model-net-mpi-replay.c 82.7 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 39 40 41 42
/* 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;
43
tw_lpid TRACK_LP = -1;
44

45
static double total_syn_data = 0;
46
static int unmatched = 0;
47 48 49 50 51
char workload_type[128];
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 59 60

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

66 67 68 69
/* variables for loading multiple applications */
char workloads_conf_file[8192];
char alloc_file[8192];
int num_traces_of_job[5];
70 71 72
tw_stime soft_delay_mpi = 2500;
tw_stime nic_delay = 1000;
tw_stime copy_per_byte_eager = 0.55;
73 74 75 76 77
char file_name_of_job[5][8192];

struct codes_jobmap_ctx *jobmap_ctx;
struct codes_jobmap_params_list jobmap_p;

78 79
/* Variables for Cortex Support */
/* Matthieu's additions start */
80
#ifdef ENABLE_CORTEX_PYTHON
81 82 83
static char cortex_file[512] = "\0";
static char cortex_class[512] = "\0";
static char cortex_gen[512] = "\0";
84
#endif
85 86
/* Matthieu's additions end */

87 88
typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
89
typedef unsigned int dumpi_req_id;
90 91

static int net_id = 0;
92
static float noise = 2.0;
93 94 95
static int num_nw_lps = 0, num_mpi_lps = 0;

static int num_syn_clients;
96

97
FILE * workload_log = NULL;
98
FILE * msg_size_log = NULL;
99 100 101 102
FILE * workload_agg_log = NULL;
FILE * workload_meta_log = NULL;

static uint64_t sample_bytes_written = 0;
103

104 105 106
long long num_bytes_sent=0;
long long num_bytes_recvd=0;

107 108 109
long long num_syn_bytes_sent = 0;
long long num_syn_bytes_recvd = 0;

110 111 112 113 114 115
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;
116 117 118
static int enable_sampling = 0;
static double sampling_interval = 5000000;
static double sampling_end_time = 3000000000;
119
static int enable_debug = 0;
120

121 122 123
/* set group context */
struct codes_mctx group_ratio;

124
/* MPI_OP_GET_NEXT is for getting next MPI operation when the previous operation completes.
125
* 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.
126 127 128 129 130 131 132
* 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,
133 134 135 136 137 138
    MPI_REND_ARRIVED,
    MPI_REND_ACK_ARRIVED,
    CLI_BCKGND_FIN,
    CLI_BCKGND_ARRIVE,
    CLI_BCKGND_GEN,
    CLI_NBR_FINISH,
139 140
};

141 142 143 144
struct mpi_workload_sample
{
    /* Sampling data */
    int nw_id;
145
    int app_id;
146 147 148 149 150
    unsigned long num_sends_sample;
    unsigned long num_bytes_sample;
    unsigned long num_waits_sample;
    double sample_end_time;
};
151 152 153 154 155 156 157
/* 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;
158
    int64_t num_bytes;
159 160 161 162 163 164 165 166
    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
{
167
	unsigned int req_id;
168
    struct qlist_head ql;
169
    int index;
170 171 172 173 174 175
};

/* for wait operations, store the pending operation and number of completed waits so far. */
struct pending_waits
{
    int op_type;
176
    unsigned int req_ids[MAX_WAIT_REQS];
177
	int num_completed;
178 179
	int count;
    tw_stime start_time;
180 181 182
    struct qlist_head ql;
};

183 184 185 186 187 188 189 190 191
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; 
};
192 193 194 195 196 197 198 199 200 201
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;
202 203
    int app_id;
    int local_rank;
204

205 206 207
    int is_finished;
    int neighbor_completed;

208
    struct rc_stack * processed_ops;
209
    struct rc_stack * processed_wait_op;
210
    struct rc_stack * matched_reqs;
211
//    struct rc_stack * indices;
212 213 214 215 216 217 218 219 220 221

    /* 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;

