model-net-mpi-replay.c 74.8 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/*
 * Copyright (C) 2014 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
 */
#include <ross.h>
#include <inttypes.h>

#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"
16
#include "codes/quickhash.h"
17
#include "codes/codes-jobmap.h"
18

19
/* turning on track lp will generate a lot of output messages */
20
#define MN_LP_NM "modelnet_dragonfly_custom"
21
#define CONTROL_MSG_SZ 64
22
#define TRACE -1
23
#define MAX_WAIT_REQS 512
24
#define CS_LP_DBG 0
25
#define EAGER_THRESHOLD 81920000
26 27 28
#define RANK_HASH_TABLE_SZ 2000
#define NOISE 3.0
#define NW_LP_NM "nw-lp"
29 30 31
#define lprintf(_fmt, ...) \
        do {if (CS_LP_DBG) printf(_fmt, __VA_ARGS__);} while (0)
#define MAX_STATS 65536
32
#define PAYLOAD_SZ 1024
33

34 35 36
static int msg_size_hash_compare(
            void *key, struct qhash_head *link);

37
int enable_msg_tracking = 0;
38
tw_lpid TRACK_LP = -1;
39 40

int unmatched = 0;
41 42 43 44 45
char workload_type[128];
char workload_file[8192];
char offset_file[8192];
static int wrkld_id;
static int num_net_traces = 0;
46 47
static int num_dumpi_traces = 0;

48
static int alloc_spec = 0;
49 50
static tw_stime self_overhead = 10.0;
static tw_stime mean_interval = 100000;
51 52 53 54 55 56 57

/* Doing LP IO*/
static char lp_io_dir[256] = {'\0'};
static lp_io_handle io_handle;
static unsigned int lp_io_use_suffix = 0;
static int do_lp_io = 0;

58 59 60 61
/* variables for loading multiple applications */
char workloads_conf_file[8192];
char alloc_file[8192];
int num_traces_of_job[5];
62 63 64
tw_stime soft_delay_mpi = 2500;
tw_stime nic_delay = 1000;
tw_stime copy_per_byte_eager = 0.55;
65 66 67 68 69
char file_name_of_job[5][8192];

struct codes_jobmap_ctx *jobmap_ctx;
struct codes_jobmap_params_list jobmap_p;

70 71
/* Variables for Cortex Support */
/* Matthieu's additions start */
72
#ifdef ENABLE_CORTEX_PYTHON
73 74 75
static char cortex_file[512] = "\0";
static char cortex_class[512] = "\0";
static char cortex_gen[512] = "\0";
76
#endif
77 78
/* Matthieu's additions end */

79 80
typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
81
typedef int32_t dumpi_req_id;
82 83 84

static int net_id = 0;
static float noise = 5.0;
85 86 87
static int num_nw_lps = 0, num_mpi_lps = 0;

static int num_syn_clients;
88

89
FILE * workload_log = NULL;
90
FILE * msg_size_log = NULL;
91 92 93 94
FILE * workload_agg_log = NULL;
FILE * workload_meta_log = NULL;

static uint64_t sample_bytes_written = 0;
95

96 97 98
long long num_bytes_sent=0;
long long num_bytes_recvd=0;

99 100 101
long long num_syn_bytes_sent = 0;
long long num_syn_bytes_recvd = 0;

102 103 104 105 106 107
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;
108 109 110
static int enable_sampling = 0;
static double sampling_interval = 5000000;
static double sampling_end_time = 3000000000;
111
static int enable_debug = 0;
112

113 114 115
/* set group context */
struct codes_mctx group_ratio;

116
/* MPI_OP_GET_NEXT is for getting next MPI operation when the previous operation completes.
117
* 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.
118 119 120 121 122 123 124
* 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,
125 126 127 128 129 130
    MPI_REND_ARRIVED,
    MPI_REND_ACK_ARRIVED,
    CLI_BCKGND_FIN,
    CLI_BCKGND_ARRIVE,
    CLI_BCKGND_GEN,
    CLI_NBR_FINISH,
131 132
};

