model-net-mpi-replay.c 77.5 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
tw_lpid TRACK_LP = -1;
40 41

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

50
static int alloc_spec = 0;
51 52
static tw_stime self_overhead = 10.0;
static tw_stime mean_interval = 100000;
53 54 55

/* Doing LP IO*/
static char lp_io_dir[256] = {'\0'};
56
static char sampling_dir[32] = {'\0'};
57 58 59 60
static lp_io_handle io_handle;
static unsigned int lp_io_use_suffix = 0;
static int do_lp_io = 0;

61 62 63 64
/* variables for loading multiple applications */
char workloads_conf_file[8192];
char alloc_file[8192];
int num_traces_of_job[5];
65 66 67
tw_stime soft_delay_mpi = 2500;
tw_stime nic_delay = 1000;
tw_stime copy_per_byte_eager = 0.55;
68 69 70 71 72
char file_name_of_job[5][8192];

struct codes_jobmap_ctx *jobmap_ctx;
struct codes_jobmap_params_list jobmap_p;

73 74
/* Variables for Cortex Support */
/* Matthieu's additions start */
75
#ifdef ENABLE_CORTEX_PYTHON
76 77 78
static char cortex_file[512] = "\0";
static char cortex_class[512] = "\0";
static char cortex_gen[512] = "\0";
79
#endif
80 81
/* Matthieu's additions end */

82 83
typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
84
typedef int32_t dumpi_req_id;
85 86 87

static int net_id = 0;
static float noise = 5.0;
88 89 90
static int num_nw_lps = 0, num_mpi_lps = 0;

static int num_syn_clients;
91

92
FILE * workload_log = NULL;
93
FILE * msg_size_log = NULL;
94 95 96 97
FILE * workload_agg_log = NULL;
FILE * workload_meta_log = NULL;

static uint64_t sample_bytes_written = 0;
98

99 100 101
long long num_bytes_sent=0;
long long num_bytes_recvd=0;

102 103 104
long long num_syn_bytes_sent = 0;
long long num_syn_bytes_recvd = 0;

105 106 107 108 109 110
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;
111 112 113
static int enable_sampling = 0;
static double sampling_interval = 5000000;
static double sampling_end_time = 3000000000;
114
static int enable_debug = 0;
115

116 117 118
/* set group context */
struct codes_mctx group_ratio;

119
/* MPI_OP_GET_NEXT is for getting next MPI operation when the previous operation completes.
120
* 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.
121 122 123 124 125 126 127
* 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,
128 129 130 131 132 133
    MPI_REND_ARRIVED,
    MPI_REND_ACK_ARRIVED,
    CLI_BCKGND_FIN,
    CLI_BCKGND_ARRIVE,
    CLI_BCKGND_GEN,
    CLI_NBR_FINISH,
134 135
};

136 137 138 139
struct mpi_workload_sample
{
    /* Sampling data */
    int nw_id;
140
    int app_id;
141 142 143 144 145
    unsigned long num_sends_sample;
    unsigned long num_bytes_sample;
    unsigned long num_waits_sample;
    double sample_end_time;
};
146 147 148 149 150 151 152
/* 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;
153
    uint64_t num_bytes;
154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169
    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;
170
    int32_t req_ids[MAX_WAIT_REQS];
171
	int num_completed;
172 173
	int count;
    tw_stime start_time;
174 175 176
    struct qlist_head ql;
};

177 178 179 180 181 182 183 184 185
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; 
};
186 187 188 189 190 191 192 193 194 195
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;
196 197
    int app_id;
    int local_rank;
198

199 200 201
    int is_finished;
    int neighbor_completed;

202
    struct rc_stack * processed_ops;
203
    struct rc_stack * processed_wait_op;
204
    struct rc_stack * matched_reqs;
205 206 207 208 209 210 211 212 213 214

    /* 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;

215

216 217
	/* time spent by the LP in executing the app trace*/
	double start_time;
218 219 220 221 222 223 224 225 226

    double col_time;

    double reduce_time;
    int num_reduce;

    double all_reduce_time;
    int num_all_reduce;

227 228 229 230 231 232 233 234 235 236 237 238 239 240 241
	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;
242

243
    tw_stime cur_interval_end;
244
    
245 246
    /* Pending wait operation */
    struct pending_waits * wait_op;
247

248 249 250 251 252 253
    /* Message size latency information */
    struct qhash_table * msg_sz_table;
    struct qlist_head msg_sz_list;

    /* quick hash for maintaining message latencies */

254 255 256
    unsigned long num_bytes_sent;
    unsigned long num_bytes_recvd;

