model-net-mpi-replay.c 77.7 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
int enable_msg_tracking = 0;
39
int is_synthetic = 0;
40
tw_lpid TRACK_LP = -1;
41 42

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

51
static int alloc_spec = 0;
52 53
static tw_stime self_overhead = 10.0;
static tw_stime mean_interval = 100000;
54 55 56

/* Doing LP IO*/
static char lp_io_dir[256] = {'\0'};
57
static char sampling_dir[32] = {'\0'};
58 59 60 61
static lp_io_handle io_handle;
static unsigned int lp_io_use_suffix = 0;
static int do_lp_io = 0;

62 63 64 65
/* variables for loading multiple applications */
char workloads_conf_file[8192];
char alloc_file[8192];
int num_traces_of_job[5];
66 67 68
tw_stime soft_delay_mpi = 2500;
tw_stime nic_delay = 1000;
tw_stime copy_per_byte_eager = 0.55;
69 70 71 72 73
char file_name_of_job[5][8192];

struct codes_jobmap_ctx *jobmap_ctx;
struct codes_jobmap_params_list jobmap_p;

74 75
/* Variables for Cortex Support */
/* Matthieu's additions start */
76
#ifdef ENABLE_CORTEX_PYTHON
77 78 79
static char cortex_file[512] = "\0";
static char cortex_class[512] = "\0";
static char cortex_gen[512] = "\0";
80
#endif
81 82
/* Matthieu's additions end */

83 84
typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
85
typedef int32_t dumpi_req_id;
86 87 88

static int net_id = 0;
static float noise = 5.0;
89 90 91
static int num_nw_lps = 0, num_mpi_lps = 0;

static int num_syn_clients;
92

93
FILE * workload_log = NULL;
94
FILE * msg_size_log = NULL;
95 96 97 98
FILE * workload_agg_log = NULL;
FILE * workload_meta_log = NULL;

static uint64_t sample_bytes_written = 0;
99

100 101 102
long long num_bytes_sent=0;
long long num_bytes_recvd=0;

103 104 105
long long num_syn_bytes_sent = 0;
long long num_syn_bytes_recvd = 0;

106 107 108 109 110 111
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;
112 113 114
static int enable_sampling = 0;
static double sampling_interval = 5000000;
static double sampling_end_time = 3000000000;
115
static int enable_debug = 0;
116

117 118 119
/* set group context */
struct codes_mctx group_ratio;

120
/* MPI_OP_GET_NEXT is for getting next MPI operation when the previous operation completes.
121
* 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.
122 123 124 125 126 127 128
* 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,
129 130 131 132 133 134
    MPI_REND_ARRIVED,
    MPI_REND_ACK_ARRIVED,
    CLI_BCKGND_FIN,
    CLI_BCKGND_ARRIVE,
    CLI_BCKGND_GEN,
    CLI_NBR_FINISH,
135 136
};

137 138 139 140
struct mpi_workload_sample
{
    /* Sampling data */
    int nw_id;
141
    int app_id;
142 143 144 145 146
    unsigned long num_sends_sample;
    unsigned long num_bytes_sample;
    unsigned long num_waits_sample;
    double sample_end_time;
};
147 148 149 150 151 152 153
/* 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;
154
    uint64_t num_bytes;
155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170
    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;
171
    int32_t req_ids[MAX_WAIT_REQS];
172
	int num_completed;
173 174
	int count;
    tw_stime start_time;
175 176 177
    struct qlist_head ql;
};

178 179 180 181 182 183 184 185 186
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; 
};
187 188 189 190 191 192 193 194 195 196
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;
197 198
    int app_id;
    int local_rank;
199

200 201 202
    int is_finished;
    int neighbor_completed;

203
    struct rc_stack * processed_ops;
204
    struct rc_stack * processed_wait_op;
205
    struct rc_stack * matched_reqs;
206 207 208 209 210 211 212 213 214 215

    /* 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;

216

217 218
	/* time spent by the LP in executing the app trace*/
	double start_time;
219 220 221 222 223 224 225 226 227

    double col_time;

    double reduce_time;
    int num_reduce;

    double all_reduce_time;
    int num_all_reduce;

228 229 230 231 232 233 234 235 236 237 238 239 240 241 242
	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;
243