133 134 135 136
struct mpi_workload_sample
{
    /* Sampling data */
    int nw_id;
137
    int app_id;
138 139 140 141 142
    unsigned long num_sends_sample;
    unsigned long num_bytes_sample;
    unsigned long num_waits_sample;
    double sample_end_time;
};
143 144 145 146 147 148 149
/* 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;
150
    uint64_t num_bytes;
151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166
    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
{
	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;
167
    int32_t req_ids[MAX_WAIT_REQS];
168
	int num_completed;
169 170
	int count;
    tw_stime start_time;
171 172 173
    struct qlist_head ql;
};

174 175 176 177 178 179 180 181 182
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; 
};
183 184 185 186 187 188 189 190 191 192
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;
193 194
    int app_id;
    int local_rank;
195

196 197 198
    int is_finished;
    int neighbor_completed;

199
    struct rc_stack * processed_ops;
200
    struct rc_stack * matched_reqs;
201 202 203 204 205 206 207 208 209 210

    /* 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;

211

212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228
	/* time spent by the LP in executing the app trace*/
	double start_time;
	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;
229

230 231
    tw_stime cur_interval_end;

232 233
    /* Pending wait operation */
    struct pending_waits * wait_op;
234

235 236 237 238 239 240
    /* Message size latency information */
    struct qhash_table * msg_sz_table;
    struct qlist_head msg_sz_list;

    /* quick hash for maintaining message latencies */

241 242 243
    unsigned long num_bytes_sent;
    unsigned long num_bytes_recvd;

244 245 246
    unsigned long syn_data;
    unsigned long gen_data;
    
247 248 249 250
    /* For sampling data */
    int sampling_indx;
    int max_arr_size;
    struct mpi_workload_sample * mpi_wkld_samples;
251
    char output_buf[512];
252 253 254 255
};

/* data for handling reverse computation.
* saved_matched_req holds the request ID of matched receives/sends for wait operations.
256
* 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.
257 258 259
* 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
{
260
   // forward message handler
261
   int msg_type;
262
   int op_type;
263
   model_net_event_return event_rc;
264

265 266 267
   struct
   {
       tw_lpid src_rank;
268
       int dest_rank;
269
       int64_t num_bytes;
270 271 272 273 274
       int num_matched;
       int data_type;
       double sim_start_time;
       // for callbacks - time message was received
       double msg_send_time;
275
       int16_t req_id;
276
       int tag;
277
       int app_id;
278 279 280 281 282 283 284 285 286
       int found_match;
       short wait_completed;
   } fwd;
   struct
   {
       double saved_send_time;
       double saved_recv_time;
       double saved_wait_time;
       double saved_delay;
287
       int16_t saved_num_bytes;
288
       struct codes_workload_op * saved_op;
289
   } rc;
290 291
};

292
static void send_ack_back(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, mpi_msgs_queue * mpi_op);
293 294

static void send_ack_back_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
295 296
/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
297
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op, int is_rend);
298 299
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
300
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
301 302
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
303
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp);
304 305
/* execute the computational delay */
static void codes_exec_comp_delay(
306
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
307 308 309 310 311 312 313 314 315 316 317 318 319 320 321
/* 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 */
322 323 324 325 326 327 328 329
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);
330 331 332 333 334 335 336 337 338 339 340 341 342 343 344
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);

345 346 347 348 349 350
static void update_message_size_rc(
        struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
351 352 353 354 355
/*TODO: Complete reverse handler */
    (void)ns;
    (void)lp;
    (void)bf;
    (void)m;
356
}
357 358 359 360 361 362 363 364 365 366
/* 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)
{
367 368 369
            (void)bf;
            (void)is_eager;

370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386
            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;
387
                msg_info->agg_latency = tw_now(lp) - msg_init_time;
388
                msg_info->avg_latency = msg_info->agg_latency;
389
                qhash_add(ns->msg_sz_table, &(msg_info->msg_size), msg_info->hash_link);
390 391 392 393 394 395 396
                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++;
397
                tmp->agg_latency += tw_now(lp) - msg_init_time;  
398 399 400 401 402 403 404 405 406 407
                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)
{
408 409 410
    (void)ns;
    (void)bf;
    (void)m;
411 412 413 414 415 416 417 418 419
    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)
{
420 421 422
        (void)bf;
        (void)m;