257 258 259
    unsigned long syn_data;
    unsigned long gen_data;
    
260 261 262 263
    /* For sampling data */
    int sampling_indx;
    int max_arr_size;
    struct mpi_workload_sample * mpi_wkld_samples;
264
    char output_buf[512];
265 266 267 268
};

/* data for handling reverse computation.
* saved_matched_req holds the request ID of matched receives/sends for wait operations.
269
* 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.
270 271 272
* 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
{
273
   // forward message handler
274
   int msg_type;
275
   int op_type;
276
   model_net_event_return event_rc;
277

278 279 280
   struct
   {
       tw_lpid src_rank;
281
       int dest_rank;
282
       uint64_t num_bytes;
283 284 285 286 287
       int num_matched;
       int data_type;
       double sim_start_time;
       // for callbacks - time message was received
       double msg_send_time;
288
       int32_t req_id;
289
       int tag;
290
       int app_id;
291 292 293 294 295 296 297 298 299
       int found_match;
       short wait_completed;
   } fwd;
   struct
   {
       double saved_send_time;
       double saved_recv_time;
       double saved_wait_time;
       double saved_delay;
300
       int32_t saved_num_bytes;
301
   } rc;
302 303
};

304
static void send_ack_back(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, mpi_msgs_queue * mpi_op);
305 306

static void send_ack_back_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
307 308
/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
309
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op, int is_rend);
310 311
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
312
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
313 314
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
315
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp);
316 317
/* execute the computational delay */
static void codes_exec_comp_delay(
318
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
319 320 321 322 323 324 325 326 327 328 329 330 331 332 333
/* 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 */
334 335 336 337 338 339 340 341
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);
342 343 344 345 346 347 348 349 350 351 352 353 354 355 356
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);

357
/*static void update_message_size_rc(
358 359 360 361
        struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
362
{*/
363
/*TODO: Complete reverse handler */
364
/*    (void)ns;
365 366 367
    (void)lp;
    (void)bf;
    (void)m;
368
}*/
369 370 371 372 373 374 375 376 377 378
/* 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)
{
379 380 381
            (void)bf;
            (void)is_eager;

382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398
            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;
399
                msg_info->agg_latency = tw_now(lp) - msg_init_time;
400
                msg_info->avg_latency = msg_info->agg_latency;
401
                qhash_add(ns->msg_sz_table, &(msg_info->msg_size), msg_info->hash_link);
402 403 404 405 406 407 408
                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++;
409
                tmp->agg_latency += tw_now(lp) - msg_init_time;  
410 411 412 413 414 415 416 417 418 419
                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)
{
420 421 422
    (void)ns;
    (void)bf;
    (void)m;
423 424 425 426 427 428 429 430 431
    tw_rand_reverse_unif(lp->rng); 
}

static void notify_background_traffic(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
432 433 434
        (void)bf;
        (void)m;

435 436 437 438 439 440 441 442 443 444 445 446 447 448 449
        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);

450
            lprintf("\n Other ranks %d ", num_other_ranks);
451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532
            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)
    {
        printf("\n All workloads completed, notifying background traffic ");
        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;

        printf("\n Local rank %d notifying neighbor %d ", ns->local_rank, ns->local_rank+1);
        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)
{
533 534 535 536
        (void)b;
        (void)msg;
        (void)lp;

537 538 539 540 541 542 543 544 545
        ns->is_finished = 0;
        return;
}
void finish_bckgnd_traffic(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
546 547
        (void)b;
        (void)msg;
548
        ns->is_finished = 1;
549
        lprintf("\n LP %llu completed sending data %lu completed at time %lf ", lp->gid, ns->gen_data, tw_now(lp));
550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632
        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)
{
    printf("\n Workload completed, notifying neighbor ");
    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 */  
633
    tw_stime ts = mean_interval  + tw_rand_exponential(lp->rng, noise); 
634 635 636 637 638 639 640 641 642 643
    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)
{
644 645 646
    (void)bf;
    (void)m;
    (void)lp;
647 648 649 650 651 652 653
//    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)
{
654 655 656
    (void)bf;
    (void)lp;

