model-net-mpi-replay.c 90.1 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
#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

32 33 34
static int msg_size_hash_compare(
            void *key, struct qhash_head *link);

35
static unsigned long perm_switch_thresh = 8388608;
36

37
/* NOTE: Message tracking works in sequential mode only! */
38
static int debug_cols = 0;
39
static int synthetic_pattern = 1;
40 41 42 43 44
/* 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;
45
tw_lpid TRACK_LP = -1;
46
int nprocs = 0;
47
static double total_syn_data = 0;
48
static int unmatched = 0;
49
char workload_type[128];
50
char workload_name[128];
51 52 53 54
char workload_file[8192];
char offset_file[8192];
static int wrkld_id;
static int num_net_traces = 0;
55
static int num_dumpi_traces = 0;
56
static int64_t EAGER_THRESHOLD = 8192;
57

58
static long num_ops = 0;
59
static int upper_threshold = 1048576;
60
static int alloc_spec = 0;
61 62
static tw_stime self_overhead = 10.0;
static tw_stime mean_interval = 100000;
63
static int payload_sz = 1024;
64 65

/* Doing LP IO*/
66
static char * params = NULL;
67
static char lp_io_dir[256] = {'\0'};
68
static char sampling_dir[32] = {'\0'};
69
static char mpi_msg_dir[32] = {'\0'};
70 71 72 73
static lp_io_handle io_handle;
static unsigned int lp_io_use_suffix = 0;
static int do_lp_io = 0;

74 75 76 77
/* variables for loading multiple applications */
char workloads_conf_file[8192];
char alloc_file[8192];
int num_traces_of_job[5];
78 79 80
tw_stime soft_delay_mpi = 2500;
tw_stime nic_delay = 1000;
tw_stime copy_per_byte_eager = 0.55;
81 82 83 84 85
char file_name_of_job[5][8192];

struct codes_jobmap_ctx *jobmap_ctx;
struct codes_jobmap_params_list jobmap_p;

86 87
/* Variables for Cortex Support */
/* Matthieu's additions start */
88
#ifdef ENABLE_CORTEX_PYTHON
89 90 91
static char cortex_file[512] = "\0";
static char cortex_class[512] = "\0";
static char cortex_gen[512] = "\0";
92
#endif
93 94
/* Matthieu's additions end */

95 96
typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
97
typedef unsigned int dumpi_req_id;
98 99

static int net_id = 0;
100
static float noise = 1.0;
101 102 103
static int num_nw_lps = 0, num_mpi_lps = 0;

static int num_syn_clients;
104
static int syn_type = 0;
105

106
FILE * workload_log = NULL;
107
FILE * msg_size_log = NULL;
108 109 110 111
FILE * workload_agg_log = NULL;
FILE * workload_meta_log = NULL;

static uint64_t sample_bytes_written = 0;
112

113 114 115
long long num_bytes_sent=0;
long long num_bytes_recvd=0;

116 117 118
long long num_syn_bytes_sent = 0;
long long num_syn_bytes_recvd = 0;

119 120 121 122 123 124
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;
125 126 127
static int enable_sampling = 0;
static double sampling_interval = 5000000;
static double sampling_end_time = 3000000000;
128
static int enable_debug = 0;
129

130 131 132
/* set group context */
struct codes_mctx group_ratio;

133
/* MPI_OP_GET_NEXT is for getting next MPI operation when the previous operation completes.
134
* 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.
135 136 137 138 139 140 141
* 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,
142 143 144 145 146 147
    MPI_REND_ARRIVED,
    MPI_REND_ACK_ARRIVED,
    CLI_BCKGND_FIN,
    CLI_BCKGND_ARRIVE,
    CLI_BCKGND_GEN,
    CLI_NBR_FINISH,
148 149
};

150 151 152 153 154 155
/* 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 */
156 157
    STENCIL = 4, /* sends message to 4 nearby neighbors */
    PERMUTATION = 5
158
};
159 160 161 162
struct mpi_workload_sample
{
    /* Sampling data */
    int nw_id;
163
    int app_id;
164 165 166 167 168
    unsigned long num_sends_sample;
    unsigned long num_bytes_sample;
    unsigned long num_waits_sample;
    double sample_end_time;
};
169 170 171 172 173 174 175
/* 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;
176
    int64_t num_bytes;
177
    int64_t seq_id;
178 179 180 181 182 183 184 185
    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
{
186
	unsigned int req_id;
187
    struct qlist_head ql;
188
    int index;
189 190 191 192 193 194
};

