model-net-mpi-replay.c 82.3 KB
Newer Older
1 2 3 4 5 6 7
/*
 * Copyright (C) 2014 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
 */
#include <ross.h>
#include <inttypes.h>
8
#include <sys/stat.h>
9 10 11 12 13 14 15 16

#include "codes/codes-workload.h"
#include "codes/codes.h"
#include "codes/configuration.h"
#include "codes/codes_mapping.h"
#include "codes/model-net.h"
#include "codes/rc-stack.h"
#include "codes/quicklist.h"
17
#include "codes/quickhash.h"
18
#include "codes/codes-jobmap.h"
19

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

33 34 35
static int msg_size_hash_compare(
            void *key, struct qhash_head *link);

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

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

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

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

struct codes_jobmap_ctx *jobmap_ctx;
struct codes_jobmap_params_list jobmap_p;

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

87 88
typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
Misbah Mubarak's avatar
Misbah Mubarak committed
89
typedef unsigned int dumpi_req_id;
90 91

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

static int num_syn_clients;
96

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

static uint64_t sample_bytes_written = 0;
103

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

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

110 111 112 113 114 115
double max_time = 0,  max_comm_time = 0, max_wait_time = 0, max_send_time = 0, max_recv_time = 0;
double avg_time = 0, avg_comm_time = 0, avg_wait_time = 0, avg_send_time = 0, avg_recv_time = 0;


/* runtime option for disabling computation time simulation */
static int disable_delay = 0;
116 117 118
static int enable_sampling = 0;
static double sampling_interval = 5000000;
static double sampling_end_time = 3000000000;
119
static int enable_debug = 0;
120

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

124
/* MPI_OP_GET_NEXT is for getting next MPI operation when the previous operation completes.
125
* MPI_SEND_ARRIVED is issued when a MPI message arrives at its destination (the message is transported by model-net and an event is invoked when it arrives.
126 127 128 129 130 131 132
* MPI_SEND_POSTED is issued when a MPI message has left the source LP (message is transported via model-net). */
enum MPI_NW_EVENTS
{
	MPI_OP_GET_NEXT=1,
	MPI_SEND_ARRIVED,
    MPI_SEND_ARRIVED_CB, // for tracking message times on sender
	MPI_SEND_POSTED,
133 134 135 136 137 138
    MPI_REND_ARRIVED,
    MPI_REND_ACK_ARRIVED,
    CLI_BCKGND_FIN,
    CLI_BCKGND_ARRIVE,
    CLI_BCKGND_GEN,
    CLI_NBR_FINISH,
139 140
};

141 142 143 144
struct mpi_workload_sample
{
    /* Sampling data */
    int nw_id;
145
    int app_id;
146 147 148 149 150
    unsigned long num_sends_sample;
    unsigned long num_bytes_sample;
    unsigned long num_waits_sample;
    double sample_end_time;
};
151 152 153 154 155 156 157
/* stores pointers of pending MPI operations to be matched with their respective sends/receives. */
struct mpi_msgs_queue
{
    int op_type;
    int tag;
    int source_rank;
    int dest_rank;
158
    int64_t num_bytes;
159
    int64_t seq_id;
160 161 162 163 164 165 166 167
    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
{
Misbah Mubarak's avatar
Misbah Mubarak committed
168
	unsigned int req_id;
169
    struct qlist_head ql;
170
    int index;
171 172 173 174 175 176
};

/* for wait operations, store the pending operation and number of completed waits so far. */
struct pending_waits
{
    int op_type;
Misbah Mubarak's avatar
Misbah Mubarak committed
177
    unsigned int req_ids[MAX_WAIT_REQS];
178
	int num_completed;
179 180
	int count;
    tw_stime start_time;
181 182 183
    struct qlist_head ql;
};

184 185 186 187 188 189 190 191 192
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; 
};
193 194 195 196 197 198 199 200 201 202
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;
203 204
    int app_id;
    int local_rank;
205

206 207 208
    int is_finished;
    int neighbor_completed;

