model-net-mpi-replay.c 91.9 KB
Newer Older
1
/*
2
* Copyright (C) 2014 University of Chicago.
3 4 5 6 7
 * See COPYRIGHT notice in top-level directory.
 *
 */
#include <ross.h>
#include <inttypes.h>
8
#include <sys/stat.h>
9
#include <sys/resource.h>
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 25
#define MAX_SYN_SENDS 10
#define MAX_WAIT_REQS 1024
26
#define CS_LP_DBG 1
27
#define RANK_HASH_TABLE_SZ 2000
28
#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

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

36
static unsigned long perm_switch_thresh = 8388608;
37

38
/* NOTE: Message tracking works in sequential mode only! */
39
static int debug_cols = 0;
40
static int synthetic_pattern = 1;
41 42 43
/* 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;
44
static int enable_msg_tracking = 0;
45
static int is_synthetic = 0;
46
tw_lpid TRACK_LP = -1;
47
int nprocs = 0;
48
static double total_syn_data = 0;
49
static int unmatched = 0;
50
char workload_type[128];
51
char workload_name[128];
52 53 54 55
char workload_file[8192];
char offset_file[8192];
static int wrkld_id;
static int num_net_traces = 0;
56
static int num_dumpi_traces = 0;
57
static int64_t EAGER_THRESHOLD = INT_MAX;
58

59
static long num_ops = 0;
60
static int upper_threshold = 1048576;
61
static int alloc_spec = 0;
62
static tw_stime self_overhead = 10.0;
63
static tw_stime mean_interval = 100;
64
static int payload_sz = 1024;
65 66

/* Doing LP IO*/
67
static char * params = NULL;
68
static char lp_io_dir[256] = {'\0'};
69
static char sampling_dir[32] = {'\0'};
70
static char mpi_msg_dir[32] = {'\0'};
71 72 73 74
static lp_io_handle io_handle;
static unsigned int lp_io_use_suffix = 0;
static int do_lp_io = 0;

75 76 77 78
/* variables for loading multiple applications */
char workloads_conf_file[8192];
char alloc_file[8192];
int num_traces_of_job[5];
79 80 81
tw_stime soft_delay_mpi = 2500;
tw_stime nic_delay = 1000;
tw_stime copy_per_byte_eager = 0.55;
82 83 84 85 86
char file_name_of_job[5][8192];

struct codes_jobmap_ctx *jobmap_ctx;
struct codes_jobmap_params_list jobmap_p;

87 88
/* Variables for Cortex Support */
/* Matthieu's additions start */
89
#ifdef ENABLE_CORTEX_PYTHON
90 91 92
static char cortex_file[512] = "\0";
static char cortex_class[512] = "\0";
static char cortex_gen[512] = "\0";
93
#endif
94 95
/* Matthieu's additions end */

96 97
typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
98
typedef unsigned int dumpi_req_id;
99 100

static int net_id = 0;
101
static float noise = 1.0;
102 103 104
static int num_nw_lps = 0, num_mpi_lps = 0;

static int num_syn_clients;
105
static int syn_type = 0;
106

107
FILE * workload_log = NULL;
108
FILE * msg_size_log = NULL;
109 110 111 112
FILE * workload_agg_log = NULL;
FILE * workload_meta_log = NULL;

static uint64_t sample_bytes_written = 0;
113

114 115 116
long long num_bytes_sent=0;
long long num_bytes_recvd=0;

117 118 119
long long num_syn_bytes_sent = 0;
long long num_syn_bytes_recvd = 0;

120 121 122 123 124 125
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;
126 127 128
static int enable_sampling = 0;
static double sampling_interval = 5000000;
static double sampling_end_time = 3000000000;
129
static int enable_debug = 0;
130

131 132 133
/* set group context */
struct codes_mctx group_ratio;

134
/* MPI_OP_GET_NEXT is for getting next MPI operation when the previous operation completes.
135
* 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.
136 137 138 139 140 141 142
* 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,
143 144 145 146 147 148
    MPI_REND_ARRIVED,
    MPI_REND_ACK_ARRIVED,
    CLI_BCKGND_FIN,
    CLI_BCKGND_ARRIVE,
    CLI_BCKGND_GEN,
    CLI_NBR_FINISH,
149 150
};