/* for wait operations, store the pending operation and number of completed waits so far. */
struct pending_waits
{
    int op_type;
195
    unsigned int req_ids[MAX_WAIT_REQS];
196
	int num_completed;
197 198
	int count;
    tw_stime start_time;
199 200 201
    struct qlist_head ql;
};

202 203 204 205 206 207
struct msg_size_info
{
    int64_t msg_size;
    int num_msgs;
    tw_stime agg_latency;
    tw_stime avg_latency;
208
    struct qhash_head  hash_link;
209 210
    struct qlist_head ql; 
};
211 212 213 214 215 216 217 218 219 220
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;
221 222
    int app_id;
    int local_rank;
223

224
    int synthetic_pattern;
225 226 227
    int is_finished;
    int neighbor_completed;

228
    struct rc_stack * processed_ops;
229
    struct rc_stack * processed_wait_op;
230
    struct rc_stack * matched_reqs;
231
//    struct rc_stack * indices;
232 233 234 235 236 237 238 239 240 241

    /* 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;

242

243 244
	/* time spent by the LP in executing the app trace*/
	double start_time;
245 246 247 248 249 250 251 252 253

    double col_time;

    double reduce_time;
    int num_reduce;

    double all_reduce_time;
    int num_all_reduce;

254 255 256 257 258 259 260 261 262 263 264 265 266 267 268
	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;
269

270
    tw_stime cur_interval_end;
271
    
272 273
    /* Pending wait operation */
    struct pending_waits * wait_op;
274

275 276 277 278 279 280
    /* Message size latency information */
    struct qhash_table * msg_sz_table;
    struct qlist_head msg_sz_list;

    /* quick hash for maintaining message latencies */

281 282 283
    unsigned long num_bytes_sent;
    unsigned long num_bytes_recvd;

284 285
    unsigned long syn_data;
    unsigned long gen_data;
286 287 288 289 290
  
    unsigned long prev_switch;
    unsigned long saved_perm_dest;
    unsigned long rc_perm;

291 292 293 294
    /* For sampling data */
    int sampling_indx;
    int max_arr_size;
    struct mpi_workload_sample * mpi_wkld_samples;
295
    char output_buf[512];
296
    char col_stats[64];
297 298 299 300
};

/* data for handling reverse computation.
* saved_matched_req holds the request ID of matched receives/sends for wait operations.
301
* 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.
302 303 304
* 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
{
305
   // forward message handler
306
   int msg_type;
307
   int op_type;
308
   model_net_event_return event_rc;
309
   struct codes_workload_op * mpi_op;
310

311 312 313
   struct
   {
       tw_lpid src_rank;
314
       int dest_rank;
315
       int64_t num_bytes;
316 317 318 319 320
       int num_matched;
       int data_type;
       double sim_start_time;
       // for callbacks - time message was received
       double msg_send_time;
321
       unsigned int req_id;
322
       int matched_req;
323
       int tag;
324
       int app_id;
325 326
       int found_match;
       short wait_completed;
327
       short rend_send;
328 329 330 331 332 333 334
   } fwd;
   struct
   {
       double saved_send_time;
       double saved_recv_time;
       double saved_wait_time;
       double saved_delay;
335
       int64_t saved_num_bytes;
336
       int saved_syn_length;
337
       unsigned long saved_prev_switch;
338
   } rc;
339 340
};

341
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);
342 343

static void send_ack_back_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
344 345
/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
346
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op, int is_rend);
347 348
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
349
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
350 351
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
352
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp);
353 354
/* execute the computational delay */
static void codes_exec_comp_delay(
355
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
356 357 358 359 360 361 362 363 364 365 366 367 368 369 370
/* 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 */
371 372 373 374 375 376 377 378
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);
379 380 381 382 383 384 385 386 387 388 389 390 391 392 393
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);

394
/*static void update_message_size_rc(
395 396 397 398
        struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
399
{*/
400
/*TODO: Complete reverse handler */
401
/*    (void)ns;
402 403 404
    (void)lp;
    (void)bf;
    (void)m;
405
}*/
406 407 408 409 410 411 412 413 414 415
/* 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)
{
416 417 418
            (void)bf;
            (void)is_eager;

