model-net-mpi-replay.c 88.6 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

32 33 34
static int msg_size_hash_compare(
            void *key, struct qhash_head *link);

35
/* NOTE: Message tracking works in sequential mode only! */
36
static int debug_cols = 0;
37 38 39 40 41
/* Turning on this option slows down optimistic mode substantially. Only turn
 * on if you get issues with wait-all completion with traces. */
static int preserve_wait_ordering = 0;
static int enable_msg_tracking = 0;
static int is_synthetic = 0;
42
tw_lpid TRACK_LP = -1;
43
int nprocs = 0;
44
static double total_syn_data = 0;
45
static int unmatched = 0;
46
char workload_type[128];
47
char workload_name[128];
48 49 50 51
char workload_file[8192];
char offset_file[8192];
static int wrkld_id;
static int num_net_traces = 0;
52
static int num_dumpi_traces = 0;
53
static int64_t EAGER_THRESHOLD = 8192;
54

55
static int alloc_spec = 0;
56 57
static tw_stime self_overhead = 10.0;
static tw_stime mean_interval = 100000;
58
static int payload_sz = 1024;
59 60

/* Doing LP IO*/
61
static char * params = NULL;
62
static char lp_io_dir[256] = {'\0'};
63
static char sampling_dir[32] = {'\0'};
64 65 66 67
static lp_io_handle io_handle;
static unsigned int lp_io_use_suffix = 0;
static int do_lp_io = 0;

68 69 70 71
/* variables for loading multiple applications */
char workloads_conf_file[8192];
char alloc_file[8192];
int num_traces_of_job[5];
72 73 74
tw_stime soft_delay_mpi = 2500;
tw_stime nic_delay = 1000;
tw_stime copy_per_byte_eager = 0.55;
75 76 77 78 79
char file_name_of_job[5][8192];

struct codes_jobmap_ctx *jobmap_ctx;
struct codes_jobmap_params_list jobmap_p;

80 81
/* Variables for Cortex Support */
/* Matthieu's additions start */
82
#ifdef ENABLE_CORTEX_PYTHON
83 84 85
static char cortex_file[512] = "\0";
static char cortex_class[512] = "\0";
static char cortex_gen[512] = "\0";
86
#endif
87 88
/* Matthieu's additions end */

89 90
typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
91
typedef unsigned int dumpi_req_id;
92 93

static int net_id = 0;
94
static float noise = 0.1;
95 96 97
static int num_nw_lps = 0, num_mpi_lps = 0;

static int num_syn_clients;
98

99
FILE * workload_log = NULL;
100
FILE * msg_size_log = NULL;
101 102 103 104
FILE * workload_agg_log = NULL;
FILE * workload_meta_log = NULL;

static uint64_t sample_bytes_written = 0;
105

106 107 108
long long num_bytes_sent=0;
long long num_bytes_recvd=0;

109 110 111
long long num_syn_bytes_sent = 0;
long long num_syn_bytes_recvd = 0;

112 113 114 115 116 117
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;
118 119 120
static int enable_sampling = 0;
static double sampling_interval = 5000000;
static double sampling_end_time = 3000000000;
121
static int enable_debug = 0;
122

123 124 125
/* set group context */
struct codes_mctx group_ratio;

126
/* MPI_OP_GET_NEXT is for getting next MPI operation when the previous operation completes.
127
* 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.
128 129 130 131 132 133 134
* 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,
135 136 137 138 139 140
    MPI_REND_ARRIVED,
    MPI_REND_ACK_ARRIVED,
    CLI_BCKGND_FIN,
    CLI_BCKGND_ARRIVE,
    CLI_BCKGND_GEN,
    CLI_NBR_FINISH,
141 142
};