244
    tw_stime cur_interval_end;
245
    
246 247
    /* Pending wait operation */
    struct pending_waits * wait_op;
248

249 250 251 252 253 254
    /* Message size latency information */
    struct qhash_table * msg_sz_table;
    struct qlist_head msg_sz_list;

    /* quick hash for maintaining message latencies */

255 256 257
    unsigned long num_bytes_sent;
    unsigned long num_bytes_recvd;

258 259 260
    unsigned long syn_data;
    unsigned long gen_data;
    
261 262 263 264
    /* For sampling data */
    int sampling_indx;
    int max_arr_size;
    struct mpi_workload_sample * mpi_wkld_samples;
265
    char output_buf[512];
266
    char col_stats[64];
267 268 269 270
};

/* data for handling reverse computation.
* saved_matched_req holds the request ID of matched receives/sends for wait operations.
271
* 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.
272 273 274
* 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
{
275
   // forward message handler
276
   int msg_type;
277
   int op_type;
278
   model_net_event_return event_rc;
279

280 281 282
   struct
   {
       tw_lpid src_rank;
283
       int dest_rank;
284
       uint64_t num_bytes;
285 286 287 288 289
       int num_matched;
       int data_type;
       double sim_start_time;
       // for callbacks - time message was received
       double msg_send_time;
290
       int32_t req_id;
291
       int tag;
292
       int app_id;
293 294 295 296 297 298 299 300 301
       int found_match;
       short wait_completed;
   } fwd;
   struct
   {
       double saved_send_time;
       double saved_recv_time;
       double saved_wait_time;
       double saved_delay;
302
       int32_t saved_num_bytes;
303
   } rc;
304 305
};

306
static void send_ack_back(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, mpi_msgs_queue * mpi_op);
307 308

static void send_ack_back_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
309 310
/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
311
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op, int is_rend);
312 313
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
314
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
315 316
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
317
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp);
318 319
/* execute the computational delay */
static void codes_exec_comp_delay(
320
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
321 322 323 324 325 326 327 328 329 330 331 332 333 334 335
/* 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 */
336 337 338 339 340 341 342 343
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);
344 345 346 347 348 349 350 351 352 353 354 355 356 357 358
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);

359
/*static void update_message_size_rc(
360 361 362 363
        struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
364
{*/
365
/*TODO: Complete reverse handler */
366
/*    (void)ns;
367 368 369
    (void)lp;
    (void)bf;
    (void)m;
370
}*/
371 372 373 374 375 376 377 378 379 380
/* 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)
{
381 382 383
            (void)bf;
            (void)is_eager;

384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400
            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;
401
                msg_info->agg_latency = tw_now(lp) - msg_init_time;
402
                msg_info->avg_latency = msg_info->agg_latency;
403
                qhash_add(ns->msg_sz_table, &(msg_info->msg_size), msg_info->hash_link);
404 405 406 407 408 409 410
                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++;
411
                tmp->agg_latency += tw_now(lp) - msg_init_time;  
412 413 414 415 416 417 418 419 420 421
                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)
{
422 423 424
    (void)ns;
    (void)bf;
    (void)m;
425 426 427 428 429 430 431 432 433
    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)
{
434 435 436
        (void)bf;
        (void)m;

437 438 439 440 441 442 443 444 445 446 447 448 449 450 451
        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);

452
            lprintf("\n Other ranks %d ", num_other_ranks);
453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534
            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)
{
535 536 537 538
        (void)b;
        (void)msg;
        (void)lp;

539 540 541 542 543 544 545 546 547
        ns->is_finished = 0;
        return;
}
void finish_bckgnd_traffic(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
548 549
        (void)b;
        (void)msg;
550
        ns->is_finished = 1;
551
        lprintf("\n LP %llu completed sending data %lu completed at time %lf ", lp->gid, ns->gen_data, tw_now(lp));
552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633
        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 */  
634
    tw_stime ts = mean_interval  + tw_rand_exponential(lp->rng, noise); 
635 636 637 638 639 640 641 642 643 644
    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)
{
645 646 647
    (void)bf;
    (void)m;
    (void)lp;
648 649 650 651 652 653 654
//    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)
{
655 656 657
    (void)bf;
    (void)lp;