222

223 224
	/* time spent by the LP in executing the app trace*/
	double start_time;
225 226 227 228 229 230 231 232 233

    double col_time;

    double reduce_time;
    int num_reduce;

    double all_reduce_time;
    int num_all_reduce;

234 235 236 237 238 239 240 241 242 243 244 245 246 247 248
	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;
249

250
    tw_stime cur_interval_end;
251
    
252 253
    /* Pending wait operation */
    struct pending_waits * wait_op;
254

255 256 257 258 259 260
    /* Message size latency information */
    struct qhash_table * msg_sz_table;
    struct qlist_head msg_sz_list;

    /* quick hash for maintaining message latencies */

261 262 263
    unsigned long num_bytes_sent;
    unsigned long num_bytes_recvd;

264 265 266
    unsigned long syn_data;
    unsigned long gen_data;
    
267 268 269 270
    /* For sampling data */
    int sampling_indx;
    int max_arr_size;
    struct mpi_workload_sample * mpi_wkld_samples;
271
    char output_buf[512];
272
    char col_stats[64];
273 274 275 276
};

/* data for handling reverse computation.
* saved_matched_req holds the request ID of matched receives/sends for wait operations.
277
* 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.
278 279 280
* 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
{
281
   // forward message handler
282
   int msg_type;
283
   int op_type;
284
   model_net_event_return event_rc;
285

286 287 288
   struct
   {
       tw_lpid src_rank;
289
       int dest_rank;
290
       int64_t num_bytes;
291 292 293 294 295
       int num_matched;
       int data_type;
       double sim_start_time;
       // for callbacks - time message was received
       double msg_send_time;
296
       unsigned int req_id;
297
       int matched_req;
298
       int tag;
299
       int app_id;
300 301
       int found_match;
       short wait_completed;
302
       short rend_send;
303 304 305 306 307 308 309
   } fwd;
   struct
   {
       double saved_send_time;
       double saved_recv_time;
       double saved_wait_time;
       double saved_delay;
310
       int64_t saved_num_bytes;
311
   } rc;
312 313
};

314
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);
315 316

static void send_ack_back_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
317 318
/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
319
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op, int is_rend);
320 321
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
322
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
323 324
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
325
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp);
326 327
/* execute the computational delay */
static void codes_exec_comp_delay(
328
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
329 330 331 332 333 334 335 336 337 338 339 340 341 342 343
/* 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 */
344 345 346 347 348 349 350 351
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);
352 353 354 355 356 357 358 359 360 361 362 363 364 365 366
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);

367
/*static void update_message_size_rc(
368 369 370 371
        struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
372
{*/
373
/*TODO: Complete reverse handler */
374
/*    (void)ns;
375 376 377
    (void)lp;
    (void)bf;
    (void)m;
378
}*/
379 380 381 382 383 384 385 386 387 388
/* 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)
{
389 390 391
            (void)bf;
            (void)is_eager;

392 393 394 395 396 397 398 399 400 401 402 403 404 405
            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)
            {
406
                struct msg_size_info * msg_info = (struct msg_size_info*)malloc(sizeof(struct msg_size_info));
407 408
                msg_info->msg_size = qitem->num_bytes;
                msg_info->num_msgs = 1;
409
                msg_info->agg_latency = tw_now(lp) - msg_init_time;
410
                msg_info->avg_latency = msg_info->agg_latency;
411
                qhash_add(ns->msg_sz_table, &(msg_info->msg_size), msg_info->hash_link);
412 413 414 415 416 417 418
                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++;
419
                tmp->agg_latency += tw_now(lp) - msg_init_time;  
420 421 422 423 424 425 426 427 428 429
                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)
{
430 431 432
    (void)ns;
    (void)bf;
    (void)m;
433 434 435 436 437
        
    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); 
438 439 440 441 442 443 444 445
}

static void notify_background_traffic(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
446 447 448
        (void)bf;
        (void)m;

449 450 451 452 453 454 455 456 457 458 459 460 461 462 463
        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);

