static int rm_matching_send(nw_state * ns,
1044 1045 1046
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp, mpi_msgs_queue * qitem)
1047 1048 1049 1050 1051
{
    int matched = 0;
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;

1052
    int index = 0;
1053 1054
    qlist_for_each(ent, &ns->arrival_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
1055 1056 1057
        if(//(qi->num_bytes == qitem->num_bytes) // it is not a requirement in MPI that the send and receive sizes match
                // && 
		(qi->tag == qitem->tag || qitem->tag == -1)
1058 1059
                && ((qi->source_rank == qitem->source_rank) || qitem->source_rank == -1))
        {
1060
            qitem->num_bytes = qi->num_bytes;
1061 1062 1063
            matched = 1;
            break;
        }
1064
        ++index;
1065 1066 1067 1068
    }

    if(matched)
    {
1069 1070 1071 1072 1073 1074 1075
        if(enable_msg_tracking && (qi->num_bytes < EAGER_THRESHOLD))
            update_message_size(ns, lp, bf, m, qi, 1, 0);
        
        if(qitem->num_bytes >= EAGER_THRESHOLD)
        {
            /* Matching receive found, need to notify the sender to transmit
             * the data */
1076
            bf->c10 = 1;
1077 1078
            send_ack_back(ns, bf, m, lp, qi);
        }
1079 1080
	    rc_stack_push(lp, qi, free, ns->processed_ops);

1081
        m->rc.saved_recv_time = ns->recv_time;
1082 1083
        ns->recv_time += (tw_now(lp) - qitem->req_init_time);

1084
//        printf("\n Completed req id %d ", qitem->req_id);
1085 1086 1087
        if(qitem->op_type == CODES_WK_IRECV)
            update_completed_queue(ns, bf, m, lp, qitem->req_id);

1088
        qlist_del(&qi->ql);
1089

1090
        return index;
1091 1092 1093 1094 1095
    }
    return -1;
}
static void codes_issue_next_event_rc(tw_lp * lp)
{
1096
	    tw_rand_reverse_unif(lp->rng);
1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107
}

/* 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);
1108
   assert(ts > 0);
1109 1110 1111 1112 1113 1114 1115 1116 1117
   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(
1118
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
1119 1120 1121 1122 1123
{
	tw_event* e;
	tw_stime ts;
	nw_message* msg;

1124
    m->rc.saved_delay = s->compute_time;
1125 1126
    s->compute_time += s_to_ns(mpi_op->u.delay.seconds);
    ts = s_to_ns(mpi_op->u.delay.seconds);
1127 1128

	ts += g_tw_lookahead + 0.1 + tw_rand_exponential(lp->rng, noise);
1129
    assert(ts > 0);
1130

1131 1132 1133
	e = tw_event_new( lp->gid, ts , lp );
	msg = tw_event_data(e);
	msg->msg_type = MPI_OP_GET_NEXT;
1134 1135
	tw_event_send(e);

1136 1137 1138
}

/* reverse computation operation for MPI irecv */
1139
static void codes_exec_mpi_recv_rc(
1140 1141 1142
        nw_state* ns,
        tw_bf * bf,
        nw_message* m,
1143
        tw_lp* lp)
1144
{
1145
	ns->recv_time = m->rc.saved_recv_time;
1146 1147 1148 1149

    if(bf->c10)
        send_ack_back_rc(ns, bf, m, lp);

1150
	if(m->fwd.found_match >= 0)
1151
	  {
1152
		ns->recv_time = m->rc.saved_recv_time;
1153 1154 1155 1156
        int queue_count = qlist_count(&ns->arrival_queue);

        mpi_msgs_queue * qi = rc_stack_pop(ns->processed_ops);

1157
        if(!m->fwd.found_match)
1158 1159 1160
        {
            qlist_add(&qi->ql, &ns->arrival_queue);
        }
1161
        else if(m->fwd.found_match >= queue_count)
1162 1163 1164
        {
            qlist_add_tail(&qi->ql, &ns->arrival_queue);
        }
1165
        else if(m->fwd.found_match > 0 && m->fwd.found_match < queue_count)
1166
        {
1167 1168 1169 1170
            int index = 1;
            struct qlist_head * ent = NULL;
            qlist_for_each(ent, &ns->arrival_queue)
            {
1171
               if(index == m->fwd.found_match)
1172 1173 1174 1175
               {
                 qlist_add(&qi->ql, ent);
                 break;
               }
1176
               index++;
1177
            }
1178
        }
1179
        if(m->op_type == CODES_WK_IRECV)
1180
        {
1181
            update_completed_queue_rc(ns, bf, m, lp);
1182
        }
1183 1184
        codes_issue_next_event_rc(lp);
      }
1185
	else if(m->fwd.found_match < 0)
1186
	    {
1187
            struct qlist_head * ent = qlist_pop_back(&ns->pending_recvs_queue);
1188 1189
            mpi_msgs_queue * qi = qlist_entry(ent, mpi_msgs_queue, ql);
            free(qi);
1190