143 144 145 146 147 148 149 150
/* type of synthetic traffic */
enum TRAFFIC
{
    UNIFORM = 1, /* sends message to a randomly selected node */
    NEAREST_NEIGHBOR = 2, /* sends message to the next node (potentially connected to the same router) */
    ALLTOALL = 3, /* sends message to all other nodes */
    STENCIL = 4  /* sends message to 4 nearby neighbors */
};
151 152 153 154
struct mpi_workload_sample
{
    /* Sampling data */
    int nw_id;
155
    int app_id;
156 157 158 159 160
    unsigned long num_sends_sample;
    unsigned long num_bytes_sample;
    unsigned long num_waits_sample;
    double sample_end_time;
};
161 162 163 164 165 166 167
/* 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;
168
    int64_t num_bytes;
169
    int64_t seq_id;
170 171 172 173 174 175 176 177
    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
{
178
	unsigned int req_id;
179
    struct qlist_head ql;
180
    int index;
181 182 183 184 185 186
};

/* for wait operations, store the pending operation and number of completed waits so far. */
struct pending_waits
{
    int op_type;
187
    unsigned int req_ids[MAX_WAIT_REQS];
188
	int num_completed;
189 190
	int count;
    tw_stime start_time;
191 192 193
    struct qlist_head ql;
};

194 195 196 197 198 199 200 201 202
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; 
};
203 204 205 206 207 208 209 210 211 212
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;
213 214
    int app_id;
    int local_rank;
215

216
    int synthetic_pattern;
217 218 219
    int is_finished;
    int neighbor_completed;

220
    struct rc_stack * processed_ops;
221
    struct rc_stack * processed_wait_op;
222
    struct rc_stack * matched_reqs;
223
//    struct rc_stack * indices;
224 225 226 227 228 229 230 231 232 233

    /* 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;

234

235 236
	/* time spent by the LP in executing the app trace*/
	double start_time;
237 238 239 240 241 242 243 244 245

    double col_time;

    double reduce_time;
    int num_reduce;

    double all_reduce_time;
    int num_all_reduce;

246 247 248 249 250 251 252 253 254 255 256 257 258 259 260
	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;
261

262
    tw_stime cur_interval_end;
263
    
264 265
    /* Pending wait operation */
    struct pending_waits * wait_op;
266

267 268 269 270 271 272
    /* Message size latency information */
    struct qhash_table * msg_sz_table;
    struct qlist_head msg_sz_list;

    /* quick hash for maintaining message latencies */

273 274 275
    unsigned long num_bytes_sent;
    unsigned long num_bytes_recvd;

276 277 278
    unsigned long syn_data;
    unsigned long gen_data;
    
279 280 281 282
    /* For sampling data */
    int sampling_indx;
    int max_arr_size;
    struct mpi_workload_sample * mpi_wkld_samples;
283
    char output_buf[512];
284
    char col_stats[64];
285 286 287 288
};

/* data for handling reverse computation.
* saved_matched_req holds the request ID of matched receives/sends for wait operations.
289
* 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.
290 291 292
* 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
{
293
   // forward message handler
294
   int msg_type;
295
   int op_type;
296
   model_net_event_return event_rc;
297

298 299 300
   struct
   {
       tw_lpid src_rank;
301
       int dest_rank;
302
       int64_t num_bytes;
303 304 305 306 307
       int num_matched;
       int data_type;
       double sim_start_time;
       // for callbacks - time message was received
       double msg_send_time;
308
       unsigned int req_id;
309
       int matched_req;
310
       int tag;
311
       int app_id;
312 313
       int found_match;
       short wait_completed;
314
       short rend_send;
315 316 317 318 319 320 321
   } fwd;
   struct
   {
       double saved_send_time;
       double saved_recv_time;
       double saved_wait_time;
       double saved_delay;
322
       int64_t saved_num_bytes;
323
       int saved_syn_length;
324
       struct codes_workload_op * mpi_op;
325
   } rc;
326 327
};

328
static void send_ack_back(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, mpi_msgs_queue * mpi_op, int matched_req);
329 330

static void send_ack_back_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
331 332
/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
333
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op, int is_rend);
334 335
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
336
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
337 338
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
339
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp);
340 341
/* execute the computational delay */
static void codes_exec_comp_delay(
342
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
343 344 345 346 347 348 349 350 351 352 353 354 355 356 357
/* 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 */
358 359 360 361 362 363 364 365
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);
366 367 368 369 370 371 372 373 374 375 376 377 378 379 380
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);