657 658 659 660 661
//    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;
}
662
/* Debugging functions, may generate unused function warning */
663
static void print_waiting_reqs(int32_t * reqs, int count)
664
{
665
    lprintf("\n Waiting reqs: %d count", count);
666 667
    int i;
    for(i = 0; i < count; i++ )
668
        lprintf(" %d ", reqs[i]);
669
}
670 671 672 673 674 675 676 677 678 679 680 681
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);
682
            printf(" \n Source %d Dest %d bytes %llu tag %d ", current->source_rank, current->dest_rank, current->num_bytes, current->tag);
683 684
       }
}
685 686 687 688 689 690 691 692
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);
693
            printf(" %d ", current->req_id);
694 695
       }
}
696
static int clear_completed_reqs(nw_state * s,
697
        tw_lp * lp,
698
        int32_t * reqs, int count)
699
{
700 701 702
    (void)s;
    (void)lp;

703
    int i, matched = 0;
704

705 706 707
    for( i = 0; i < count; i++)
    {
      struct qlist_head * ent = NULL;
708 709 710
      struct completed_requests * current = NULL;
      struct completed_requests * prev = NULL;

711 712
      qlist_for_each(ent, &s->completed_reqs)
       {
713 714 715 716 717
            current = qlist_entry(ent, completed_requests, ql);
            
            if(prev)
              rc_stack_push(lp, prev, free, s->matched_reqs);
            
718 719
            if(current->req_id == reqs[i])
            {
720
                ++matched;
721
                qlist_del(&current->ql);
722
                prev = current;
723
            }
724 725
            else
                prev = NULL;
726
       }
727 728 729

      if(prev)
          rc_stack_push(lp, prev, free, s->matched_reqs);
730
    }
731
    return matched;
732
}
733
static void add_completed_reqs(nw_state * s,
734 735
        tw_lp * lp,
        int count)
736
{
737
    (void)lp;
738 739 740
    int i;
    for( i = 0; i < count; i++)
    {
741 742
       struct completed_requests * req = rc_stack_pop(s->matched_reqs);
       qlist_add(&req->ql, &s->completed_reqs);
743 744
    }
}
745

746 747 748 749 750 751
/* 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);
}

752
static int notify_posted_wait(nw_state* s,
753
        tw_bf * bf, nw_message * m, tw_lp * lp,
754
        dumpi_req_id completed_req)
755
{
756 757
    (void)bf;

758 759
    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;
760

761
    m->fwd.wait_completed = 0;
762

763 764
    if(!wait_elem)
        return 0;
765

766
    int op_type = wait_elem->op_type;
767

768 769 770 771 772
    if(op_type == CODES_WK_WAIT &&
            (wait_elem->req_ids[0] == completed_req))
    {
            wait_completed = 1;
    }
773 774
    else if(op_type == CODES_WK_WAITALL
            || op_type == CODES_WK_WAITANY
775 776 777 778 779 780
            || op_type == CODES_WK_WAITSOME)
    {
        int i;
        for(i = 0; i < wait_elem->count; i++)
        {
            if(wait_elem->req_ids[i] == completed_req)
781
            {
782
                wait_elem->num_completed++;
783
                if(wait_elem->num_completed > wait_elem->count)
784
                    printf("\n Num completed %d count %d LP %llu ",
785 786 787
                            wait_elem->num_completed,
                            wait_elem->count,
                            lp->gid);
788 789
//                if(wait_elem->num_completed > wait_elem->count)
//                    tw_lp_suspend(lp, 1, 0);
790

791
                if(wait_elem->num_completed >= wait_elem->count)
792
                {
793
                    if(enable_debug)
794
                        fprintf(workload_log, "\n(%lf) APP ID %d MPI WAITALL COMPLETED AT %llu ", tw_now(lp), s->app_id, s->nw_id);
795
                    wait_completed = 1;
796
                }
797

798
                m->fwd.wait_completed = 1;
799
            }
800
        }
801
    }
802
    return wait_completed;
803
}
804

805
/* reverse handler of MPI wait operation */
806
static void codes_exec_mpi_wait_rc(nw_state* s, tw_bf * bf, tw_lp* lp)
807
{
808
    if(bf->c2)
809
     {
810 811 812
         struct pending_waits * wait_op = s->wait_op;
         free(wait_op);
         s->wait_op = NULL;
813
     }
814
   if(bf->c1)
815 816
    {
        codes_issue_next_event_rc(lp);
817
        completed_requests * qi = rc_stack_pop(s->processed_ops);
818
        qlist_add(&qi->ql, &s->completed_reqs);
819
    }
820
    return;
821
}
822

823
/* execute MPI wait operation */
824
static void codes_exec_mpi_wait(nw_state* s, tw_bf * bf, tw_lp* lp, struct codes_workload_op * mpi_op)
825
{
826 827
    /* check in the completed receives queue if the request ID has already been completed.*/
    assert(!s->wait_op);
828
    dumpi_req_id req_id = mpi_op->u.wait.req_id;
829