209
    struct rc_stack * processed_ops;
210
    struct rc_stack * processed_wait_op;
211
    struct rc_stack * matched_reqs;
212
//    struct rc_stack * indices;
213 214 215 216 217 218 219 220 221 222

    /* 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;

223

224 225
	/* time spent by the LP in executing the app trace*/
	double start_time;
226 227 228 229 230 231 232 233 234

    double col_time;

    double reduce_time;
    int num_reduce;

    double all_reduce_time;
    int num_all_reduce;

235 236 237 238 239 240 241 242 243 244 245 246 247 248 249
	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;
250

251
    tw_stime cur_interval_end;
252
    
253 254
    /* Pending wait operation */
    struct pending_waits * wait_op;
255

256 257 258 259 260 261
    /* Message size latency information */
    struct qhash_table * msg_sz_table;
    struct qlist_head msg_sz_list;

    /* quick hash for maintaining message latencies */

262 263 264
    unsigned long num_bytes_sent;
    unsigned long num_bytes_recvd;

265 266 267
    unsigned long syn_data;
    unsigned long gen_data;
    
268 269 270 271
    /* For sampling data */
    int sampling_indx;
    int max_arr_size;
    struct mpi_workload_sample * mpi_wkld_samples;
272
    char output_buf[512];
273
    char col_stats[64];
274 275 276 277
};

/* data for handling reverse computation.
* saved_matched_req holds the request ID of matched receives/sends for wait operations.
278
* 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.
279 280 281
* 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
{
282
   // forward message handler
283
   int msg_type;
284
   int op_type;
285
   model_net_event_return event_rc;
286

287 288 289
   struct
   {
       tw_lpid src_rank;
290
       int dest_rank;
291
       int64_t num_bytes;
292 293 294 295 296
       int num_matched;
       int data_type;
       double sim_start_time;
       // for callbacks - time message was received
       double msg_send_time;
Misbah Mubarak's avatar
Misbah Mubarak committed
297
       unsigned int req_id;
298
       int matched_req;
299
       int tag;
300
       int app_id;
301 302
       int found_match;
       short wait_completed;
303
       short rend_send;
304 305 306 307 308 309 310
   } fwd;
   struct
   {
       double saved_send_time;
       double saved_recv_time;
       double saved_wait_time;
       double saved_delay;
311
       int64_t saved_num_bytes;
312
   } rc;
313 314
};

315
static void send_ack_back(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, mpi_msgs_queue * mpi_op, int matched_req);
316 317

static void send_ack_back_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
318 319
/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
320
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op, int is_rend);
321 322
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
323
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
324 325
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
326
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp);
327 328
/* execute the computational delay */
static void codes_exec_comp_delay(
329
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
330 331 332 333 334 335 336 337 338 339 340 341 342 343 344
/* 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 */
345 346 347 348 349 350 351 352
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);
353 354 355 356 357 358 359 360 361 362 363 364 365 366 367
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);

368
/*static void update_message_size_rc(
369 370 371 372
        struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
373
{*/
374
/*TODO: Complete reverse handler */
375
/*    (void)ns;
376 377 378
    (void)lp;
    (void)bf;
    (void)m;
379
}*/
380 381 382 383 384 385 386 387 388 389
/* 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)
{
390 391 392
            (void)bf;
            (void)is_eager;

393 394 395 396 397 398 399 400 401 402 403 404 405 406
            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)
            {
407
                struct msg_size_info * msg_info = (struct msg_size_info*)malloc(sizeof(struct msg_size_info));
408 409
                msg_info->msg_size = qitem->num_bytes;
                msg_info->num_msgs = 1;
410
                msg_info->agg_latency = tw_now(lp) - msg_init_time;
411
                msg_info->avg_latency = msg_info->agg_latency;
412
                qhash_add(ns->msg_sz_table, &(msg_info->msg_size), msg_info->hash_link);
413 414 415 416 417 418 419
                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++;
420
                tmp->agg_latency += tw_now(lp) - msg_init_time;  
421 422 423 424 425 426 427 428 429 430
                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)
{
431 432 433
    (void)ns;
    (void)bf;
    (void)m;
434 435 436 437 438
        
    int num_jobs = codes_jobmap_get_num_jobs(jobmap_ctx); 
    
    for(int i = 0; i < num_jobs - 1; i++)
        tw_rand_reverse_unif(lp->rng); 
439 440 441 442 443 444 445 446
}

static void notify_background_traffic(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
447 448 449
        (void)bf;
        (void)m;