381
/*static void update_message_size_rc(
382 383 384 385
        struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
386
{*/
387
/*TODO: Complete reverse handler */
388
/*    (void)ns;
389 390 391
    (void)lp;
    (void)bf;
    (void)m;
392
}*/
393 394 395 396 397 398 399 400 401 402
/* 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)
{
403 404 405
            (void)bf;
            (void)is_eager;

406 407 408 409 410 411 412 413 414 415 416 417 418 419
            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)
            {
420
                struct msg_size_info * msg_info = (struct msg_size_info*)calloc(1, sizeof(struct msg_size_info));
421 422
                msg_info->msg_size = qitem->num_bytes;
                msg_info->num_msgs = 1;
423
                msg_info->agg_latency = tw_now(lp) - msg_init_time;
424
                msg_info->avg_latency = msg_info->agg_latency;
425
                qhash_add(ns->msg_sz_table, &(msg_info->msg_size), msg_info->hash_link);
426 427 428 429 430 431 432
                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++;
433
                tmp->agg_latency += tw_now(lp) - msg_init_time;  
434 435 436 437 438 439 440 441 442 443
                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)
{
444 445 446
    (void)ns;
    (void)bf;
    (void)m;
447 448 449 450 451
        
    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); 
452 453 454 455 456 457 458 459
}

static void notify_background_traffic(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
460 461 462
        (void)bf;
        (void)m;

463 464 465 466 467 468 469 470 471 472 473 474 475 476 477
        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);

478
            lprintf("\n Other ranks %d ", num_other_ranks);
479 480 481 482 483 484 485 486 487 488 489 490
            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);
491
                m_new = (struct nw_message*)tw_event_data(e);
492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524
                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)
    {
525
//        printf("\n All workloads completed, notifying background traffic ");
526 527 528 529 530 531 532 533 534 535 536 537 538
        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;

539
//        printf("\n Local rank %d notifying neighbor %d ", ns->local_rank, ns->local_rank+1);
540 541 542 543 544 545 546 547 548 549
        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);
550
        m_new = (struct nw_message*)tw_event_data(e); 
551 552 553 554 555 556 557 558 559 560
        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)
{
561 562 563 564
        (void)b;
        (void)msg;
        (void)lp;

565 566 567 568 569 570 571 572 573
        ns->is_finished = 0;
        return;
}
void finish_bckgnd_traffic(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
574 575
        (void)b;
        (void)msg;
576
        ns->is_finished = 1;
577
        lprintf("\n LP %llu completed sending data %lu completed at time %lf ", LLU(lp->gid), ns->gen_data, tw_now(lp));
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
        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;

607 608 609 610 611 612 613
    int i;
    for (i=0; i < m->rc.saved_syn_length; i++){
        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);
614 615 616 617 618 619 620 621 622 623 624 625 626 627

}

/* 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;
628 629
    int intm_dest_id;
    nw_message remote_m;
630 631 632 633 634 635

    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);

636 637 638 639
    /* Find destination */
    int* dest_svr = NULL; 
    int i, length=0;
    switch(s->synthetic_pattern)
640
    {
641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725
        case UNIFORM:
        {
            length = 1;
            dest_svr = (int*) calloc(1, sizeof(int));
            dest_svr[0] = tw_rand_integer(lp->rng, 0, num_clients - 1);
            if(dest_svr[0] == s->local_rank)
                dest_svr[0] = (s->local_rank + 1) % num_clients;
        }
        break;
        case NEAREST_NEIGHBOR:
        {
            length = 1;
            dest_svr = (int*) calloc(1, sizeof(int));
            dest_svr[0] = (s->local_rank + 1) % num_clients;
        }
        break;
        case ALLTOALL:
        {
            dest_svr = (int*) calloc(num_clients-1, sizeof(int));
            int index = 0;
            for (i=0;i<num_clients;i++)
            {
                if(i!=s->local_rank) 
                {
                    dest_svr[index] = i;
                    index++;
                    length++;
                }
            }
        }
        break;
        case STENCIL:  //2D 4-point stencil
        {
            /* I think this code snippet is coming from the LLNL stencil patterns. */
            int digits, x=1, y=1, row, col, temp=num_clients;
            length = 4;
            dest_svr = (int*) calloc(4, sizeof(int));
            for (digits = 0; temp > 0; temp >>= 1)
                digits++;
            digits = digits/2;
            for (i = 0; i < digits; i++)
                x = x * 2;
            y = num_clients / x;
            //printf("\nStencil Syn: x=%d, y=%d", x, y);
            row = s->local_rank / y;
            col = s->local_rank % y;