464
            lprintf("\n Other ranks %d ", num_other_ranks);
465 466 467 468 469 470 471 472 473 474 475 476
            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);
477
                m_new = (struct nw_message*)tw_event_data(e);
478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510
                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)
    {
511
//        printf("\n All workloads completed, notifying background traffic ");
512 513 514 515 516 517 518 519 520 521 522 523 524
        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;

525
//        printf("\n Local rank %d notifying neighbor %d ", ns->local_rank, ns->local_rank+1);
526 527 528 529 530 531 532 533 534 535
        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);
536
        m_new = (struct nw_message*)tw_event_data(e); 
537 538 539 540 541 542 543 544 545 546
        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)
{
547 548 549 550
        (void)b;
        (void)msg;
        (void)lp;

551 552 553 554 555 556 557 558 559
        ns->is_finished = 0;
        return;
}
void finish_bckgnd_traffic(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
560 561
        (void)b;
        (void)msg;
562
        ns->is_finished = 1;
563
        lprintf("\n LP %llu completed sending data %lu completed at time %lf ", LLU(lp->gid), ns->gen_data, tw_now(lp));
564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645
        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 */  
646
    tw_stime ts = mean_interval  + tw_rand_exponential(lp->rng, noise); 
647 648 649
    tw_event * e;
    nw_message * m_new;
    e = tw_event_new(lp->gid, ts, lp);
650
    m_new = (struct nw_message*)tw_event_data(e);
651 652 653 654 655 656
    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)
{
657 658 659
    (void)bf;
    (void)m;
    (void)lp;
660 661 662 663 664 665 666
//    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)
{
667 668 669
    (void)bf;
    (void)lp;

670 671 672 673 674
//    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;
}
675
/* Debugging functions, may generate unused function warning */
676
/*static void print_waiting_reqs(uint32_t * reqs, int count)
677
{
678
    lprintf("\n Waiting reqs: %d count", count);
679 680
    int i;
    for(i = 0; i < count; i++ )
681
        lprintf(" %d ", reqs[i]);
682
}*/
683 684 685 686 687 688 689 690 691 692 693 694
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);
695
            printf(" \n Source %d Dest %d bytes %"PRId64" tag %d ", current->source_rank, current->dest_rank, current->num_bytes, current->tag);
696 697
       }
}
698
static void print_completed_queue(tw_lp * lp, struct qlist_head * head)
699
{
700
//    printf("\n Completed queue: ");
701 702
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
703
      tw_output(lp, "\n");
704 705 706
      qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, completed_requests, ql);
707
            tw_output(lp, " %llu ", current->req_id);
708
       }
709
}
710
static int clear_completed_reqs(nw_state * s,
711
        tw_lp * lp,
712
        unsigned int * reqs, int count)
713
{
714 715 716
    (void)s;
    (void)lp;

717
    int i, matched = 0;
718

719 720 721
    for( i = 0; i < count; i++)
    {
      struct qlist_head * ent = NULL;
722 723 724
      struct completed_requests * current = NULL;
      struct completed_requests * prev = NULL;

725
      int index = 0;
726 727
      qlist_for_each(ent, &s->completed_reqs)
       {
728 729
           if(prev)
           {
730
              rc_stack_push(lp, prev, free, s->matched_reqs);
731 732
              prev = NULL;
           }
733
            
734 735
           current = qlist_entry(ent, completed_requests, ql);
           current->index = index; 
736 737
            if(current->req_id == reqs[i])
            {
738
                ++matched;
739
                qlist_del(&current->ql);
740
                prev = current;
741
            }
742
            ++index;
743
       }
744 745

      if(prev)
746 747 748 749
      {
         rc_stack_push(lp, prev, free, s->matched_reqs);
         prev = NULL;
      }
750
    }
751
    return matched;
752
}
753
static void add_completed_reqs(nw_state * s,
754 755
        tw_lp * lp,
        int count)
756
{
757
    (void)lp;
758
    for(int i = 0; i < count; i++)
759
    {
760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786
       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.
       if(preserve_wait_ordering)
       {
            if(req->index == 0)
            {
                qlist_add(&req->ql, &s->completed_reqs);
            }
            else
            {
                int index = 1;
                struct qlist_head * ent = NULL;
                qlist_for_each(ent, &s->completed_reqs) 
                {
                    if(index == req->index)
                    {
                        qlist_add(&req->ql, ent);
                    }
                }//end qlist
            }// end else*/
       }
       else
       {
               qlist_add(&req->ql, &s->completed_reqs);
       }
    }//end for
787
}
788