450 451 452 453 454 455 456 457 458 459 460 461 462 463 464
        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);

465
            lprintf("\n Other ranks %d ", num_other_ranks);
466 467 468 469 470 471 472 473 474 475 476 477
            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);
478
                m_new = (struct nw_message*)tw_event_data(e);
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
                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)
    {
512
//        printf("\n All workloads completed, notifying background traffic ");
513 514 515 516 517 518 519 520 521 522 523 524 525
        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;

526
//        printf("\n Local rank %d notifying neighbor %d ", ns->local_rank, ns->local_rank+1);
527 528 529 530 531 532 533 534 535 536
        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);
537
        m_new = (struct nw_message*)tw_event_data(e); 
538 539 540 541 542 543 544 545 546 547
        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)
{
548 549 550 551
        (void)b;
        (void)msg;
        (void)lp;

552 553 554 555 556 557 558 559 560
        ns->is_finished = 0;
        return;
}
void finish_bckgnd_traffic(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
561 562
        (void)b;
        (void)msg;
563
        ns->is_finished = 1;
564
        lprintf("\n LP %llu completed sending data %lu completed at time %lf ", LLU(lp->gid), ns->gen_data, tw_now(lp));
565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646
        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 */  
647
    tw_stime ts = mean_interval  + tw_rand_exponential(lp->rng, noise); 
648 649 650
    tw_event * e;
    nw_message * m_new;
    e = tw_event_new(lp->gid, ts, lp);
651
    m_new = (struct nw_message*)tw_event_data(e);
652 653 654 655 656 657
    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)
{
658 659 660
    (void)bf;
    (void)m;
    (void)lp;
661 662 663 664 665 666 667
//    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)
{
668 669 670
    (void)bf;
    (void)lp;

671 672 673 674 675
//    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;
}
676
/* Debugging functions, may generate unused function warning */
677
/*static void print_waiting_reqs(uint32_t * reqs, int count)
678
{
679
    lprintf("\n Waiting reqs: %d count", count);
680 681
    int i;
    for(i = 0; i < count; i++ )
682
        lprintf(" %d ", reqs[i]);
683
}*/
684 685 686 687 688 689 690 691 692 693 694 695
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);
696
            printf(" \n Source %d Dest %d bytes %"PRId64" tag %d ", current->source_rank, current->dest_rank, current->num_bytes, current->tag);
697 698
       }
}
Misbah Mubarak's avatar
Misbah Mubarak committed
699
static void print_completed_queue(tw_lp * lp, struct qlist_head * head)
700
{
701
//    printf("\n Completed queue: ");
702 703
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
704
      tw_output(lp, "\n");
705 706 707
      qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, completed_requests, ql);
708
            tw_output(lp, " %llu ", current->req_id);
709
       }
Misbah Mubarak's avatar
Misbah Mubarak committed
710
}
711
static int clear_completed_reqs(nw_state * s,
712
        tw_lp * lp,
Misbah Mubarak's avatar
Misbah Mubarak committed
713
        unsigned int * reqs, int count)
714
{
715 716 717
    (void)s;
    (void)lp;

718
    int i, matched = 0;
719

720 721 722
    for( i = 0; i < count; i++)
    {
      struct qlist_head * ent = NULL;
723 724 725
      struct completed_requests * current = NULL;
      struct completed_requests * prev = NULL;

726
      int index = 0;
727 728
      qlist_for_each(ent, &s->completed_reqs)
       {
729 730
           if(prev)
           {
731
              rc_stack_push(lp, prev, free, s->matched_reqs);
732 733
              prev = NULL;
           }
734
            