830
    struct completed_requests* current = NULL;
831

832 833 834 835 836 837
    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)
        {
838
            bf->c1=1;
839
            qlist_del(&current->ql);
840
            rc_stack_push(lp, current, free, s->processed_ops);
841 842 843 844
            codes_issue_next_event(lp);
            return;
        }
    }
845 846

    bf->c2 = 1;
847 848 849 850 851
    /* 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;
852 853
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
854
    s->wait_op = wait_op;
855

856
    return;
857 858
}

859
static void codes_exec_mpi_wait_all_rc(
860
        nw_state* s,
861 862
        tw_bf * bf,
        nw_message * m,
863
        tw_lp* lp)
864
{
865 866 867 868 869 870 871 872 873 874 875
  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--;
    }
  }
876 877 878 879 880 881 882 883
  if(s->wait_op)
  {
      struct pending_waits * wait_op = s->wait_op;
      free(wait_op);
      s->wait_op = NULL;
  }
  else
  {
884
      add_completed_reqs(s, lp, m->fwd.num_matched);
885 886 887
      codes_issue_next_event_rc(lp);
  }
  return;
888
}
889

890
static void codes_exec_mpi_wait_all(
891
        nw_state* s,
892 893
        tw_bf * bf,
        nw_message * m,
894
        tw_lp* lp,
895
        struct codes_workload_op * mpi_op)
896
{
897
  if(enable_debug)
898
    fprintf(workload_log, "\n MPI WAITALL POSTED AT %llu ", s->nw_id);
899

900 901 902 903 904 905 906 907
  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;
908
        s->mpi_wkld_samples[indx].app_id = s->app_id;
909 910 911 912 913 914
        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)
    {
915
        struct mpi_workload_sample * tmp = calloc((MAX_STATS + s->max_arr_size), sizeof(struct mpi_workload_sample));
916 917 918 919 920 921 922 923
        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++;
  }
924
  int count = mpi_op->u.waits.count;
925 926
  /* If the count is not less than max wait reqs then stop */
  assert(count < MAX_WAIT_REQS);
927

928
  int i = 0, num_matched = 0;
929
  m->fwd.num_matched = 0;
930

931
  if(lp->gid == TRACK_LP)
932
  {
933
      printf("\n MPI Wait all posted ");
934 935
      print_waiting_reqs(mpi_op->u.waits.req_ids, count);
      print_completed_queue(&s->completed_reqs);
936
  }
937
      /* check number of completed irecvs in the completion queue */
938 939 940 941 942 943 944
  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)
       {
945
            current = qlist_entry(ent, struct completed_requests, ql);
946 947 948 949
            if(current->req_id == req_id)
                num_matched++;
       }
  }
950

951
  m->fwd.found_match = num_matched;
952 953 954 955
  if(num_matched == count)
  {
    /* No need to post a MPI Wait all then, issue next event */
      /* Remove all completed requests from the list */
956 957 958
      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);
959 960
      s->wait_op = NULL;
      codes_issue_next_event(lp);
961 962
  }
  else
963 964 965 966 967 968 969 970 971 972 973
  {
      /* 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;
974
	  wait_op->start_time = tw_now(lp);
975
      s->wait_op = wait_op;
976
  }
977 978
  return;
}
979

980 981
/* search for a matching mpi operation and remove it from the list.
 * Record the index in the list from where the element got deleted.
982
 * Index is used for inserting the element once again in the queue for reverse computation. */
983
static int rm_matching_rcv(nw_state * ns,
984
        tw_bf * bf,
985 986
        nw_message * m,
        tw_lp * lp,
987
        mpi_msgs_queue * qitem)
988 989
{
    int matched = 0;
990
    int index = 0;
991 992
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;
993

994 995
    qlist_for_each(ent, &ns->pending_recvs_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
996 997
        if(//(qi->num_bytes == qitem->num_bytes)
                //&& 
998
               ((qi->tag == qitem->tag) || qi->tag == -1)
999
                && ((qi->source_rank == qitem->source_rank) || qi->source_rank == -1))
1000 1001
        {
            matched = 1;
1002
            qi->num_bytes = qitem->num_bytes;
1003 1004
            break;
        }
1005
        ++index;
1006
    }
1007

1008 1009
    if(matched)
    {
1010 1011 1012 1013 1014 1015 1016
        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
1017 1018
             * the data * (only works in sequential mode)*/
            bf->c10 = 1;
1019 1020
            send_ack_back(ns, bf, m, lp, qitem);
        }
1021
        m->rc.saved_recv_time = ns->recv_time;
1022
        ns->recv_time += (tw_now(lp) - m->fwd.sim_start_time);
1023