            dest_svr[0] = row * y + ((col-1+y)%y);   /* left neighbor */
            dest_svr[1] = row * y + ((col+1+y)%y);   /* right neighbor */
            dest_svr[2] = ((row-1+x)%x) * y + col;   /* bottom neighbor */
            dest_svr[3] = ((row+1+x)%x) * y + col;   /* up neighbor */
        }
        break;
        default:
            tw_error(TW_LOC, "Undefined traffic pattern");
    }   
    /* Record length for reverse handler*/
    m->rc.saved_syn_length = length;

    if(length > 0)
    {
        // m->event_array_rc = (model_net_event_return) malloc(length * sizeof(model_net_event_return));
        //printf("\nRANK %d Dests %d", s->local_rank, length);
        for (i = 0; i < length; i++)
        {
            /* Generate synthetic traffic */
            jid.rank = dest_svr[i];
            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);

            remote_m.fwd.sim_start_time = tw_now(lp);
            remote_m.fwd.dest_rank = dest_svr[i];
            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;

            // printf("\nAPP %d SRC %d Dest %d (twid %llu)", jid.job, s->local_rank, dest_svr[i], global_dest_id);
            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; 
        }
726 727
    }
    /* New event after MEAN_INTERVAL */  
728
    tw_stime ts = mean_interval  + tw_rand_exponential(lp->rng, noise); 
729 730 731
    tw_event * e;
    nw_message * m_new;
    e = tw_event_new(lp->gid, ts, lp);
732
    m_new = (struct nw_message*)tw_event_data(e);
733 734
    m_new->msg_type = CLI_BCKGND_GEN;
    tw_event_send(e);
735 736

    free(dest_svr);
737 738 739 740
}

void arrive_syn_tr_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
741 742 743
    (void)bf;
    (void)m;
    (void)lp;
744 745 746 747 748 749 750
//    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)
{
751 752 753
    (void)bf;
    (void)lp;

754
//    printf("\n Data arrived %lld rank %llu total data %ld ", m->fwd.num_bytes, s->nw_id, s->syn_data);
755 756 757 758
    int data = m->fwd.num_bytes;
    s->syn_data += data;
    num_syn_bytes_recvd += data;
}
759
/* Debugging functions, may generate unused function warning */
760
/*static void print_waiting_reqs(uint32_t * reqs, int count)
761
{
762
    lprintf("\n Waiting reqs: %d count", count);
763 764
    int i;
    for(i = 0; i < count; i++ )
765
        lprintf(" %d ", reqs[i]);
766
}*/
767 768 769 770 771 772 773 774 775 776 777 778
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);
779
            printf(" \n Source %d Dest %d bytes %"PRId64" tag %d ", current->source_rank, current->dest_rank, current->num_bytes, current->tag);
780 781
       }
}
782
static void print_completed_queue(tw_lp * lp, struct qlist_head * head)
783
{
784
//    printf("\n Completed queue: ");
785 786
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
787
      tw_output(lp, "\n");
788 789 790
      qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, completed_requests, ql);
791
            tw_output(lp, " %llu ", current->req_id);
792
       }
793
}
794
static int clear_completed_reqs(nw_state * s,
795
        tw_lp * lp,
796
        unsigned int * reqs, int count)
797
{
798 799 800
    (void)s;
    (void)lp;