789 790 791 792 793 794
/* 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);
}

795
static int notify_posted_wait(nw_state* s,
796
        tw_bf * bf, nw_message * m, tw_lp * lp,
797
        unsigned int completed_req)
798
{
799 800
    (void)bf;

801 802
    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;
803

804
    m->fwd.wait_completed = 0;
805

806 807
    if(!wait_elem)
        return 0;
808

809
    int op_type = wait_elem->op_type;
810

811 812 813
    if(op_type == CODES_WK_WAIT &&
            (wait_elem->req_ids[0] == completed_req))
    {
814
            m->fwd.wait_completed = 1;
815 816
            wait_completed = 1;
    }
817 818
    else if(op_type == CODES_WK_WAITALL
            || op_type == CODES_WK_WAITANY
819 820 821 822 823 824
            || op_type == CODES_WK_WAITSOME)
    {
        int i;
        for(i = 0; i < wait_elem->count; i++)
        {
            if(wait_elem->req_ids[i] == completed_req)
825
            {
826
                wait_elem->num_completed++;
827
                if(wait_elem->num_completed > wait_elem->count)
828
                    printf("\n Num completed %d count %d LP %llu ",
829 830
                            wait_elem->num_completed,
                            wait_elem->count,
831
                            LLU(lp->gid));
832 833
//                if(wait_elem->num_completed > wait_elem->count)
//                    tw_lp_suspend(lp, 1, 0);
834

835
                if(wait_elem->num_completed >= wait_elem->count)
836
                {
837
                    if(enable_debug)
838
                        fprintf(workload_log, "\n(%lf) APP ID %d MPI WAITALL COMPLETED AT %llu ", tw_now(lp), s->app_id, LLU(s->nw_id));
839
                    wait_completed = 1;
840
                }
841

842
                m->fwd.wait_completed = 1;
843
            }
844
        }
845
    }
846
    return wait_completed;
847
}
848

849
/* reverse handler of MPI wait operation */
850
static void codes_exec_mpi_wait_rc(nw_state* s, tw_bf * bf, tw_lp* lp, nw_message * m)
851
{
852
   if(bf->c1)
853
    {
854
        completed_requests * qi = (completed_requests*)rc_stack_pop(s->processed_ops);
855 856
        if(m->fwd.found_match == 0)
        {
857
            qlist_add(&qi->ql, &s->completed_reqs);
858
        }
859 860 861 862 863 864 865 866 867 868 869 870 871
        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++;
           }
872
        }
873
        codes_issue_next_event_rc(lp);
874
        return;
875
    }
876 877 878
         struct pending_waits * wait_op = s->wait_op;
         free(wait_op);
         s->wait_op = NULL;
879
}
880

881
/* execute MPI wait operation */
882
static void codes_exec_mpi_wait(nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
883
{
884
    /* check in the completed receives queue if the request ID has already been completed.*/
885 886
                
//    printf("\n Wait posted rank id %d ", s->nw_id);
887
    assert(!s->wait_op);
888
    unsigned int req_id = mpi_op->u.wait.req_id;
889