423 424 425 426 427 428 429 430 431 432 433 434 435 436 437
        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);

438
            lprintf("\n Other ranks %d ", num_other_ranks);
439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520
            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)
{
521 522 523 524
        (void)b;
        (void)msg;
        (void)lp;

525 526 527 528 529 530 531 532 533
        ns->is_finished = 0;
        return;
}
void finish_bckgnd_traffic(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
534 535
        (void)b;
        (void)msg;
536
        ns->is_finished = 1;
537
        lprintf("\n LP %llu completed sending data %lu completed at time %lf ", lp->gid, ns->gen_data, tw_now(lp));
538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631
        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 */  
    tw_stime ts = mean_interval  + tw_rand_exponential(lp->rng, NOISE); 
    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)
{
632 633 634
    (void)bf;
    (void)m;
    (void)lp;
635 636 637 638 639 640 641
//    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)
{
642 643 644
    (void)bf;
    (void)lp;

645 646 647 648 649
//    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;
}
650
/* Debugging functions, may generate unused function warning */
651
static void print_waiting_reqs(int32_t * reqs, int count)
652
{
653
    lprintf("\n Waiting reqs: %d count", count);
654 655
    int i;
    for(i = 0; i < count; i++ )
656
        lprintf(" %d ", reqs[i]);
657
}
658 659 660 661 662 663 664 665 666 667 668 669
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);
670
            printf(" \n Source %d Dest %d bytes %llu tag %d ", current->source_rank, current->dest_rank, current->num_bytes, current->tag);
671 672
       }
}
673 674 675 676 677 678 679 680
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);
681
            printf(" %d ", current->req_id);
682 683
       }
}
684
static int clear_completed_reqs(nw_state * s,
685
        tw_lp * lp,
686
        int32_t * reqs, int count)
687
{
688 689 690
    (void)s;
    (void)lp;

691
    int i, matched = 0;
692

693 694 695
    for( i = 0; i < count; i++)
    {
      struct qlist_head * ent = NULL;
696 697 698
      struct completed_requests * current = NULL;
      struct completed_requests * prev = NULL;

699 700
      qlist_for_each(ent, &s->completed_reqs)
       {
701 702 703 704 705
            current = qlist_entry(ent, completed_requests, ql);
            
            if(prev)
              rc_stack_push(lp, prev, free, s->matched_reqs);
            
706 707
            if(current->req_id == reqs[i])
            {
708
                ++matched;
709
                qlist_del(&current->ql);
710
                prev = current;
711
            }
712 713
            else
                prev = NULL;
714
       }
715 716 717

      if(prev)
          rc_stack_push(lp, prev, free, s->matched_reqs);
718
    }
719
    return matched;
720
}
721
static void add_completed_reqs(nw_state * s,
722 723
        tw_lp * lp,
        int count)
724
{
725
    (void)lp;
726 727 728
    int i;
    for( i = 0; i < count; i++)
    {
729 730
       struct completed_requests * req = rc_stack_pop(s->matched_reqs);
       qlist_add(&req->ql, &s->completed_reqs);
731 732
    }
}
733

734 735 736 737 738 739
/* 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);
}

740
static int notify_posted_wait(nw_state* s,
741
        tw_bf * bf, nw_message * m, tw_lp * lp,
742
        dumpi_req_id completed_req)
743
{
744 745
    (void)bf;

746 747
    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;
748

749
    m->fwd.wait_completed = 0;
750

751 752
    if(!wait_elem)
        return 0;
753

754
    int op_type = wait_elem->op_type;
755

756 757 758 759 760
    if(op_type == CODES_WK_WAIT &&
            (wait_elem->req_ids[0] == completed_req))
    {
            wait_completed = 1;
    }
761 762
    else if(op_type == CODES_WK_WAITALL
            || op_type == CODES_WK_WAITANY
763 764 765 766 767 768
            || op_type == CODES_WK_WAITSOME)
    {
        int i;
        for(i = 0; i < wait_elem->count; i++)
        {
            if(wait_elem->req_ids[i] == completed_req)
769
            {
770
                wait_elem->num_completed++;
771
                if(wait_elem->num_completed > wait_elem->count)
772
                    printf("\n Num completed %d count %d LP %llu ",
773 774 775
                            wait_elem->num_completed,
                            wait_elem->count,
                            lp->gid);
776 777
//                if(wait_elem->num_completed > wait_elem->count)
//                    tw_lp_suspend(lp, 1, 0);
778