801
    int i, matched = 0;
802

803 804 805
    for( i = 0; i < count; i++)
    {
      struct qlist_head * ent = NULL;
806 807 808
      struct completed_requests * current = NULL;
      struct completed_requests * prev = NULL;

809
      int index = 0;
810 811
      qlist_for_each(ent, &s->completed_reqs)
       {
812 813
           if(prev)
           {
814
              rc_stack_push(lp, prev, free, s->matched_reqs);
815 816
              prev = NULL;
           }
817
            
818 819
           current = qlist_entry(ent, completed_requests, ql);
           current->index = index; 
820 821
            if(current->req_id == reqs[i])
            {
822
                ++matched;
823
                qlist_del(&current->ql);
824
                prev = current;
825
            }
826
            ++index;
827
       }
828 829

      if(prev)
830 831 832 833
      {
         rc_stack_push(lp, prev, free, s->matched_reqs);
         prev = NULL;
      }
834
    }
835
    return matched;
836
}
837
static void add_completed_reqs(nw_state * s,
838 839
        tw_lp * lp,
        int count)
840
{
841
    (void)lp;
842
    for(int i = 0; i < count; i++)
843
    {
844 845
       struct completed_requests * req = (struct completed_requests*)rc_stack_pop(s->matched_reqs);
       // turn on only if wait-all unmatched error arises in optimistic mode.
846
       qlist_add(&req->ql, &s->completed_reqs);
847
    }//end for
848
}
849

850 851 852 853 854 855
/* 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);
}

856
static int notify_posted_wait(nw_state* s,
857
        tw_bf * bf, nw_message * m, tw_lp * lp,
858
        unsigned int completed_req)
859
{
860 861
    (void)bf;

862 863
    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;
864

865
    m->fwd.wait_completed = 0;
866

867 868
    if(!wait_elem)
        return 0;
869

870
    int op_type = wait_elem->op_type;
871

872 873 874
    if(op_type == CODES_WK_WAIT &&
            (wait_elem->req_ids[0] == completed_req))
    {
875
            m->fwd.wait_completed = 1;
876 877
            wait_completed = 1;
    }
878 879
    else if(op_type == CODES_WK_WAITALL
            || op_type == CODES_WK_WAITANY
880 881 882 883 884 885
            || op_type == CODES_WK_WAITSOME)
    {
        int i;
        for(i = 0; i < wait_elem->count; i++)
        {
            if(wait_elem->req_ids[i] == completed_req)
886
            {
887
                wait_elem->num_completed++;
888
                if(wait_elem->num_completed > wait_elem->count)
889
                    printf("\n Num completed %d count %d LP %llu ",
890 891
                            wait_elem->num_completed,
                            wait_elem->count,
892
                            LLU(lp->gid));
893 894
//                if(wait_elem->num_completed > wait_elem->count)
//                    tw_lp_suspend(lp, 1, 0);
895

896
                if(wait_elem->num_completed >= wait_elem->count)
897
                {
898
                    if(enable_debug)
899
                        fprintf(workload_log, "\n(%lf) APP ID %d MPI WAITALL COMPLETED AT %llu ", tw_now(lp), s->app_id, LLU(s->nw_id));
900
                    wait_completed = 1;
901
                }
902

903
                m->fwd.wait_completed = 1;
904
            }
905
        }
906
    }
907
    return wait_completed;
908
}
909

910
/* reverse handler of MPI wait operation */
911
static void codes_exec_mpi_wait_rc(nw_state* s, tw_bf * bf, tw_lp* lp, nw_message * m)
912
{
913
   if(bf->c1)
914
    {
915
        completed_requests * qi = (completed_requests*)rc_stack_pop(s->processed_ops);
916 917
        if(m->fwd.found_match == 0)
        {
918
            qlist_add(&qi->ql, &s->completed_reqs);
919
        }
920 921 922 923 924 925 926 927 928 929 930 931 932
        else
        {
           int index = 1;
           struct qlist_head * ent = NULL;
           qlist_for_each(ent, &s->completed_reqs)
           {
                if(index == m->fwd.found_match)
                {
                    qlist_add(&qi->ql, ent);
                    break;
                }
                index++;
           }
933
        }
934
        //get_next_mpi_operation_rc(s, bf, lp, m);
935
        codes_issue_next_event_rc(lp);
936
        return;
937
    }
938 939 940
         struct pending_waits * wait_op = s->wait_op;
         free(wait_op);
         s->wait_op = NULL;
941
}
942