890
    struct completed_requests* current = NULL;
891

892
    struct qlist_head * ent = NULL;
893
    int index = 0;
894 895 896 897 898
    qlist_for_each(ent, &s->completed_reqs)
    {
        current = qlist_entry(ent, completed_requests, ql);
        if(current->req_id == req_id)
        {
899
            bf->c1=1;
900
            qlist_del(&current->ql);
901
            rc_stack_push(lp, current, free, s->processed_ops);
902
            codes_issue_next_event(lp);
903
            m->fwd.found_match = index;
904
            if(s->nw_id == (tw_lpid)TRACK_LP)
905 906 907
            {
                tw_output(lp, "\n wait matched at post %d ", req_id);
                print_completed_queue(lp, &s->completed_reqs);
908
            }
909 910
            return;
        }
911
        ++index;
912
    }
913

914 915 916 917 918
    /*if(s->nw_id == (tw_lpid)TRACK_LP)
    {
        tw_output(lp, "\n wait posted %llu ", req_id);
        print_completed_queue(lp, &s->completed_reqs);
    }*/
919
    /* If not, add the wait operation in the pending 'waits' list. */
920
    struct pending_waits* wait_op = (struct pending_waits*)malloc(sizeof(struct pending_waits));
921 922 923
    wait_op->op_type = mpi_op->op_type;
    wait_op->req_ids[0] = req_id;
    wait_op->count = 1;
924 925
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
926
    s->wait_op = wait_op;
927

928
    return;
929 930
}

931
static void codes_exec_mpi_wait_all_rc(
932
        nw_state* s,
933 934
        tw_bf * bf,
        nw_message * m,
935
        tw_lp* lp)
936
{
937 938 939 940 941 942 943 944 945 946 947
  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--;
    }
  }
948 949 950 951 952 953 954 955
  if(s->wait_op)
  {
      struct pending_waits * wait_op = s->wait_op;
      free(wait_op);
      s->wait_op = NULL;
  }
  else
  {
956
      add_completed_reqs(s, lp, m->fwd.num_matched);
957 958 959
      codes_issue_next_event_rc(lp);
  }
  return;
960
}
961

962
static void codes_exec_mpi_wait_all(
963
        nw_state* s,
964 965
        tw_bf * bf,
        nw_message * m,
966
        tw_lp* lp,
967
        struct codes_workload_op * mpi_op)
968
{
969
  if(enable_debug)
970
    fprintf(workload_log, "\n MPI WAITALL POSTED AT %llu ", LLU(s->nw_id));
971

972 973 974 975 976 977 978 979
  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;
980
        s->mpi_wkld_samples[indx].app_id = s->app_id;
981 982 983 984 985 986
        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)
    {
987
        struct mpi_workload_sample * tmp = (struct mpi_workload_sample*)calloc((MAX_STATS + s->max_arr_size), sizeof(struct mpi_workload_sample));
988 989 990 991 992 993 994 995
        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++;
  }
996
  int count = mpi_op->u.waits.count;
997 998
  /* If the count is not less than max wait reqs then stop */
  assert(count < MAX_WAIT_REQS);
999

1000
  int i = 0, num_matched = 0;
1001
  m->fwd.num_matched = 0;
1002

1003
  /*if(lp->gid == TRACK_LP)
1004
  {
1005
      printf("\n MPI Wait all posted ");
1006
      print_waiting_reqs(mpi_op->u.waits.req_ids, count);
1007 1008
      print_completed_queue(lp, &s->completed_reqs);
  }*/
1009
      /* check number of completed irecvs in the completion queue */
1010 1011
  for(i = 0; i < count; i++)
  {
1012
      unsigned int req_id = mpi_op->u.waits.req_ids[i];
1013 1014 1015 1016
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
      qlist_for_each(ent, &s->completed_reqs)
       {
1017
            current = qlist_entry(ent, struct completed_requests, ql);
1018 1019 1020 1021
            if(current->req_id == req_id)
                num_matched++;
       }
  }
1022