151 152 153 154 155 156
/* type of synthetic traffic */
enum TRAFFIC
{
    UNIFORM = 1, /* sends message to a randomly selected node */
    NEAREST_NEIGHBOR = 2, /* sends message to the next node (potentially connected to the same router) */
    ALLTOALL = 3, /* sends message to all other nodes */
157 158
    STENCIL = 4, /* sends message to 4 nearby neighbors */
    PERMUTATION = 5
159
};
160 161 162 163
struct mpi_workload_sample
{
    /* Sampling data */
    int nw_id;
164
    int app_id;
165 166 167 168 169
    unsigned long num_sends_sample;
    unsigned long num_bytes_sample;
    unsigned long num_waits_sample;
    double sample_end_time;
};
170 171 172 173 174 175 176
/* 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;
177
    int64_t num_bytes;
178
    int64_t seq_id;
179 180 181 182 183 184 185 186
    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
{
187
	unsigned int req_id;
188
    struct qlist_head ql;
189
    int index;
190 191 192 193 194 195
};

/* for wait operations, store the pending operation and number of completed waits so far. */
struct pending_waits
{
    int op_type;
196
    unsigned int req_ids[MAX_WAIT_REQS];
197
	int num_completed;
198 199
	int count;
    tw_stime start_time;
200 201 202
    struct qlist_head ql;
};

203 204 205 206 207 208
struct msg_size_info
{
    int64_t msg_size;
    int num_msgs;
    tw_stime agg_latency;
    tw_stime avg_latency;
209
    struct qhash_head  hash_link;
210 211
    struct qlist_head ql; 
};
212 213 214 215 216 217 218 219 220 221
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;
222 223
    int app_id;
    int local_rank;
224

225
    int synthetic_pattern;
226 227 228
    int is_finished;
    int neighbor_completed;

229
    struct rc_stack * processed_ops;
230
    struct rc_stack * processed_wait_op;
231
    struct rc_stack * matched_reqs;
232
//    struct rc_stack * indices;
233 234 235 236 237 238 239 240 241 242

    /* 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;

243

244 245
	/* time spent by the LP in executing the app trace*/
	double start_time;
246 247 248 249 250 251 252 253 254

    double col_time;

    double reduce_time;
    int num_reduce;

    double all_reduce_time;
    int num_all_reduce;

255 256 257 258 259
	double elapsed_time;
	/* time spent in compute operations */
	double compute_time;
	/* time spent in message send/isend */
	double send_time;
260 261
    /* max time for synthetic traffic message */
    double max_time;
262 263 264 265 266 267 268 269 270 271
	/* 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;
272

273
    tw_stime cur_interval_end;
274
    
275 276
    /* Pending wait operation */
    struct pending_waits * wait_op;
277

278 279 280 281 282 283
    /* Message size latency information */
    struct qhash_table * msg_sz_table;
    struct qlist_head msg_sz_list;

    /* quick hash for maintaining message latencies */

284 285 286
    unsigned long num_bytes_sent;
    unsigned long num_bytes_recvd;

287 288
    unsigned long syn_data;
    unsigned long gen_data;
289 290
  
    unsigned long prev_switch;
291
    int saved_perm_dest;
292
    unsigned long rc_perm;
293

294 295 296 297
    /* For sampling data */
    int sampling_indx;
    int max_arr_size;
    struct mpi_workload_sample * mpi_wkld_samples;
298
    char output_buf[512];
299
    char col_stats[64];
300 301 302 303
};

/* data for handling reverse computation.
* saved_matched_req holds the request ID of matched receives/sends for wait operations.
304
* 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.
305 306 307
* 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
{
308
   // forward message handler
309
   int msg_type;
310
   int op_type;
311
   model_net_event_return event_rc;
312
   struct codes_workload_op * mpi_op;
313