943
/* execute MPI wait operation */
944
static void codes_exec_mpi_wait(nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
945
{
946
    /* check in the completed receives queue if the request ID has already been completed.*/
947 948
                
//    printf("\n Wait posted rank id %d ", s->nw_id);
949
    assert(!s->wait_op);
950
    unsigned int req_id = mpi_op->u.wait.req_id;
951

952
    struct completed_requests* current = NULL;
953

954
    struct qlist_head * ent = NULL;
955
    int index = 0;
956 957 958 959 960
    qlist_for_each(ent, &s->completed_reqs)
    {
        current = qlist_entry(ent, completed_requests, ql);
        if(current->req_id == req_id)
        {
961
            bf->c1=1;
962
            qlist_del(&current->ql);
963
            rc_stack_push(lp, current, free, s->processed_ops);
964
            codes_issue_next_event(lp);
965
            m->fwd.found_match = index;
966
            if(s->nw_id == (tw_lpid)TRACK_LP)
967 968 969
            {
                tw_output(lp, "\n wait matched at post %d ", req_id);
                print_completed_queue(lp, &s->completed_reqs);
970
            }
971
            //get_next_mpi_operation(s, bf, m, lp);
972 973
            return;
        }
974
        ++index;
975
    }
976

977 978 979 980 981
    /*if(s->nw_id == (tw_lpid)TRACK_LP)
    {
        tw_output(lp, "\n wait posted %llu ", req_id);
        print_completed_queue(lp, &s->completed_reqs);
    }*/
982
    /* If not, add the wait operation in the pending 'waits' list. */
983
    struct pending_waits* wait_op = (struct pending_waits*)calloc(1, sizeof(struct pending_waits));
984 985 986
    wait_op->op_type = mpi_op->op_type;
    wait_op->req_ids[0] = req_id;
    wait_op->count = 1;
987 988
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
989
    s->wait_op = wait_op;
990

991
    return;
992 993
}

994
static void codes_exec_mpi_wait_all_rc(
995
        nw_state* s,
996 997
        tw_bf * bf,
        nw_message * m,
998
        tw_lp* lp)
999
{
1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010
  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--;
    }
  }
1011 1012 1013 1014 1015 1016 1017 1018
  if(s->wait_op)
  {
      struct pending_waits * wait_op = s->wait_op;
      free(wait_op);
      s->wait_op = NULL;
  }
  else
  {
1019
      add_completed_reqs(s, lp, m->fwd.num_matched);
1020
      codes_issue_next_event_rc(lp);
1021
      //get_next_mpi_operation_rc(s, bf, lp, m);
1022 1023
  }
  return;
1024
}
1025

1026
static void codes_exec_mpi_wait_all(
1027
        nw_state* s,
1028 1029
        tw_bf * bf,
        nw_message * m,
1030
        tw_lp* lp,
1031
        struct codes_workload_op * mpi_op)
1032
{
1033
  if(enable_debug)
1034
    fprintf(workload_log, "\n MPI WAITALL POSTED AT %llu ", LLU(s->nw_id));
1035

1036 1037 1038 1039 1040 1041 1042 1043
  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;
1044
        s->mpi_wkld_samples[indx].app_id = s->app_id;
1045 1046 1047 1048 1049 1050
        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)
    {
1051
        struct mpi_workload_sample * tmp = (struct mpi_workload_sample*)calloc((MAX_STATS + s->max_arr_size), sizeof(struct mpi_workload_sample));
1052 1053 1054 1055 1056 1057 1058 1059
        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++;
  }
1060
  int count = mpi_op->u.waits.count;
1061 1062
  /* If the count is not less than max wait reqs then stop */
  assert(count < MAX_WAIT_REQS);
1063