735 736
           current = qlist_entry(ent, completed_requests, ql);
           current->index = index; 
737 738
            if(current->req_id == reqs[i])
            {
739
                ++matched;
740
                qlist_del(&current->ql);
741
                prev = current;
742
            }
743
            ++index;
744
       }
745 746

      if(prev)
747 748 749 750
      {
         rc_stack_push(lp, prev, free, s->matched_reqs);
         prev = NULL;
      }
751
    }
752
    return matched;
753
}
754
static void add_completed_reqs(nw_state * s,
755 756
        tw_lp * lp,
        int count)
757
{
758
    (void)lp;
759
    for(int i = 0; i < count; i++)
760
    {
761 762
       struct completed_requests * req = (struct completed_requests*)rc_stack_pop(s->matched_reqs);
       // turn on only if wait-all unmatched error arises in optimistic mode.
763
       qlist_add(&req->ql, &s->completed_reqs);
764
    }//end for
765
}
766

767 768 769 770 771 772
/* 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);
}

773
static int notify_posted_wait(nw_state* s,
774
        tw_bf * bf, nw_message * m, tw_lp * lp,
Misbah Mubarak's avatar
Misbah Mubarak committed
775
        unsigned int completed_req)
776
{
777 778
    (void)bf;

779 780
    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;
781

782
    m->fwd.wait_completed = 0;
783

784 785
    if(!wait_elem)
        return 0;
786

787
    int op_type = wait_elem->op_type;
788

789 790 791
    if(op_type == CODES_WK_WAIT &&
            (wait_elem->req_ids[0] == completed_req))
    {
792
            m->fwd.wait_completed = 1;
793 794
            wait_completed = 1;
    }
795 796
    else if(op_type == CODES_WK_WAITALL
            || op_type == CODES_WK_WAITANY
797 798 799 800 801 802
            || op_type == CODES_WK_WAITSOME)
    {
        int i;
        for(i = 0; i < wait_elem->count; i++)
        {
            if(wait_elem->req_ids[i] == completed_req)
803
            {
804
                wait_elem->num_completed++;
805
                if(wait_elem->num_completed > wait_elem->count)
806
                    printf("\n Num completed %d count %d LP %llu ",
807 808
                            wait_elem->num_completed,
                            wait_elem->count,
809
                            LLU(lp->gid));
810 811
//                if(wait_elem->num_completed > wait_elem->count)
//                    tw_lp_suspend(lp, 1, 0);
812

813
                if(wait_elem->num_completed >= wait_elem->count)
814
                {
815
                    if(enable_debug)
816
                        fprintf(workload_log, "\n(%lf) APP ID %d MPI WAITALL COMPLETED AT %llu ", tw_now(lp), s->app_id, LLU(s->nw_id));
817
                    wait_completed = 1;
818
                }
819

820
                m->fwd.wait_completed = 1;
821
            }
822
        }
823
    }
824
    return wait_completed;
825
}
826

827
/* reverse handler of MPI wait operation */
828
static void codes_exec_mpi_wait_rc(nw_state* s, tw_bf * bf, tw_lp* lp, nw_message * m)
829
{
830
   if(bf->c1)
831
    {
832
        completed_requests * qi = (completed_requests*)rc_stack_pop(s->processed_ops);
833 834
        if(m->fwd.found_match == 0)
        {
835
            qlist_add(&qi->ql, &s->completed_reqs);
836
        }
837 838 839 840 841 842 843 844 845 846 847 848 849
        else
        {
           int index = 1;
           struct qlist_head * ent = NULL;
           qlist_for_each(ent, &s->completed_reqs)
           {
                if(index == m->fwd.found_match)
                {
                    qlist_add(&qi->ql, ent);
                    break;
                }
                index++;
           }
850
        }
851
        codes_issue_next_event_rc(lp);
852
        return;
853
    }
854 855 856
         struct pending_waits * wait_op = s->wait_op;
         free(wait_op);
         s->wait_op = NULL;
857
}
858