314 315 316
   struct
   {
       tw_lpid src_rank;
317
       int dest_rank;
318
       int64_t num_bytes;
319 320 321 322 323
       int num_matched;
       int data_type;
       double sim_start_time;
       // for callbacks - time message was received
       double msg_send_time;
324
       unsigned int req_id;
325
       int matched_req;
326
       int tag;
327
       int app_id;
328 329
       int found_match;
       short wait_completed;
330
       short rend_send;
331 332 333
   } fwd;
   struct
   {
334
       int saved_perm;
335 336 337 338
       double saved_send_time;
       double saved_recv_time;
       double saved_wait_time;
       double saved_delay;
339
       int64_t saved_num_bytes;
340
       int saved_syn_length;
341
       unsigned long saved_prev_switch;
342
   } rc;
343 344
};

345
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);
346 347

static void send_ack_back_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
348 349
/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
350
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op, int is_rend);
351 352
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
353
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
354 355
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
356
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp);
357 358
/* execute the computational delay */
static void codes_exec_comp_delay(
359
        nw_state* s, tw_bf *bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
360 361 362 363 364 365 366 367 368 369 370 371 372 373 374
/* 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 */
375 376 377 378 379 380 381 382
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);
383 384 385 386 387 388 389 390 391 392 393 394 395 396 397
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);

398
/*static void update_message_size_rc(
399 400 401 402
        struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
403
{*/
404
/*TODO: Complete reverse handler */
405
/*    (void)ns;
406 407 408
    (void)lp;
    (void)bf;
    (void)m;
409
}*/
410 411 412 413 414 415 416 417 418 419
/* 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)
{
420 421 422
            (void)bf;
            (void)is_eager;

423 424 425
            struct qhash_head * hash_link = NULL;
            tw_stime msg_init_time = qitem->req_init_time;
        
426
            if(ns->msg_sz_table == NULL)
427 428 429 430 431 432 433 434 435 436
                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)
            {
437
                struct msg_size_info * msg_info = (struct msg_size_info*)malloc(sizeof(struct msg_size_info));
438 439
                msg_info->msg_size = qitem->num_bytes;
                msg_info->num_msgs = 1;
440
                msg_info->agg_latency = tw_now(lp) - msg_init_time;
441
                msg_info->avg_latency = msg_info->agg_latency;
442 443
                assert(ns->msg_sz_table);
                qhash_add(ns->msg_sz_table, &(msg_info->msg_size), &(msg_info->hash_link));
444
                qlist_add(&msg_info->ql, &ns->msg_sz_list);
445
                //printf("\n Msg size %d aggregate latency %f num messages %d ", m->fwd.num_bytes, msg_info->agg_latency, msg_info->num_msgs);
446 447 448 449 450
            }
            else
            {
                struct msg_size_info * tmp = qhash_entry(hash_link, struct msg_size_info, hash_link);
                tmp->num_msgs++;
451
                tmp->agg_latency += tw_now(lp) - msg_init_time;  
452
                tmp->avg_latency = (tmp->agg_latency / tmp->num_msgs);
453
//                printf("\n Msg size %lld aggregate latency %f num messages %d ", qitem->num_bytes, tmp->agg_latency, tmp->num_msgs);
454 455 456 457 458 459 460 461
            }
}
static void notify_background_traffic_rc(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
462 463 464
    (void)ns;
    (void)bf;
    (void)m;
465 466 467 468 469
        
    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); 
470 471 472 473 474 475 476 477
}

static void notify_background_traffic(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
478 479 480
        (void)bf;
        (void)m;

481 482 483 484 485 486 487 488 489 490 491 492 493 494 495
        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);

496
            lprintf("\n Other ranks %d ", num_other_ranks);
497
            tw_stime ts = (1.1 * g_tw_lookahead) + tw_rand_exponential(lp->rng, noise);
498 499 500 501 502 503 504 505 506 507 508
            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);
509
                m_new = (struct nw_message*)tw_event_data(e);
510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542
                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)
    {
543
//        printf("\n All workloads completed, notifying background traffic ");
544 545 546 547 548 549 550 551 552 553 554 555 556
        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;