1064
  int i = 0, num_matched = 0;
1065
  m->fwd.num_matched = 0;
1066

1067
  /*if(lp->gid == TRACK_LP)
1068
  {
1069
      printf("\n MPI Wait all posted ");
1070
      print_waiting_reqs(mpi_op->u.waits.req_ids, count);
1071 1072
      print_completed_queue(lp, &s->completed_reqs);
  }*/
1073
      /* check number of completed irecvs in the completion queue */
1074 1075
  for(i = 0; i < count; i++)
  {
1076
      unsigned int req_id = mpi_op->u.waits.req_ids[i];
1077 1078 1079 1080
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
      qlist_for_each(ent, &s->completed_reqs)
       {
1081
            current = qlist_entry(ent, struct completed_requests, ql);
1082 1083 1084 1085
            if(current->req_id == req_id)
                num_matched++;
       }
  }
1086

1087
  m->fwd.found_match = num_matched;
1088 1089 1090 1091
  if(num_matched == count)
  {
    /* No need to post a MPI Wait all then, issue next event */
      /* Remove all completed requests from the list */
1092 1093 1094
      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);
1095 1096
      s->wait_op = NULL;
      codes_issue_next_event(lp);
1097
      //get_next_mpi_operation(s, bf, lp, m);
1098 1099
  }
  else
1100 1101
  {
      /* If not, add the wait operation in the pending 'waits' list. */
1102
	  struct pending_waits* wait_op = (struct pending_waits*)calloc(1, sizeof(struct pending_waits));
1103 1104 1105 1106 1107 1108 1109 1110
	  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;
1111
	  wait_op->start_time = tw_now(lp);
1112
      s->wait_op = wait_op;
1113
  }
1114 1115
  return;
}
1116

1117 1118
/* search for a matching mpi operation and remove it from the list.
 * Record the index in the list from where the element got deleted.
1119
 * Index is used for inserting the element once again in the queue for reverse computation. */
1120
static int rm_matching_rcv(nw_state * ns,
1121
        tw_bf * bf,
1122 1123
        nw_message * m,
        tw_lp * lp,
1124
        mpi_msgs_queue * qitem)
1125 1126
{
    int matched = 0;
1127
    int index = 0;
1128
    int is_rend = 0;
1129 1130
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;
1131

1132 1133
    qlist_for_each(ent, &ns->pending_recvs_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
1134 1135
        if(//(qi->num_bytes == qitem->num_bytes)
                //&& 
1136
               ((qi->tag == qitem->tag) || qi->tag == -1)
1137
                && ((qi->source_rank == qitem->source_rank) || qi->source_rank == -1))
1138 1139
        {
            matched = 1;
1140
            qi->num_bytes = qitem->num_bytes;
1141 1142
            break;
        }
1143
        ++index;
1144
    }
1145

1146 1147
    if(matched)
    {
1148 1149 1150 1151 1152 1153 1154
        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
1155 1156
             * the data * (only works in sequential mode)*/
            bf->c10 = 1;
1157 1158
            is_rend = 1;
            send_ack_back(ns, bf, m, lp, qitem, qi->req_id);
1159
        }
1160 1161
        else
        {
1162
            bf->c12 = 1;
1163 1164 1165 1166
            m->rc.saved_recv_time = ns->recv_time;
            ns->recv_time += (tw_now(lp) - m->fwd.sim_start_time);
        }
        if(qi->op_type == CODES_WK_IRECV && !is_rend)
1167
        {
1168
            bf->c9 = 1;
1169 1170 1171 1172
            /*if(ns->nw_id == (tw_lpid)TRACK_LP)
            {
                printf("\n Completed irecv req id %d ", qi->req_id);
            }*/
1173
            update_completed_queue(ns, bf, m, lp, qi->req_id);
1174
        }
1175 1176 1177
        else if(qi->op_type == CODES_WK_RECV && !is_rend)
        {
            bf->c8 = 1;
1178
            codes_issue_next_event(lp);
1179
            //get_next_mpi_operation(s, bf, lp, m);
1180
        }
1181