1024 1025
        if(qi->op_type == CODES_WK_IRECV)
            update_completed_queue(ns, bf, m, lp, qi->req_id);
1026 1027
        else if(qi->op_type == CODES_WK_RECV)
            codes_issue_next_event(lp);
1028

1029
        qlist_del(&qi->ql);
1030

1031
        rc_stack_push(lp, qi, free, ns->processed_ops);
1032
        return index;
1033 1034 1035 1036
    }
    return -1;
}

1037
static int rm_matching_send(nw_state * ns,
1038 1039 1040
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp, mpi_msgs_queue * qitem)
1041 1042 1043 1044 1045
{
    int matched = 0;
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;

1046
    int index = 0;
1047 1048
    qlist_for_each(ent, &ns->arrival_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
1049 1050 1051
        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)
1052 1053
                && ((qi->source_rank == qitem->source_rank) || qitem->source_rank == -1))
        {
1054
            qitem->num_bytes = qi->num_bytes;
1055 1056 1057
            matched = 1;
            break;
        }
1058
        ++index;
1059 1060 1061 1062
    }

    if(matched)
    {
1063 1064 1065 1066 1067 1068 1069
        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 */
1070
            bf->c10 = 1;
1071 1072
            send_ack_back(ns, bf, m, lp, qi);
        }
Misbah Mubarak's avatar
Misbah Mubarak committed
1073 1074
	    rc_stack_push(lp, qi, free, ns->processed_ops);

1075
        m->rc.saved_recv_time = ns->recv_time;
1076 1077
        ns->recv_time += (tw_now(lp) - qitem->req_init_time);

1078
//        printf("\n Completed req id %d ", qitem->req_id);
1079 1080 1081
        if(qitem->op_type == CODES_WK_IRECV)
            update_completed_queue(ns, bf, m, lp, qitem->req_id);

1082
        qlist_del(&qi->ql);
1083

1084
        return index;
1085 1086 1087 1088 1089
    }
    return -1;
}
static void codes_issue_next_event_rc(tw_lp * lp)
{
1090
	    tw_rand_reverse_unif(lp->rng);
1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101
}

/* 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);
1102
   assert(ts > 0);
1103 1104 1105 1106 1107 1108 1109 1110 1111
   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(
1112
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
1113 1114 1115 1116 1117
{
	tw_event* e;
	tw_stime ts;
	nw_message* msg;

1118
    m->rc.saved_delay = s->compute_time;
1119 1120
    s->compute_time += s_to_ns(mpi_op->u.delay.seconds);
    ts = s_to_ns(mpi_op->u.delay.seconds);
1121 1122

	ts += g_tw_lookahead + 0.1 + tw_rand_exponential(lp->rng, noise);
1123
    assert(ts > 0);
1124

1125 1126 1127
	e = tw_event_new( lp->gid, ts , lp );
	msg = tw_event_data(e);
	msg->msg_type = MPI_OP_GET_NEXT;
1128 1129
	tw_event_send(e);

1130 1131 1132
}

/* reverse computation operation for MPI irecv */
1133
static void codes_exec_mpi_recv_rc(
1134 1135 1136
        nw_state* ns,
        tw_bf * bf,
        nw_message* m,
1137
        tw_lp* lp)
1138
{
1139
	ns->recv_time = m->rc.saved_recv_time;
1140 1141 1142 1143

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

1144
	if(m->fwd.found_match >= 0)
1145
	  {
1146
		ns->recv_time = m->rc.saved_recv_time;
1147 1148 1149 1150
        int queue_count = qlist_count(&ns->arrival_queue);

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

1151
        if(!m->fwd.found_match)
1152 1153 1154
        {
            qlist_add(&qi->ql, &ns->arrival_queue);
        }
1155
        else if(m->fwd.found_match >= queue_count)
1156 1157 1158
        {
            qlist_add_tail(&qi->ql, &ns->arrival_queue);
        }
1159
        else if(m->fwd.found_match > 0 && m->fwd.found_match < queue_count)
1160
        {
1161 1162 1163 1164
            int index = 1;
            struct qlist_head * ent = NULL;
            qlist_for_each(ent, &ns->arrival_queue)
            {
1165
               if(index == m->fwd.found_match)
1166 1167 1168 1169
               {
                 qlist_add(&qi->ql, ent);
                 break;
               }
1170
               index++;
1171
            }
1172
        }
1173
        if(m->op_type == CODES_WK_IRECV)
1174
        {
1175
            update_completed_queue_rc(ns,