557
//        printf("\n Local rank %d notifying neighbor %d ", ns->local_rank, ns->local_rank+1);
558
        tw_stime ts = (1.1 * g_tw_lookahead) + tw_rand_exponential(lp->rng, noise);
559 560 561 562 563 564 565 566 567
        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);
568
        m_new = (struct nw_message*)tw_event_data(e); 
569 570 571 572 573 574 575 576 577 578
        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)
{
579 580 581 582
        (void)b;
        (void)msg;
        (void)lp;

583 584 585 586 587 588 589 590 591
        ns->is_finished = 0;
        return;
}
void finish_bckgnd_traffic(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
592 593
        (void)b;
        (void)msg;
594
        ns->is_finished = 1;
595
        lprintf("\n LP %llu completed sending data %lu completed at time %lf ", LLU(lp->gid), ns->gen_data, tw_now(lp));
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
        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;

626 627 628 629 630 631
    if(bf->c1)
    {
        tw_rand_reverse_unif(lp->rng);
    }
    if(bf->c2)
    {
632
        s->prev_switch = m->rc.saved_prev_switch;
633
        s->saved_perm_dest = m->rc.saved_perm;
634 635
        tw_rand_reverse_unif(lp->rng);
    }
636 637 638 639 640
    int i;
    for (i=0; i < m->rc.saved_syn_length; i++){
        model_net_event_rc2(lp, &m->event_rc);
        s->gen_data -= payload_sz;
        num_syn_bytes_sent -= payload_sz;
641
        s->num_bytes_sent -= payload_sz;
642 643
    }
        tw_rand_reverse_unif(lp->rng);
644
        s->num_sends--;
645

646 647
     if(bf->c5)
         s->is_finished = 0;
648 649 650 651 652 653 654 655 656 657 658 659 660
}

/* 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;
661 662
    int intm_dest_id;
    nw_message remote_m;
663 664 665 666 667 668

    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);

669 670 671 672
    /* Find destination */
    int* dest_svr = NULL; 
    int i, length=0;
    switch(s->synthetic_pattern)
673
    {
674 675
        case UNIFORM:
        {
676
            bf->c1 = 1;
677 678 679 680 681 682 683
            length = 1;
            dest_svr = (int*) calloc(1, sizeof(int));
            dest_svr[0] = tw_rand_integer(lp->rng, 0, num_clients - 1);
            if(dest_svr[0] == s->local_rank)
                dest_svr[0] = (s->local_rank + 1) % num_clients;
        }
        break;
684 685 686

        case PERMUTATION:
        {
687 688
            m->rc.saved_prev_switch = s->prev_switch; //for reverse computation

689 690
            length = 1;
            dest_svr = (int*) calloc(1, sizeof(int));
691
            if(s->gen_data - s->prev_switch >= perm_switch_thresh)
692
            {
693
                // printf("%d - %d >= %d\n",s->gen_data,s->prev_switch,perm_switch_thresh);
694
                bf->c2 = 1;
695
                s->prev_switch = s->gen_data; //Amount of data pushed at time when switch initiated
696 697
                dest_svr[0] = tw_rand_integer(lp->rng, 0, num_clients - 1);
                if(dest_svr[0] == s->local_rank)
698
                    dest_svr[0] = (s->local_rank + num_clients/2) % num_clients;
699
                /* TODO: Fix random number generation code */
700 701
                m->rc.saved_perm = s->saved_perm_dest;
                s->saved_perm_dest = dest_svr[0];
702
                assert(s->saved_perm_dest != s->local_rank);
703 704
            }
            else
705
                dest_svr[0] = s->saved_perm_dest;
706 707

            assert(dest_svr[0] != s->local_rank);
708 709
        }
        break;
710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778
        case NEAREST_NEIGHBOR:
        {
            length = 1;
            dest_svr = (int*) calloc(1, sizeof(int));
            dest_svr[0] = (s->local_rank + 1) % num_clients;
        }
        break;
        case ALLTOALL:
        {
            dest_svr = (int*) calloc(num_clients-1, sizeof(int));
            int index = 0;
            for (i=0;i<num_clients;i++)
            {
                if(i!=s->local_rank) 
                {
                    dest_svr[index] = i;
                    index++;
                    length++;
                }
            }
        }
        break;
        case STENCIL:  //2D 4-point stencil
        {
            /* I think this code snippet is coming from the LLNL stencil patterns. */
            int digits, x=1, y=1, row, col, temp=num_clients;
            length = 4;
            dest_svr = (int*) calloc(4, sizeof(int));
            for (digits = 0; temp > 0; temp >>= 1)
                digits++;
            digits = digits/2;
            for (i = 0; i < digits; i++)
                x = x * 2;
            y = num_clients / x;
            //printf("\nStencil Syn: x=%d, y=%d", x, y);
            row = s->local_rank / y;
            col = s->local_rank % y;