1191
            if(m->op_type == CODES_WK_IRECV)
1192
                codes_issue_next_event_rc(lp);
1193 1194 1195
	    }
}

1196
/* Execute MPI Irecv operation (non-blocking receive) */
1197
static void codes_exec_mpi_recv(
1198
        nw_state* s,
1199
        tw_bf * bf,
1200 1201
        nw_message * m,
        tw_lp* lp,
1202
        struct codes_workload_op * mpi_op)
1203 1204 1205 1206 1207
{
/* 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. */

1208
	m->rc.saved_recv_time = s->recv_time;
1209 1210
    m->rc.saved_num_bytes = mpi_op->u.recv.num_bytes;

1211 1212 1213 1214 1215 1216 1217 1218 1219
    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;

1220

1221
    //printf("\n Req id %d bytes %d source %d tag %d ", recv_op->req_id, recv_op->num_bytes, recv_op->source_rank, recv_op->tag);
1222 1223
    if(s->nw_id == (tw_lpid)TRACK_LP)
        printf("\n Receive op posted num bytes %llu source %d ", recv_op->num_bytes,
1224 1225
                recv_op->source_rank);

1226
	int found_matching_sends = rm_matching_send(s, bf, m, lp, recv_op);
1227 1228 1229 1230

	/* save the req id inserted in the completed queue for reverse computation. */
	if(found_matching_sends < 0)
	  {
1231
	   	  m->fwd.found_match = -1;
1232
          qlist_add_tail(&recv_op->ql, &s->pending_recvs_queue);
1233

1234 1235 1236
	       /* 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)
		   {
1237
			codes_issue_next_event(lp);
1238 1239
			return;
		   }
1240
      }
1241 1242
	else
	  {
1243
        m->fwd.found_match = found_matching_sends;
1244
        codes_issue_next_event(lp);
1245
      }
1246 1247
}

1248 1249 1250 1251 1252 1253 1254 1255
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;
}
1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271
static void codes_exec_mpi_send_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
        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;
           }
        }
        model_net_event_rc2(lp, &m->event_rc);
1272 1273

        if(bf->c4)
1274 1275
            codes_issue_next_event_rc(lp);
        s->num_sends--;
1276 1277 1278 1279 1280 1281

        if(bf->c3)
        {
            s->num_bytes_sent -= m->rc.saved_num_bytes;
            num_bytes_sent -= m->rc.saved_num_bytes;
        }
1282
}
1283
/* executes MPI send and isend operations */
1284
static void codes_exec_mpi_send(nw_state* s,
1285 1286
        tw_bf * bf,
        nw_message * m,
1287
        tw_lp* lp,
1288 1289
        struct codes_workload_op * mpi_op,
        int is_rend)
1290
{
1291 1292 1293 1294
    bf->c3 = 0;
    bf->c1 = 0;
    bf->c4 = 0;

1295 1296
	/* model-net event */
    int global_dest_rank = mpi_op->u.send.dest_rank;
1297

1298 1299 1300 1301 1302
    if(alloc_spec)
    {
        global_dest_rank = get_global_id_of_job_rank(mpi_op->u.send.dest_rank, s->app_id);
    }

1303 1304
    if(lp->gid == TRACK_LP)
        printf("\n Sender rank %llu global dest rank %d dest-rank %d bytes %lld Tag %d", s->nw_id, global_dest_rank, mpi_op->u.send.dest_rank, mpi_op->u.send.num_bytes, mpi_op->u.send.tag);
1305
    m->rc.saved_num_bytes = mpi_op->u.send.num_bytes;
1306
	/* model-net event */
1307
	tw_lpid dest_rank = codes_mapping_get_lpid_from_relative(global_dest_rank, NULL, "nw-lp", NULL, 0);
1308

1309
    if(is_rend == 1 || (!is_rend && mpi_op->u.send.num_bytes < EAGER_THRESHOLD))
1310
    {
1311
        bf->c3 = 1;
1312 1313 1314
        num_bytes_sent += mpi_op->u.send.num_bytes;
        s->num_bytes_sent += mpi_op->u.send.num_bytes;
    }
1315 1316 1317 1318 1319 1320 1321
    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;
1322