1023
  m->fwd.found_match = num_matched;
1024 1025 1026 1027
  if(num_matched == count)
  {
    /* No need to post a MPI Wait all then, issue next event */
      /* Remove all completed requests from the list */
1028 1029 1030
      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);
1031 1032
      s->wait_op = NULL;
      codes_issue_next_event(lp);
1033 1034
  }
  else
1035 1036
  {
      /* If not, add the wait operation in the pending 'waits' list. */
1037
	  struct pending_waits* wait_op = (struct pending_waits*)malloc(sizeof(struct pending_waits));
1038 1039 1040 1041 1042 1043 1044 1045
	  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;
1046
	  wait_op->start_time = tw_now(lp);
1047
      s->wait_op = wait_op;
1048
  }
1049 1050
  return;
}
1051

1052 1053
/* search for a matching mpi operation and remove it from the list.
 * Record the index in the list from where the element got deleted.
1054
 * Index is used for inserting the element once again in the queue for reverse computation. */
1055
static int rm_matching_rcv(nw_state * ns,
1056
        tw_bf * bf,
1057 1058
        nw_message * m,
        tw_lp * lp,
1059
        mpi_msgs_queue * qitem)
1060 1061
{
    int matched = 0;
1062
    int index = 0;
1063
    int is_rend = 0;
1064 1065
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;
1066

1067 1068
    qlist_for_each(ent, &ns->pending_recvs_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
1069 1070
        if(//(qi->num_bytes == qitem->num_bytes)
                //&& 
1071
               ((qi->tag == qitem->tag) || qi->tag == -1)
1072
                && ((qi->source_rank == qitem->source_rank) || qi->source_rank == -1))
1073 1074
        {
            matched = 1;
1075
            qi->num_bytes = qitem->num_bytes;
1076 1077
            break;
        }
1078
        ++index;
1079
    }
1080

1081 1082
    if(matched)
    {
1083 1084 1085 1086 1087 1088 1089
        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
1090 1091
             * the data * (only works in sequential mode)*/
            bf->c10 = 1;
1092 1093
            is_rend = 1;
            send_ack_back(ns, bf, m, lp, qitem, qi->req_id);
1094
        }
1095 1096 1097 1098 1099 1100
        else
        {
            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)
1101
        {
1102
            bf->c9 = 1;
1103 1104 1105 1106
            /*if(ns->nw_id == (tw_lpid)TRACK_LP)
            {
                printf("\n Completed irecv req id %d ", qi->req_id);
            }*/
1107
            update_completed_queue(ns, bf, m, lp, qi->req_id);
1108
        }
1109 1110 1111
        else if(qi->op_type == CODES_WK_RECV && !is_rend)
        {
            bf->c8 = 1;
1112
            codes_issue_next_event(lp);
1113
        }
1114

1115
        qlist_del(&qi->ql);
1116

1117
        rc_stack_push(lp, qi, free, ns->processed_ops);
1118
        return index;
1119 1120 1121 1122
    }
    return -1;
}

1123
static int rm_matching_send(nw_state * ns,
1124 1125 1126
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp, mpi_msgs_queue * qitem)
1127 1128 1129 1130 1131
{
    int matched = 0;
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;

1132
    int index = 0;
1133 1134
    qlist_for_each(ent, &ns->arrival_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
1135 1136 1137
        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)
1138 1139
                && ((qi->source_rank == qitem->source_rank) || qitem->source_rank == -1))
        {
1140
            qitem->num_bytes = qi->num_bytes;
1141 1142 1143
            matched = 1;
            break;
        }
1144
        ++index;
1145 1146 1147 1148
    }

    if(matched)
    {
1149 1150 1151
        if(enable_msg_tracking && (qi->num_bytes < EAGER_THRESHOLD))
            update_message_size(ns, lp, bf, m, qi, 1, 0);
        
1152 1153
        m->fwd.matched_req = qitem->req_id;
        int is_rend = 0;
1154 1155 1156 1157
        if(qitem->num_bytes >= EAGER_THRESHOLD)
        {
            /* Matching receive found, need to notify the sender to transmit
             * the data */
1158
            bf->c10 = 1;
1159 1160
            is_rend = 1;
            send_ack_back(ns, bf, m, lp, qi, qitem->req_id);
1161
        }
1162