            dest_svr[0] = row * y + ((col-1+y)%y);   /* left neighbor */
            dest_svr[1] = row * y + ((col+1+y)%y);   /* right neighbor */
            dest_svr[2] = ((row-1+x)%x) * y + col;   /* bottom neighbor */
            dest_svr[3] = ((row+1+x)%x) * y + col;   /* up neighbor */
        }
        break;
        default:
            tw_error(TW_LOC, "Undefined traffic pattern");
    }   
    /* Record length for reverse handler*/
    m->rc.saved_syn_length = length;

    if(length > 0)
    {
        // m->event_array_rc = (model_net_event_return) malloc(length * sizeof(model_net_event_return));
        //printf("\nRANK %d Dests %d", s->local_rank, length);
        for (i = 0; i < length; i++)
        {
            /* Generate synthetic traffic */
            jid.rank = dest_svr[i];
            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);

            remote_m.fwd.sim_start_time = tw_now(lp);
            remote_m.fwd.dest_rank = dest_svr[i];
            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;

            // printf("\nAPP %d SRC %d Dest %d (twid %llu)", jid.job, s->local_rank, dest_svr[i], global_dest_id);
779
            m->event_rc = model_net_event(net_id, "high", global_dest_id, payload_sz, 0.0, 
780 781 782 783
                    sizeof(nw_message), (const void*)&remote_m, 
                    0, NULL, lp);
            
            s->gen_data += payload_sz;
784
            s->num_bytes_sent += payload_sz;
785 786
            num_syn_bytes_sent += payload_sz; 
        }
787
    }
788 789
    s->num_sends++;

790
    /* New event after MEAN_INTERVAL */  
791
    tw_stime ts = mean_interval  + tw_rand_exponential(lp->rng, noise); 
792 793 794
    tw_event * e;
    nw_message * m_new;
    e = tw_event_new(lp->gid, ts, lp);
795
    m_new = (struct nw_message*)tw_event_data(e);
796 797
    m_new->msg_type = CLI_BCKGND_GEN;
    tw_event_send(e);
798 799
    
    if(s->num_sends == MAX_SYN_SENDS)
800 801
    {
        bf->c5 = 1;
802
        s->is_finished = 1;
803
    }
804
    free(dest_svr);
805 806 807 808
}

void arrive_syn_tr_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
809 810 811
    (void)bf;
    (void)m;
    (void)lp;
812
//    printf("\n Data arrived %d total data %ld ", m->fwd.num_bytes, s->syn_data);
813
    s->num_recvs--;
814 815 816
    int data = m->fwd.num_bytes;
    s->syn_data -= data;
    num_syn_bytes_recvd -= data;
817
    s->num_bytes_recvd -= data;
818
    s->send_time = m->rc.saved_send_time;
819 820 821
}
void arrive_syn_tr(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
822 823 824
    (void)bf;
    (void)lp;

825 826 827
//    printf("\n Data arrived %d total data %ld ", m->fwd.num_bytes, s->syn_data);
    if(s->local_rank == 0)
     {
828
//    	printf("\n Data arrived %lld rank %llu total data %ld ", m->fwd.num_bytes, s->nw_id, s->syn_data);
829
/*	if(s->syn_data > upper_threshold)
830 831 832 833
    if(s->local_rank == 0)
     {
    	printf("\n Data arrived %lld rank %llu total data %ld ", m->fwd.num_bytes, s->nw_id, s->syn_data);
	if(s->syn_data > upper_threshold)
834 835 836 837 838 839 840
	{ 
        	struct rusage mem_usage;
		int who = RUSAGE_SELF;
		int err = getrusage(who, &mem_usage);
		printf("\n Memory usage %lf gigabytes", ((double)mem_usage.ru_maxrss / (1024.0 * 1024.0)));
		upper_threshold += 1048576;
	}*/
841
	}
842
    m->rc.saved_send_time = s->send_time;
843
    if((tw_now(lp) - m->fwd.sim_start_time) > s->max_time)
844 845
        s->max_time = tw_now(lp) - m->fwd.sim_start_time;