419 420 421
            struct qhash_head * hash_link = NULL;
            tw_stime msg_init_time = qitem->req_init_time;
        
422
            if(ns->msg_sz_table == NULL)
423 424 425 426 427 428 429 430 431 432
                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)
            {
433
                struct msg_size_info * msg_info = (struct msg_size_info*)malloc(sizeof(struct msg_size_info));
434 435
                msg_info->msg_size = qitem->num_bytes;
                msg_info->num_msgs = 1;
436
                msg_info->agg_latency = tw_now(lp) - msg_init_time;
437
                msg_info->avg_latency = msg_info->agg_latency;
438 439
                assert(ns->msg_sz_table);
                qhash_add(ns->msg_sz_table, &(msg_info->msg_size), &(msg_info->hash_link));
440 441 442 443 444 445 446
                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++;
447
                tmp->agg_latency += tw_now(lp) - msg_init_time;  
448
                tmp->avg_latency = (tmp->agg_latency / tmp->num_msgs);
449
//                printf("\n Msg size %lld aggregate latency %f num messages %d ", qitem->num_bytes, tmp->agg_latency, tmp->num_msgs);
450 451 452 453 454 455 456 457
            }
}
static void notify_background_traffic_rc(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
458 459 460
    (void)ns;
    (void)bf;
    (void)m;
461 462 463 464 465
        
    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); 
466 467 468 469 470 471 472 473
}

static void notify_background_traffic(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
474 475 476
        (void)bf;
        (void)m;

477 478 479 480 481 482 483 484 485 486 487 488 489 490 491
        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);

492
            lprintf("\n Other ranks %d ", num_other_ranks);
493 494 495 496 497 498 499 500 501 502 503 504
            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);
505
                m_new = (struct nw_message*)tw_event_data(e);
506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538
                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)
    {
539
//        printf("\n All workloads completed, notifying background traffic ");
540 541 542 543 544 545 546 547 548 549 550 551 552
        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;

553
//        printf("\n Local rank %d notifying neighbor %d ", ns->local_rank, ns->local_rank+1);
554 555 556 557 558 559 560 561 562 563
        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);
564
        m_new = (struct nw_message*)tw_event_data(e); 
565 566 567 568 569 570 571 572 573 574
        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)
{
575 576 577 578
        (void)b;
        (void)msg;
        (void)lp;

579 580 581 582 583 584 585 586 587
        ns->is_finished = 0;
        return;
}
void finish_bckgnd_traffic(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
588 589
        (void)b;
        (void)msg;
590
        ns->is_finished = 1;
591
        lprintf("\n LP %llu completed sending data %lu completed at time %lf ", LLU(lp->gid), ns->gen_data, tw_now(lp));
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
        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;

621 622 623 624 625 626
    if(bf->c1)
    {
        tw_rand_reverse_unif(lp->rng);
    }
    if(bf->c2)
    {
627
        s->prev_switch = m->rc.saved_prev_switch;
628 629 630
        s->saved_perm_dest = s->rc_perm;
        tw_rand_reverse_unif(lp->rng);
    }
631 632 633 634 635 636 637
    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;
    }
        tw_rand_reverse_unif(lp->rng);
638 639 640 641 642 643 644 645 646 647 648 649 650 651

}

/* 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;
652 653
    int intm_dest_id;
    nw_message remote_m;
654 655 656 657 658 659

    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);

660 661 662 663
    /* Find destination */
    int* dest_svr = NULL; 
    int i, length=0;
    switch(s->synthetic_pattern)
664
    {
665 666
        case UNIFORM:
        {
667
            bf->c1 = 1;
668 669 670 671 672 673 674
            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;
675 676 677

        case PERMUTATION:
        {
678 679
            m->rc.saved_prev_switch = s->prev_switch; //for reverse computation