658 659 660 661 662
//    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;
}
663
/* Debugging functions, may generate unused function warning */
664
static void print_waiting_reqs(int32_t * reqs, int count)
665
{
666
    lprintf("\n Waiting reqs: %d count", count);
667 668
    int i;
    for(i = 0; i < count; i++ )
669
        lprintf(" %d ", reqs[i]);
670
}
671 672 673 674 675 676 677 678 679 680 681 682
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);
683
            printf(" \n Source %d Dest %d bytes %llu tag %d ", current->source_rank, current->dest_rank, current->num_bytes, current->tag);
684 685
       }
}
686 687 688 689 690 691 692 693
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);
694
            printf(" %d ", current->req_id);
695 696
       }
}
697
static int clear_completed_reqs(nw_state * s,
698
        tw_lp * lp,
699
        int32_t * reqs, int count)
700
{
701 702 703
    (void)s;
    (void)lp;

704
    int i, matched = 0;
705

706 707 708
    for( i = 0; i < count; i++)
    {
      struct qlist_head * ent = NULL;
709 710 711
      struct completed_requests * current = NULL;
      struct completed_requests * prev = NULL;

712 713
      qlist_for_each(ent, &s->completed_reqs)
       {
714 715 716 717 718
            current = qlist_entry(ent, completed_requests, ql);
            
            if(prev)
              rc_stack_push(lp, prev, free, s->matched_reqs);
            
719 720
            if(current->req_id == reqs[i])
            {
721
                ++matched;
722
                qlist_del(&current->ql);
723
                prev = current;
724
            }
725 726
            else
                prev = NULL;
727
       }
728 729 730

      if(prev)
          rc_stack_push(lp, prev, free, s->matched_reqs);
731
    }
732
    return matched;
733
}
734
static void add_completed_reqs(nw_state * s,
735 736
        tw_lp * lp,
        int count)
737
{
738
    (void)lp;
739 740 741
    int i;
    for( i = 0; i < count; i++)
    {
742 743
       struct completed_requests * req = rc_stack_pop(s->matched_reqs);
       qlist_add(&req->ql, &s->completed_reqs);
744 745
    }
}
746

747 748 749 750 751 752
/* 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);
}

753
static int notify_posted_wait(nw_state* s,
754
        tw_bf * bf, nw_message * m, tw_lp * lp,
755
        dumpi_req_id completed_req)
756
{
757 758
    (void)bf;

759 760
    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;
761

762
    m->fwd.wait_completed = 0;
763

764 765
    if(!wait_elem)
        return 0;
766

767
    int op_type = wait_elem->op_type;
768

769 770 771 772 773
    if(op_type == CODES_WK_WAIT &&
            (wait_elem->req_ids[0] == completed_req))
    {
            wait_completed = 1;
    }
774 775
    else if(op_type == CODES_WK_WAITALL
            || op_type == CODES_WK_WAITANY
776 777 778 779 780 781
            || op_type == CODES_WK_WAITSOME)
    {
        int i;
        for(i = 0; i < wait_elem->count; i++)
        {
            if(wait_elem->req_ids[i] == completed_req)
782
            {
783
                wait_elem->num_completed++;
784
                if(wait_elem->num_completed > wait_elem->count)
785
                    printf("\n Num completed %d count %d LP %llu ",
786 787 788
                            wait_elem->num_completed,
                            wait_elem->count,
                            lp->gid);
789 790
//                if(wait_elem->num_completed > wait_elem->count)
//                    tw_lp_suspend(lp, 1, 0);
791

792
                if(wait_elem->num_completed >= wait_elem->count)
793
                {
794
                    if(enable_debug)
795
                        fprintf(workload_log, "\n(%lf) APP ID %d MPI WAITALL COMPLETED AT %llu ", tw_now(lp), s->app_id, s->nw_id);
796
                    wait_completed = 1;
797
                }
798

799
                m->fwd.wait_completed = 1;
800
            }
801
        }
802
    }
803
    return wait_completed;
804
}
805

806
/* reverse handler of MPI wait operation */
807
static void codes_exec_mpi_wait_rc(nw_state* s, tw_bf * bf, tw_lp* lp)
808
{
809
    if(bf->c2)
810
     {
811 812 813
         struct pending_waits * wait_op = s->wait_op;
         free(wait_op);
         s->wait_op = NULL;
814
     }
815
   if(bf->c1)
816 817
    {
        codes_issue_next_event_rc(lp);
818
        completed_requests * qi = rc_stack_pop(s->processed_ops);
819
        qlist_add(&qi->ql, &s->completed_reqs);
820
    }
821
    return;
822
}
823