846 847
    s->send_time += (tw_now(lp) - m->fwd.sim_start_time);
    s->num_recvs++;
848 849
    int data = m->fwd.num_bytes;
    s->syn_data += data;
850
    s->num_bytes_recvd += data;
851 852
    num_syn_bytes_recvd += data;
}
853
/* Debugging functions, may generate unused function warning */
854
/*static void print_waiting_reqs(uint32_t * reqs, int count)
855
{
856
    lprintf("\n Waiting reqs: %d count", count);
857 858
    int i;
    for(i = 0; i < count; i++ )
859
        lprintf(" %d ", reqs[i]);
860
}*/
861 862 863 864 865 866 867 868 869 870 871 872
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);
873
            //printf(" \n Source %d Dest %d bytes %"PRId64" tag %d ", current->source_rank, current->dest_rank, current->num_bytes, current->tag);
874 875
       }
}
876
static void print_completed_queue(tw_lp * lp, struct qlist_head * head)
877
{
878
//    printf("\n Completed queue: ");
879 880
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
881
      tw_output(lp, "\n");
882 883 884
      qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, completed_requests, ql);
885
            tw_output(lp, " %llu ", current->req_id);
886
       }
887
}
888
static int clear_completed_reqs(nw_state * s,
889
        tw_lp * lp,
890
        unsigned int * reqs, int count)
891
{
892 893 894
    (void)s;
    (void)lp;

895
    int i, matched = 0;
896

897 898 899
    for( i = 0; i < count; i++)
    {
      struct qlist_head * ent = NULL;
900 901 902
      struct completed_requests * current = NULL;
      struct completed_requests * prev = NULL;

903
      int index = 0;
904 905
      qlist_for_each(ent, &s->completed_reqs)
       {
906 907
           if(prev)
           {
908
              rc_stack_push(lp, prev, free, s->matched_reqs);
909 910
              prev = NULL;
           }
911
            
912 913
           current = qlist_entry(ent, completed_requests, ql);
           current->index = index; 
914 915
            if(current->req_id == reqs[i])
            {
916
                ++matched;
917
                qlist_del(&current->ql);
918
                prev = current;
919
            }
920
            ++index;
921
       }
922 923

      if(prev)
924 925 926 927
      {
         rc_stack_push(lp, prev, free, s->matched_reqs);
         prev = NULL;
      }
928
    }
929
    return matched;
930
}
931
static void add_completed_reqs(nw_state * s,
932 933
        tw_lp * lp,
        int count)
934
{
935
    (void)lp;
936
    for(int i = 0; i < count; i++)
937
    {
938 939
       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.
940
       qlist_add(&req->ql, &s->completed_reqs);
941
    }//end for
942
}
943

944 945 946 947 948 949
/* 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);
}

950
static int notify_posted_wait(nw_state* s,
951
        tw_bf * bf, nw_message * m, tw_lp * lp,
952
        unsigned int completed_req)
953
{
954 955
    (void)bf;

956 957
    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;
958

959
    m->fwd.wait_completed = 0;
960

961 962
    if(!wait_elem)
        return 0;
963

964
    int op_type = wait_elem->op_type;
965

966 967 968
    if(op_type == CODES_WK_WAIT &&
            (wait_elem->req_ids[0] == completed_req))
    {
969
            m->fwd.wait_completed = 1;
970 971
            wait_completed = 1;
    }
972 973
    else if(op_type == CODES_WK_WAITALL
            || op_type == CODES_WK_WAITANY
974 975 976 977 978 979
            || op_type == CODES_WK_WAITSOME)
    {
        int i;
        for(i = 0; i < wait_elem->count; i++)
        {
            if(wait_elem->req_ids[i] == completed_req)
980
            {
981
                wait_elem->num_completed++;
982
                if(wait_elem->num_completed > wait_elem->count)
983
                    printf("\n Num completed %d count %d LP %llu ",
984 985
                            wait_elem->num_completed,
                            wait_elem->count,
986
                            LLU(lp->gid));
987 988
//                if(wait_elem->num_completed > wait_elem->count)
//                    tw_lp_suspend(lp, 1, 0);
989