779
                if(wait_elem->num_completed == wait_elem->count)
780
                {
781
                    if(enable_debug)
782
                        fprintf(workload_log, "\n(%lf) APP ID %d MPI WAITALL COMPLETED AT %llu ", tw_now(lp), s->app_id, s->nw_id);
783
                    wait_completed = 1;
784
                }
785

786
                m->fwd.wait_completed = 1;
787
            }
788
        }
789
    }
790
    return wait_completed;
791
}
792

793
/* reverse handler of MPI wait operation */
794
static void codes_exec_mpi_wait_rc(nw_state* s, tw_lp* lp)
795
{
796
    if(s->wait_op)
797
     {
798 799 800
         struct pending_waits * wait_op = s->wait_op;
         free(wait_op);
         s->wait_op = NULL;
801 802 803 804
     }
   else
    {
        codes_issue_next_event_rc(lp);
805
        completed_requests * qi = rc_stack_pop(s->processed_ops);
806
        qlist_add(&qi->ql, &s->completed_reqs);
807
    }
808
    return;
809
}
810

811
/* execute MPI wait operation */
812
static void codes_exec_mpi_wait(nw_state* s, tw_lp* lp, struct codes_workload_op * mpi_op)
813
{
814 815
    /* check in the completed receives queue if the request ID has already been completed.*/
    assert(!s->wait_op);
816
    dumpi_req_id req_id = mpi_op->u.wait.req_id;
817
    struct completed_requests* current = NULL;
818

819 820 821 822 823 824 825
    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)
        {
            qlist_del(&current->ql);
826
            rc_stack_push(lp, current, free, s->processed_ops);
827 828 829 830
            codes_issue_next_event(lp);
            return;
        }
    }
831 832 833 834 835
    /* 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;
836 837
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
838
    s->wait_op = wait_op;
839

840
    return;
841 842
}

843
static void codes_exec_mpi_wait_all_rc(
844
        nw_state* s,
845 846
        tw_bf * bf,
        nw_message * m,
847
        tw_lp* lp)
848
{
849 850 851 852 853 854 855 856 857 858 859
  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--;
    }
  }
860 861 862 863 864 865 866 867
  if(s->wait_op)
  {
      struct pending_waits * wait_op = s->wait_op;
      free(wait_op);
      s->wait_op = NULL;
  }
  else
  {
868
      add_completed_reqs(s, lp, m->fwd.num_matched);
869 870 871
      codes_issue_next_event_rc(lp);
  }
  return;
872
}
873

874
static void codes_exec_mpi_wait_all(
875
        nw_state* s,
876 877
        tw_bf * bf,
        nw_message * m,
878
        tw_lp* lp,
879
        struct codes_workload_op * mpi_op)
880
{
881
  if(enable_debug)
882
    fprintf(workload_log, "\n MPI WAITALL POSTED AT %llu ", s->nw_id);
883

884 885 886 887 888 889 890 891
  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;
892
        s->mpi_wkld_samples[indx].app_id = s->app_id;
893 894 895 896 897 898
        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)
    {
899
        struct mpi_workload_sample * tmp = calloc((MAX_STATS + s->max_arr_size), sizeof(struct mpi_workload_sample));
900 901 902 903 904 905 906 907
        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++;
  }
908
  int count = mpi_op->u.waits.count;
909 910
  /* If the count is not less than max wait reqs then stop */
  assert(count < MAX_WAIT_REQS);
911

912
  int i = 0, num_matched = 0;
913
  m->fwd.num_matched = 0;
914

915
  if(lp->gid == TRACK_LP)
916
  {
917
      printf("\n MPI Wait all posted ");
918 919
      print_waiting_reqs(mpi_op->u.waits.req_ids, count);
      print_completed_queue(&s->completed_reqs);
920
  }
921
      /* check number of completed irecvs in the completion queue */
922 923 924 925 926 927 928
  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)
       {
929
            current = qlist_entry(ent, struct completed_requests, ql);
930 931 932 933
            if(current->req_id == req_id)
                num_matched++;
       }
  }
934