1163
        m->rc.saved_recv_time = ns->recv_time;
1164 1165
        ns->recv_time += (tw_now(lp) - qitem->req_init_time);

1166 1167 1168 1169 1170 1171
        /*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);
        }*/
        
1172 1173 1174
        if(qitem->op_type == CODES_WK_IRECV && !is_rend)
        {
            bf->c9 = 1;
1175
            update_completed_queue(ns, bf, m, lp, qitem->req_id);
1176
        }
1177 1178 1179 1180 1181 1182
        else
         if(qitem->op_type == CODES_WK_RECV && !is_rend)
         {
            bf->c6 = 1;
            codes_issue_next_event(lp);
         }
1183

1184

1185
        qlist_del(&qi->ql);
1186

1187
	    rc_stack_push(lp, qi, free, ns->processed_ops);
1188
        return index;
1189 1190 1191 1192 1193
    }
    return -1;
}
static void codes_issue_next_event_rc(tw_lp * lp)
{
1194
	    tw_rand_reverse_unif(lp->rng);
1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205
}

/* 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);
1206
   assert(ts > 0);
1207
   e = tw_event_new( lp->gid, ts, lp );
1208
   msg = (nw_message*)tw_event_data(e);
1209 1210 1211 1212 1213 1214 1215

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

/* Simulate delays between MPI operations */
static void codes_exec_comp_delay(
1216
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
1217 1218 1219 1220 1221
{
	tw_event* e;
	tw_stime ts;
	nw_message* msg;

1222
    m->rc.saved_delay = s->compute_time;
1223 1224
    s->compute_time += s_to_ns(mpi_op->u.delay.seconds);
    ts = s_to_ns(mpi_op->u.delay.seconds);
1225 1226

	ts += g_tw_lookahead + 0.1 + tw_rand_exponential(lp->rng, noise);
1227
    assert(ts > 0);
1228

1229
	e = tw_event_new( lp->gid, ts , lp );
1230
	msg = (nw_message*)tw_event_data(e);
1231
	msg->msg_type = MPI_OP_GET_NEXT;
1232 1233
	tw_event_send(e);

1234 1235 1236
}

/* reverse computation operation for MPI irecv */
1237
static void codes_exec_mpi_recv_rc(
1238 1239 1240
        nw_state* ns,
        tw_bf * bf,
        nw_message* m,
1241
        tw_lp* lp)
1242
{
1243
	ns->recv_time = m->rc.saved_recv_time;
1244 1245 1246 1247

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

1248
	if(m->fwd.found_match >= 0)
1249
	  {
1250
		ns->recv_time = m->rc.saved_recv_time;
1251
        //int queue_count = qlist_count(&ns->arrival_queue);
1252

1253
        mpi_msgs_queue * qi = (mpi_msgs_queue*)rc_stack_pop(ns->processed_ops);
1254

1255
        if(m->fwd.found_match == 0)
1256 1257 1258
        {
            qlist_add(&qi->ql, &ns->arrival_queue);
        }
1259
        else 
1260
        {
1261 1262 1263 1264
            int index = 1;
            struct qlist_head * ent = NULL;
            qlist_for_each(ent, &ns->arrival_queue)
            {
1265
               if(index == m->fwd.found_match)
1266 1267 1268 1269
               {
                 qlist_add(&qi->ql, ent);
                 break;
               }
1270
               index++;
1271
            }
1272
        }
1273
        if(bf->c9)
1274
        {
1275
            update_completed_queue_rc(ns, bf, m, lp);
1276
        }
1277 1278
        if(bf->c6)
            codes_issue_next_event_rc(lp);
1279
      }
1280
	else if(m->fwd.found_match < 0)
1281
	    {
1282
            struct qlist_head * ent = qlist_pop(&ns->pending_recvs_queue);
1283 1284
            mpi_msgs_queue * qi = qlist_entry(ent, mpi_msgs_queue, ql);
            free(qi);