990
                if(wait_elem->num_completed >= wait_elem->count)
991
                {
992
                    if(enable_debug)
993
                        fprintf(workload_log, "\n(%lf) APP ID %d MPI WAITALL COMPLETED AT %llu ", tw_now(lp), s->app_id, LLU(s->nw_id));
994
                    wait_completed = 1;
995
                }
996

997
                m->fwd.wait_completed = 1;
998
            }
999
        }
1000
    }
1001
    return wait_completed;
1002
}
1003

1004
/* reverse handler of MPI wait operation */
1005
static void codes_exec_mpi_wait_rc(nw_state* s, tw_bf * bf, tw_lp* lp, nw_message * m)
1006
{
1007
   if(bf->c1)
1008
    {
1009
        completed_requests * qi = (completed_requests*)rc_stack_pop(s->processed_ops);
1010 1011
        if(m->fwd.found_match == 0)
        {
1012
            qlist_add(&qi->ql, &s->completed_reqs);
1013
        }
1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026
        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++;
           }
1027
        }
1028
        codes_issue_next_event_rc(lp);
1029
        return;
1030
    }
1031 1032 1033
         struct pending_waits * wait_op = s->wait_op;
         free(wait_op);
         s->wait_op = NULL;
1034
}
1035

1036
/* execute MPI wait operation */
1037
static void codes_exec_mpi_wait(nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
1038
{
1039
    /* check in the completed receives queue if the request ID has already been completed.*/
1040 1041
                
//    printf("\n Wait posted rank id %d ", s->nw_id);
1042
    assert(!s->wait_op);
1043
    unsigned int req_id = mpi_op->u.wait.req_id;
1044

1045
    struct completed_requests* current = NULL;
1046

1047
    struct qlist_head * ent = NULL;
1048
    int index = 0;
1049 1050 1051 1052 1053
    qlist_for_each(ent, &s->completed_reqs)
    {
        current = qlist_entry(ent, completed_requests, ql);
        if(current->req_id == req_id)
        {
1054
            bf->c1=1;
1055
            qlist_del(&current->ql);
1056
            rc_stack_push(lp, current, free, s->processed_ops);
1057
            codes_issue_next_event(lp);
1058
            m->fwd.found_match = index;
1059
            if(s->nw_id == (tw_lpid)TRACK_LP)
1060 1061 1062
            {
                tw_output(lp, "\n wait matched at post %d ", req_id);
                print_completed_queue(lp, &s->completed_reqs);
1063
            }
1064 1065
            return;
        }
1066
        ++index;
1067
    }
1068

1069 1070 1071 1072 1073
    /*if(s->nw_id == (tw_lpid)TRACK_LP)
    {
        tw_output(lp, "\n wait posted %llu ", req_id);
        print_completed_queue(lp, &s->completed_reqs);
    }*/
1074
    /* If not, add the wait operation in the pending 'waits' list. */
1075
    struct pending_waits* wait_op = (struct pending_waits*)malloc(sizeof(struct pending_waits));
1076 1077 1078
    wait_op->op_type = mpi_op->op_type;
    wait_op->req_ids[0] = req_id;
    wait_op->count = 1;
1079 1080
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
1081
    s->wait_op = wait_op;
1082

1083
    return;
1084 1085
}

1086
static void codes_exec_mpi_wait_all_rc(
1087
        nw_state* s,
1088 1089
        tw_bf * bf,
        nw_message * m,
1090
        tw_lp* lp)
1091
{
1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102
  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--;
    }
  }
1103 1104 1105 1106 1107 1108 1109 1110
  if(s->wait_op)
  {
      struct pending_waits * wait_op = s->wait_op;
      free(wait_op);
      s->wait_op = NULL;
  }
  else
  {
1111
      add_completed_reqs(s, lp, m->fwd.num_matched);
1112 1113 1114
      codes_issue_next_event_rc(lp);
  }
  return;
1115
}
1116