935
  m->fwd.found_match = num_matched;
936 937 938 939
  if(num_matched == count)
  {
    /* No need to post a MPI Wait all then, issue next event */
      /* Remove all completed requests from the list */
940 941 942
      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);
943 944
      s->wait_op = NULL;
      codes_issue_next_event(lp);
945 946
  }
  else
947 948 949 950 951 952 953 954 955 956 957
  {
      /* 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;
958
	  wait_op->start_time = tw_now(lp);
959
      s->wait_op = wait_op;
960
  }
961 962
  return;
}
963

964 965
/* search for a matching mpi operation and remove it from the list.
 * Record the index in the list from where the element got deleted.
966
 * Index is used for inserting the element once again in the queue for reverse computation. */
967
static int rm_matching_rcv(nw_state * ns,
968
        tw_bf * bf,
969 970
        nw_message * m,
        tw_lp * lp,
971
        mpi_msgs_queue * qitem)
972 973
{
    int matched = 0;
974
    int index = 0;
975 976
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;
977

978 979
    qlist_for_each(ent, &ns->pending_recvs_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
980 981
        if(//(qi->num_bytes == qitem->num_bytes)
                //&& 
982
               ((qi->tag == qitem->tag) || qi->tag == -1)
983
                && ((qi->source_rank == qitem->source_rank) || qi->source_rank == -1))
984 985
        {
            matched = 1;
986
            //qitem->num_bytes = qi->num_bytes;
987 988
            break;
        }
989
        ++index;
990
    }
991

992 993
    if(matched)
    {
994 995 996 997 998 999 1000
        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
1001 1002
             * the data * (only works in sequential mode)*/
            bf->c10 = 1;
1003 1004
            send_ack_back(ns, bf, m, lp, qitem);
        }
1005
        m->rc.saved_recv_time = ns->recv_time;
1006
        ns->recv_time += (tw_now(lp) - m->fwd.sim_start_time);
1007

1008 1009
        if(qi->op_type == CODES_WK_IRECV)
            update_completed_queue(ns, bf, m, lp, qi->req_id);
1010 1011
        else if(qi->op_type == CODES_WK_RECV)
            codes_issue_next_event(lp);
1012

1013
        qlist_del(&qi->ql);
1014

1015
        rc_stack_push(lp, qi, free, ns->processed_ops);
1016
        return index;
1017 1018 1019 1020
    }
    return -1;
}

1021
static int rm_matching_send(nw_state * ns,
1022 1023 1024
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp, mpi_msgs_queue * qitem)
1025 1026 1027 1028 1029
{
    int matched = 0;
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;

1030
    int index = 0;
1031 1032
    qlist_for_each(ent, &ns->arrival_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
1033 1034 1035
        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)
1036 1037
                && ((qi->source_rank == qitem->source_rank) || qitem->source_rank == -1))
        {
1038
            //qi->num_bytes = qitem->num_bytes;
1039 1040 1041
            matched = 1;
            break;
        }
1042
        ++index;
1043 1044 1045 1046
    }

    if(matched)
    {
1047 1048 1049 1050 1051 1052 1053
        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 */
1054
            bf->c10 = 1;
1055 1056
            send_ack_back(ns, bf, m, lp, qi);
        }
1057
        m->rc.saved_recv_time = ns->recv_time;
1058 1059 1060 1061 1062
        ns->recv_time += (tw_now(lp) - qitem->req_init_time);

        if(qitem->op_type == CODES_WK_IRECV)
            update_completed_queue(ns, bf, m, lp, qitem->req_id);

1063
        qlist_del(&qi->ql);
1064

1065
        return index;
1066 1067 1068 1069 1070
    }
    return -1;
}
static void codes_issue_next_event_rc(tw_lp * lp)
{
1071
	    tw_rand_reverse_unif(lp->rng);
1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091
}

/* 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);
   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(
1092
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
1093 1094 1095 1096 1097
{
	tw_event* e;
	tw_stime ts;
	nw_message* msg;