1182
        qlist_del(&qi->ql);
1183

1184
        rc_stack_push(lp, qi, free, ns->processed_ops);
1185
        return index;
1186 1187 1188 1189
    }
    return -1;
}

1190
static int rm_matching_send(nw_state * ns,
1191 1192 1193
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp, mpi_msgs_queue * qitem)
1194 1195 1196 1197 1198
{
    int matched = 0;
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;

1199
    int index = 0;
1200 1201
    qlist_for_each(ent, &ns->arrival_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
1202 1203 1204
        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)
1205 1206
                && ((qi->source_rank == qitem->source_rank) || qitem->source_rank == -1))
        {
1207
            qitem->num_bytes = qi->num_bytes;
1208 1209 1210
            matched = 1;
            break;
        }
1211
        ++index;
1212 1213 1214 1215
    }

    if(matched)
    {
1216 1217 1218
        if(enable_msg_tracking && (qi->num_bytes < EAGER_THRESHOLD))
            update_message_size(ns, lp, bf, m, qi, 1, 0);
        
1219 1220
        m->fwd.matched_req = qitem->req_id;
        int is_rend = 0;
1221 1222 1223 1224
        if(qitem->num_bytes >= EAGER_THRESHOLD)
        {
            /* Matching receive found, need to notify the sender to transmit
             * the data */
1225
            bf->c10 = 1;
1226 1227
            is_rend = 1;
            send_ack_back(ns, bf, m, lp, qi, qitem->req_id);
1228
        }
1229

1230
        m->rc.saved_recv_time = ns->recv_time;
1231 1232
        ns->recv_time += (tw_now(lp) - qitem->req_init_time);

1233 1234 1235 1236 1237 1238
        /*if(ns->nw_id == (tw_lpid)TRACK_LP && qitem->op_type == CODES_WK_IRECV)
        {
            tw_output(lp, "\n Completed recv req id %d ", qitem->req_id);
            print_completed_queue(lp, &ns->completed_reqs);
        }*/
        
1239 1240 1241
        if(qitem->op_type == CODES_WK_IRECV && !is_rend)
        {
            bf->c9 = 1;
1242
            update_completed_queue(ns, bf, m, lp, qitem->req_id);
1243
        }
1244 1245 1246 1247 1248
        else
         if(qitem->op_type == CODES_WK_RECV && !is_rend)
         {
            bf->c6 = 1;
            codes_issue_next_event(lp);
1249
            //get_next_mpi_operation(s, bf, lp, m);
1250
         }
1251

1252

1253
        qlist_del(&qi->ql);
1254

1255
	    rc_stack_push(lp, qi, free, ns->processed_ops);
1256
        return index;
1257 1258 1259 1260 1261
    }
    return -1;
}
static void codes_issue_next_event_rc(tw_lp * lp)
{
1262
	    tw_rand_reverse_unif(lp->rng);
1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273
}

/* 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);
1274
   assert(ts > 0);
1275
   e = tw_event_new( lp->gid, ts, lp );
1276
   msg = (nw_message*)tw_event_data(e);
1277 1278 1279 1280 1281 1282 1283

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

/* Simulate delays between MPI operations */
static void codes_exec_comp_delay(
1284
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
1285 1286 1287 1288 1289
{
	tw_event* e;
	tw_stime ts;
	nw_message* msg;

1290
    m->rc.saved_delay = s->compute_time;
1291 1292 1293 1294 1295
    s->compute_time += mpi_op->u.delay.nsecs;
    ts = mpi_op->u.delay.nsecs;
    if(ts <= 0)
        ts = g_tw_lookahead + 0.1 + tw_rand_exponential(lp->rng, noise);
    //ts = s_to_ns(mpi_op->u.delay.seconds);
1296

1297
//	ts += g_tw_lookahead + 0.1 + tw_rand_exponential(lp->rng, noise);
1298
    assert(ts > 0);