1117
static void codes_exec_mpi_wait_all(
1118
        nw_state* s,
1119 1120
        tw_bf * bf,
        nw_message * m,
1121
        tw_lp* lp,
1122
        struct codes_workload_op * mpi_op)
1123
{
1124
  if(enable_debug)
1125
    fprintf(workload_log, "\n MPI WAITALL POSTED AT %llu ", LLU(s->nw_id));
1126

1127 1128 1129 1130 1131 1132 1133 1134
  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;
1135
        s->mpi_wkld_samples[indx].app_id = s->app_id;
1136 1137 1138 1139 1140 1141
        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)
    {
1142
        struct mpi_workload_sample * tmp = (struct mpi_workload_sample*)calloc((MAX_STATS + s->max_arr_size), sizeof(struct mpi_workload_sample));
1143 1144 1145 1146 1147 1148 1149 1150
        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++;
  }
1151
  int count = mpi_op->u.waits.count;
1152 1153
  /* If the count is not less than max wait reqs then stop */
  assert(count < MAX_WAIT_REQS);
1154

1155
  int i = 0, num_matched = 0;
1156
  m->fwd.num_matched = 0;
1157

1158
  /*if(lp->gid == TRACK_LP)
1159
  {
1160
      printf("\n MPI Wait all posted ");
1161
      print_waiting_reqs(mpi_op->u.waits.req_ids, count);
1162 1163
      print_completed_queue(lp, &s->completed_reqs);
  }*/
1164
      /* check number of completed irecvs in the completion queue */
1165 1166
  for(i = 0; i < count; i++)
  {
1167
      unsigned int req_id = mpi_op->u.waits.req_ids[i];
1168 1169 1170 1171
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
      qlist_for_each(ent, &s->completed_reqs)
       {
1172
            current = qlist_entry(ent, struct completed_requests, ql);
1173 1174 1175 1176
            if(current->req_id == req_id)
                num_matched++;
       }
  }
1177

1178
  m->fwd.found_match = num_matched;
1179 1180 1181 1182
  if(num_matched == count)
  {
    /* No need to post a MPI Wait all then, issue next event */
      /* Remove all completed requests from the list */
1183 1184 1185
      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);
1186 1187
      s->wait_op = NULL;
      codes_issue_next_event(lp);
1188 1189
  }
  else
1190 1191
  {
      /* If not, add the wait operation in the pending 'waits' list. */
1192
	  struct pending_waits* wait_op = (struct pending_waits*)malloc(sizeof(struct pending_waits));
1193 1194 1195 1196 1197 1198 1199 1200
	  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;
1201
	  wait_op->start_time = tw_now(lp);
1202
      s->wait_op = wait_op;
1203
  }
1204 1205
  return;
}
1206

1207 1208
/* search for a matching mpi operation and remove it from the list.
 * Record the index in the list from where the element got deleted.
1209
 * Index is used for inserting the element once again in the queue for reverse computation. */
1210
static int rm_matching_rcv(nw_state * ns,
1211
        tw_bf * bf,
1212 1213
        nw_message * m,
        tw_lp * lp,
1214
        mpi_msgs_queue * qitem)
1215 1216
{
    int matched = 0;
1217
    int index = 0;
1218
    int is_rend = 0;
1219 1220
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;
1221

1222 1223
    qlist_for_each(ent, &ns->pending_recvs_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
1224 1225
        if(//(qi->num_bytes == qitem->num_bytes)
                //&& 
1226
               ((qi->tag == qitem->tag) || qi->tag == -1)
1227
                && ((qi->source_rank == qitem->source_rank) || qi->source_rank == -1))
1228 1229
        {
            matched = 1;
1230
            qi->num_bytes = qitem->num_bytes;
1231 1232
            break;
        }
1233
        ++index;
1234
    }
1235

1236 1237
    if(matched)
    {
1238 1239 1240 1241 1242 1243 1244
        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