859
/* execute MPI wait operation */
860
static void codes_exec_mpi_wait(nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
861
{
862
    /* check in the completed receives queue if the request ID has already been completed.*/
Misbah Mubarak's avatar
Misbah Mubarak committed
863 864
                
//    printf("\n Wait posted rank id %d ", s->nw_id);
865
    assert(!s->wait_op);
Misbah Mubarak's avatar
Misbah Mubarak committed
866
    unsigned int req_id = mpi_op->u.wait.req_id;
867

868
    struct completed_requests* current = NULL;
869

870
    struct qlist_head * ent = NULL;
871
    int index = 0;
872 873 874 875 876
    qlist_for_each(ent, &s->completed_reqs)
    {
        current = qlist_entry(ent, completed_requests, ql);
        if(current->req_id == req_id)
        {
877
            bf->c1=1;
878
            qlist_del(&current->ql);
879
            rc_stack_push(lp, current, free, s->processed_ops);
880
            codes_issue_next_event(lp);
881
            m->fwd.found_match = index;
Misbah Mubarak's avatar
Misbah Mubarak committed
882
            if(s->nw_id == (tw_lpid)TRACK_LP)
883 884 885
            {
                tw_output(lp, "\n wait matched at post %d ", req_id);
                print_completed_queue(lp, &s->completed_reqs);
Misbah Mubarak's avatar
Misbah Mubarak committed
886
            }
887 888
            return;
        }
889
        ++index;
890
    }
891

892 893 894 895 896
    /*if(s->nw_id == (tw_lpid)TRACK_LP)
    {
        tw_output(lp, "\n wait posted %llu ", req_id);
        print_completed_queue(lp, &s->completed_reqs);
    }*/
897
    /* If not, add the wait operation in the pending 'waits' list. */
898
    struct pending_waits* wait_op = (struct pending_waits*)malloc(sizeof(struct pending_waits));
899 900 901
    wait_op->op_type = mpi_op->op_type;
    wait_op->req_ids[0] = req_id;
    wait_op->count = 1;
902 903
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
904
    s->wait_op = wait_op;
905

906
    return;
907 908
}

909
static void codes_exec_mpi_wait_all_rc(
910
        nw_state* s,
911 912
        tw_bf * bf,
        nw_message * m,
913
        tw_lp* lp)
914
{
915 916 917 918 919 920 921 922 923 924 925
  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--;
    }
  }
926 927 928 929 930 931 932 933
  if(s->wait_op)
  {
      struct pending_waits * wait_op = s->wait_op;
      free(wait_op);
      s->wait_op = NULL;
  }
  else
  {
934
      add_completed_reqs(s, lp, m->fwd.num_matched);
935 936 937
      codes_issue_next_event_rc(lp);
  }
  return;
938
}
939

940
static void codes_exec_mpi_wait_all(
941
        nw_state* s,
942 943
        tw_bf * bf,
        nw_message * m,
944
        tw_lp* lp,
945
        struct codes_workload_op * mpi_op)
946
{
947
  if(enable_debug)
948
    fprintf(workload_log, "\n MPI WAITALL POSTED AT %llu ", LLU(s->nw_id));
949

950 951 952 953 954 955 956 957
  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;
958
        s->mpi_wkld_samples[indx].app_id = s->app_id;
959 960 961 962 963 964
        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)
    {
965
        struct mpi_workload_sample * tmp = (struct mpi_workload_sample*)calloc((MAX_STATS + s->max_arr_size), sizeof(struct mpi_workload_sample));
966 967 968 969 970 971 972 973
        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++;
  }
974
  int count = mpi_op->u.waits.count;
975 976
  /* If the count is not less than max wait reqs then stop */
  assert(count < MAX_WAIT_REQS);
977