1098
    m->rc.saved_delay = s->compute_time;
1099 1100
    s->compute_time += s_to_ns(mpi_op->u.delay.seconds);
    ts = s_to_ns(mpi_op->u.delay.seconds);
1101 1102

	ts += g_tw_lookahead + 0.1 + tw_rand_exponential(lp->rng, noise);
1103

1104 1105 1106
	e = tw_event_new( lp->gid, ts , lp );
	msg = tw_event_data(e);
	msg->msg_type = MPI_OP_GET_NEXT;
1107 1108
	tw_event_send(e);

1109 1110 1111
}

/* reverse computation operation for MPI irecv */
1112
static void codes_exec_mpi_recv_rc(
1113 1114 1115
        nw_state* ns,
        tw_bf * bf,
        nw_message* m,
1116
        tw_lp* lp)
1117
{
1118
	ns->recv_time = m->rc.saved_recv_time;
1119 1120 1121 1122

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

1123
	if(m->fwd.found_match >= 0)
1124
	  {
1125
		ns->recv_time = m->rc.saved_recv_time;
1126 1127 1128 1129
        int queue_count = qlist_count(&ns->arrival_queue);

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

1130
        if(!m->fwd.found_match)
1131 1132 1133
        {
            qlist_add(&qi->ql, &ns->arrival_queue);
        }
1134
        else if(m->fwd.found_match >= queue_count)
1135 1136 1137
        {
            qlist_add_tail(&qi->ql, &ns->arrival_queue);
        }
1138
        else if(m->fwd.found_match > 0 && m->fwd.found_match < queue_count)
1139
        {
1140 1141 1142 1143
            int index = 1;
            struct qlist_head * ent = NULL;
            qlist_for_each(ent, &ns->arrival_queue)
            {
1144
               if(index == m->fwd.found_match)
1145 1146 1147 1148
               {
                 qlist_add(&qi->ql, ent);
                 break;
               }
1149
               index++;
1150
            }
1151
        }
1152
        if(qi->op_type == CODES_WK_IRECV)
1153
        {
1154
            update_completed_queue_rc(ns, bf, m, lp);
1155
        }
1156 1157
        codes_issue_next_event_rc(lp);
      }
1158
	else if(m->fwd.found_match < 0)
1159
	    {
1160
            struct qlist_head * ent = qlist_pop_back(&ns->pending_recvs_queue);
1161 1162
            mpi_msgs_queue * qi = qlist_entry(ent, mpi_msgs_queue, ql);
            free(qi);
1163

1164
            if(m->op_type == CODES_WK_IRECV)
1165
                codes_issue_next_event_rc(lp);
1166 1167 1168
	    }
}

1169
/* Execute MPI Irecv operation (non-blocking receive) */
1170
static void codes_exec_mpi_recv(
1171
        nw_state* s,
1172
        tw_bf * bf,
1173 1174
        nw_message * m,
        tw_lp* lp,
1175
        struct codes_workload_op * mpi_op)
1176 1177 1178 1179 1180
{
/* Once an irecv is posted, list of completed sends is checked to find a matching isend.
   If no matching isend is found, the receive operation is queued in the pending queue of
   receive operations. */

1181
	m->rc.saved_recv_time = s->recv_time;
1182 1183
    m->rc.saved_num_bytes = mpi_op->u.recv.num_bytes;

1184 1185 1186 1187 1188 1189 1190 1191 1192
    mpi_msgs_queue * recv_op = (mpi_msgs_queue*) malloc(sizeof(mpi_msgs_queue));
    recv_op->req_init_time = tw_now(lp);
    recv_op->op_type = mpi_op->op_type;
    recv_op->source_rank = mpi_op->u.recv.source_rank;
    recv_op->dest_rank = mpi_op->u.recv.dest_rank;
    recv_op->num_bytes = mpi_op->u.recv.num_bytes;
    recv_op->tag = mpi_op->u.recv.tag;
    recv_op->req_id = mpi_op->u.recv.req_id;

1193 1194
    if(s->nw_id == (tw_lpid)TRACK_LP)
        printf("\n Receive op posted num bytes %llu source %d ", recv_op->num_bytes,
1195 1196
                recv_op->source_rank);