824
/* execute MPI wait operation */
825
static void codes_exec_mpi_wait(nw_state* s, tw_bf * bf, tw_lp* lp, struct codes_workload_op * mpi_op)
826
{
827 828
    /* check in the completed receives queue if the request ID has already been completed.*/
    assert(!s->wait_op);
829
    dumpi_req_id req_id = mpi_op->u.wait.req_id;
830

831
    struct completed_requests* current = NULL;
832

833 834 835 836 837 838
    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)
        {
839
            bf->c1=1;
840
            qlist_del(&current->ql);
841
            rc_stack_push(lp, current, free, s->processed_ops);
842 843 844 845
            codes_issue_next_event(lp);
            return;
        }
    }
846 847

    bf->c2 = 1;
848 849 850 851 852
    /* 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;
853 854
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
855
    s->wait_op = wait_op;
856

857
    return;
858 859
}

860
static void codes_exec_mpi_wait_all_rc(
861
        nw_state* s,
862 863
        tw_bf * bf,
        nw_message * m,
864
        tw_lp* lp)
865
{
866 867 868 869 870 871 872 873 874 875 876
  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--;
    }
  }
877 878 879 880 881 882 883 884
  if(s->wait_op)
  {
      struct pending_waits * wait_op = s->wait_op;
      free(wait_op);
      s->wait_op = NULL;
  }
  else
  {
885
      add_completed_reqs(s, lp, m->fwd.num_matched);
886 887 888
      codes_issue_next_event_rc(lp);
  }
  return;
889
}
890

891
static void codes_exec_mpi_wait_all(
892
        nw_state* s,
893 894
        tw_bf * bf,
        nw_message * m,
895
        tw_lp* lp,
896
        struct codes_workload_op * mpi_op)
897
{
898
  if(enable_debug)
899
    fprintf(workload_log, "\n MPI WAITALL POSTED AT %llu ", s->nw_id);
900

901 902 903 904 905 906 907 908
  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;
909
        s->mpi_wkld_samples[indx].app_id = s->app_id;
910 911 912 913 914 915
        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)
    {
916
        struct mpi_workload_sample * tmp = calloc((MAX_STATS + s->max_arr_size), sizeof(struct mpi_workload_sample));
917 918 919 920 921 922 923 924
        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++;
  }
925
  int count = mpi_op->u.waits.count;
926 927
  /* If the count is not less than max wait reqs then stop */
  assert(count < MAX_WAIT_REQS);
928

929
  int i = 0, num_matched = 0;
930
  m->fwd.num_matched = 0;
931

932
  if(lp->gid == TRACK_LP)
933
  {
934
      printf("\n MPI Wait all posted ");
935 936
      print_waiting_reqs(mpi_op->u.waits.req_ids, count);
      print_completed_queue(&s->completed_reqs);
937
  }
938
      /* check number of completed irecvs in the completion queue */
939 940 941 942 943 944 945
  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)
       {
946
            current = qlist_entry(ent, struct completed_requests, ql);
947 948 949 950
            if(current->req_id == req_id)
                num_matched++;
       }
  }
951

952
  m->fwd.found_match = num_matched;
953 954 955 956
  if(num_matched == count)
  {
    /* No need to post a MPI Wait all then, issue next event */
      /* Remove all completed requests from the list */
957 958 959
      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);
960 961
      s->wait_op = NULL;
      codes_issue_next_event(lp);
962 963
  }
  else
964 965 966 967 968 969 970 971 972 973 974
  {
      /* 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;
975
	  wait_op->start_time = tw_now(lp);
976
      s->wait_op = wait_op;
977
  }
978 979
  return;
}
980

981 982
/* search for a matching mpi operation and remove it from the list.
 * Record the index in the list from where the element got deleted.
983
 * Index is used for inserting the element once again in the queue for reverse computation. */
984
static int rm_matching_rcv(nw_state * ns,
985
        tw_bf * bf,
986 987
        nw_message * m,
        tw_lp * lp,
988
        mpi_msgs_queue * qitem)
989 990
{
    int matched = 0;
991
    int index = 0;
992 993
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;
994