978
  int i = 0, num_matched = 0;
979
  m->fwd.num_matched = 0;
980

981
  /*if(lp->gid == TRACK_LP)
982
  {
983
      printf("\n MPI Wait all posted ");
984
      print_waiting_reqs(mpi_op->u.waits.req_ids, count);
985 986
      print_completed_queue(lp, &s->completed_reqs);
  }*/
987
      /* check number of completed irecvs in the completion queue */
988 989
  for(i = 0; i < count; i++)
  {
Misbah Mubarak's avatar
Misbah Mubarak committed
990
      unsigned int req_id = mpi_op->u.waits.req_ids[i];
991 992 993 994
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
      qlist_for_each(ent, &s->completed_reqs)
       {
995
            current = qlist_entry(ent, struct completed_requests, ql);
996 997 998 999
            if(current->req_id == req_id)
                num_matched++;
       }
  }
1000

1001
  m->fwd.found_match = num_matched;
1002 1003 1004 1005
  if(num_matched == count)
  {
    /* No need to post a MPI Wait all then, issue next event */
      /* Remove all completed requests from the list */
1006 1007 1008
      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);
1009 1010
      s->wait_op = NULL;
      codes_issue_next_event(lp);
1011 1012
  }
  else
1013 1014
  {
      /* If not, add the wait operation in the pending 'waits' list. */
1015
	  struct pending_waits* wait_op = (struct pending_waits*)malloc(sizeof(struct pending_waits));
1016 1017 1018 1019 1020 1021 1022 1023
	  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;
1024
	  wait_op->start_time = tw_now(lp);
1025
      s->wait_op = wait_op;
1026
  }
1027 1028
  return;
}
1029

1030 1031
/* search for a matching mpi operation and remove it from the list.
 * Record the index in the list from where the element got deleted.
1032
 * Index is used for inserting the element once again in the queue for reverse computation. */
1033
static int rm_matching_rcv(nw_state * ns,
1034
        tw_bf * bf,
1035 1036
        nw_message * m,
        tw_lp * lp,
1037
        mpi_msgs_queue * qitem)
1038 1039
{
    int matched = 0;
1040
    int index = 0;
1041
    int is_rend = 0;
1042 1043
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;
1044

1045 1046
    qlist_for_each(ent, &ns->pending_recvs_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
1047 1048
        if(//(qi->num_bytes == qitem->num_bytes)
                //&& 
1049
               ((qi->tag == qitem->tag) || qi->tag == -1)
1050
                && ((qi->source_rank == qitem->source_rank) || qi->source_rank == -1))
1051 1052
        {
            matched = 1;
1053
            qi->num_bytes = qitem->num_bytes;
1054 1055
            break;
        }
1056
        ++index;
1057
    }
1058

1059 1060
    if(matched)
    {
1061 1062 1063 1064 1065 1066 1067
        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
1068 1069
             * the data * (only works in sequential mode)*/
            bf->c10 = 1;
1070 1071
            is_rend = 1;
            send_ack_back(ns, bf, m, lp, qitem, qi->req_id);
1072
        }
1073 1074
        else
        {
1075
            bf->c12 = 1;
1076 1077 1078 1079
            m->rc.saved_recv_time = ns->recv_time;
            ns->recv_time += (tw_now(lp) - m->fwd.sim_start_time);
        }
        if(qi->op_type == CODES_WK_IRECV && !is_rend)
1080
        {
1081
            bf->c9 = 1;
1082 1083 1084 1085
            /*if(ns->nw_id == (tw_lpid)TRACK_LP)
            {
                printf("\n Completed irecv req id %d ", qi->req_id);
            }*/
1086
            update_completed_queue(ns, bf, m, lp, qi->req_id);
1087
        }
1088 1089 1090
        else if(qi->op_type == CODES_WK_RECV && !is_rend)
        {
            bf->c8 = 1;
1091
            codes_issue_next_event(lp);
1092
        }
1093

1094
        qlist_del(&qi->ql);
1095

1096
        rc_stack_push(lp, qi, free, ns->processed_ops);
1097
        return index;
1098 1099 1100 1101
    }
    return -1;
}

1102
static int rm_matching_send(nw_state * ns,
1103 1104 1105
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp, mpi_msgs_queue * qitem)
1106 1107 1108 1109 1110
{
    int matched = 0;
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;

1111
    int index = 0;