995 996
    qlist_for_each(ent, &ns->pending_recvs_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
997 998
        if(//(qi->num_bytes == qitem->num_bytes)
                //&& 
999
               ((qi->tag == qitem->tag) || qi->tag == -1)
1000
                && ((qi->source_rank == qitem->source_rank) || qi->source_rank == -1))
1001 1002
        {
            matched = 1;
1003
            qi->num_bytes = qitem->num_bytes;
1004 1005
            break;
        }
1006
        ++index;
1007
    }
1008

1009 1010
    if(matched)
    {
1011 1012 1013 1014 1015 1016 1017
        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
1018 1019
             * the data * (only works in sequential mode)*/
            bf->c10 = 1;
1020 1021
            send_ack_back(ns, bf, m, lp, qitem);
        }
1022
        m->rc.saved_recv_time = ns->recv_time;
1023
        ns->recv_time += (tw_now(lp) - m->fwd.sim_start_time);
1024

1025 1026
        if(qi->op_type == CODES_WK_IRECV)
            update_completed_queue(ns, bf, m, lp, qi->req_id);
1027 1028
        else if(qi->op_type == CODES_WK_RECV)
            codes_issue_next_event(lp);
1029

1030
        qlist_del(&qi->ql);
1031

1032
        rc_stack_push(lp, qi, free, ns->processed_ops);
1033
        return index;
1034 1035 1036 1037
    }
    return -1;
}

1038
static int rm_matching_send(nw_state * ns,
1039 1040 1041
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp, mpi_msgs_queue * qitem)
1042 1043 1044 1045 1046
{
    int matched = 0;
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;

1047
    int index = 0;
1048 1049
    qlist_for_each(ent, &ns->arrival_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
1050 1051 1052
        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)
1053 1054
                && ((qi->source_rank == qitem->source_rank) || qitem->source_rank == -1))
        {
1055
            qitem->num_bytes = qi->num_bytes;
1056 1057 1058
            matched = 1;
            break;
        }
1059
        ++index;
1060 1061 1062 1063
    }

    if(matched)
    {
1064 1065 1066 1067 1068 1069 1070
        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 */
1071
            bf->c10 = 1;
1072 1073
            send_ack_back(ns, bf, m, lp, qi);
        }
Misbah Mubarak's avatar
Misbah Mubarak committed
1074 1075
	    rc_stack_push(lp, qi, free, ns->processed_ops);

1076
        m->rc.saved_recv_time = ns->recv_time;
1077 1078
        ns->recv_time += (tw_now(lp) - qitem->req_init_time);

1079
//        printf("\n Completed req id %d ", qitem->req_id);
1080 1081 1082
        if(qitem->op_type == CODES_WK_IRECV)
            update_completed_queue(ns, bf, m, lp, qitem->req_id);

1083
        qlist_del(&qi->ql);
1084

1085
        return index;
1086 1087 1088 1089 1090
    }
    return -1;
}
static void codes_issue_next_event_rc(tw_lp * lp)
{
1091
	    tw_rand_reverse_unif(lp->rng);
1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102
}

/* 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);
1103
   assert(ts > 0);
1104 1105 1106 1107 1108 1109 1110 1111 1112
   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(
1113
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
1114 1115 1116 1117 1118
{
	tw_event* e;
	tw_stime ts;
	nw_message* msg;

1119
    m->rc.saved_delay = s->compute_time;
1120 1121
    s->compute_time += s_to_ns(mpi_op->u.delay.seconds);
    ts = s_to_ns(mpi_op->u.delay.seconds);
1122 1123

	ts += g_tw_lookahead + 0.1 + tw_rand_exponential(lp->rng, noise);
1124
    assert(ts > 0);
1125

1126 1127 1128
	e = tw_event_new( lp->gid, ts , lp );
	msg = tw_event_data(e);
	msg->msg_type = MPI_OP_GET_NEXT;
1129 1130
	tw_event_send(e);

1131 1132 1133
}

/* reverse computation operation for MPI irecv */
1134
static void codes_exec_mpi_recv_rc(
1135 1136 1137
        nw_state* ns,
        tw_bf * bf,
        nw_message* m,
1138
        tw_lp* lp)
1139
{
1140
	ns->recv_time = m->rc.saved_recv_time;
1141 1142 1143 1144

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

1145
	if(m->fwd.found_match >= 0)
1146
	  {
1147
		ns->recv_time = m->rc.saved_recv_time;
1148 1149 1150 1151
        int queue_count = qlist_count(&ns->arrival_queue);

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

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