680 681
            length = 1;
            dest_svr = (int*) calloc(1, sizeof(int));
682
            if(s->gen_data - s->prev_switch >= perm_switch_thresh)
683
            {
684
                // printf("%d - %d >= %d\n",s->gen_data,s->prev_switch,perm_switch_thresh);
685
                bf->c2 = 1;
686
                s->prev_switch = s->gen_data; //Amount of data pushed at time when switch initiated
687 688 689 690 691 692 693 694 695 696
                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 + num_clients/2) % num_clients;
                s->rc_perm = s->saved_perm_dest;
                s->saved_perm_dest = dest_svr[0];
            }
            else
                dest_svr[0] = s->saved_perm_dest;
        }
        break;
697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745
        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;
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
    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);
            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; 
        }
773 774
    }
    /* New event after MEAN_INTERVAL */  
775
    tw_stime ts = mean_interval  + tw_rand_exponential(lp->rng, noise); 
776 777 778
    tw_event * e;
    nw_message * m_new;
    e = tw_event_new(lp->gid, ts, lp);
779
    m_new = (struct nw_message*)tw_event_data(e);
780 781
    m_new->msg_type = CLI_BCKGND_GEN;
    tw_event_send(e);
782 783

    free(dest_svr);
784 785 786 787
}

void arrive_syn_tr_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
788 789 790
    (void)bf;
    (void)m;
    (void)lp;
791 792 793 794 795 796 797
//    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)
{
798 799 800
    (void)bf;
    (void)lp;

801
//    printf("\n Data arrived %d total data %ld ", m->fwd.num_bytes, s->syn_data);
802 803 804 805
    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)
806 807 808 809
    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)
810 811 812 813 814 815 816
	{ 
        	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;
	}*/
817
	}
818 819 820 821
    int data = m->fwd.num_bytes;
    s->syn_data += data;
    num_syn_bytes_recvd += data;
}
822
/* Debugging functions, may generate unused function warning */
823
/*static void print_waiting_reqs(uint32_t * reqs, int count)
824
{
825
    lprintf("\n Waiting reqs: %d count", count);
826 827
    int i;
    for(i = 0; i < count; i++ )
828
        lprintf(" %d ", reqs[i]);
829
}*/
830 831 832 833 834 835 836 837 838 839 840 841
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);
842
            //printf(" \n Source %d Dest %d bytes %"PRId64" tag %d ", current->source_rank, current->dest_rank, current->num_bytes, current->tag);
843 844
       }
}
845
static void print_completed_queue(tw_lp * lp, struct qlist_head * head)
846
{
847
//    printf("\n Completed queue: ");
848 849
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
850
      tw_output(lp, "\n");
851 852 853
      qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, completed_requests, ql);
854
            tw_output(lp, " %llu ", current->req_id);
855
       }
856
}
857
static int clear_completed_reqs(nw_state * s,
858
        tw_lp * lp,
859
        unsigned int * reqs, int count)
860
{
861 862 863
    (void)s;
    (void)lp;

864
    int i, matched = 0;
865

866 867 868
    for( i = 0; i < count; i++)
    {
      struct qlist_head * ent = NULL;
869 870 871
      struct completed_requests * current = NULL;
      struct completed_requests * prev = NULL;

872
      int index = 0;
873 874
      qlist_for_each(ent, &s->completed_reqs)
       {
875 876
           if(prev)
           {
877
              rc_stack_push(lp, prev, free, s->matched_reqs);
878 879
              prev = NULL;
           }
880
            
881 882
           current = qlist_entry(ent, completed_requests, ql);
           current->index = index; 
883 884
            if(current->req_id == reqs[i])
            {
885
                ++matched;
886
                qlist_del(&current->ql);
887
                prev = current;
888
            }
889
            ++index;
890
       }
891 892

      if(prev)
893 894 895 896
      {
         rc_stack_push(lp, prev, free, s->matched_reqs);
         prev = NULL;
      }
897
    }
898
    return matched;
899
}
900
static void add_completed_reqs(nw_state * s,
901 902
        tw_lp * lp,
        int count)
903
{
904
    (void)lp;
905
    for(int i = 0; i < count; i++)
906
    {
907 908
       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.
909
       qlist_add(&req->ql, &s->completed_reqs);
910
    }//end for
911
}
912

913 914 915 916 917 918
/* 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);
}

919
static int notify_posted_wait(nw_state* s,
920
        tw_bf * bf, nw_message * m, tw_lp * lp,
921
        unsigned int completed_req)
922
{
923 924
    (void)bf;

925 926
    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;
927

928
    m->fwd.wait_completed = 0;
929

930 931
    if(!wait_elem)
        return 0;
932

933
    int op_type = wait_elem->op_type;
934

935 936 937
    if(op_type == CODES_WK_WAIT &&
            (wait_elem->req_ids[0] == completed_req))
    {
938
            m->fwd.wait_completed = 1;
939 940
            wait_completed = 1;
    }
941 942
    else if(op_type == CODES_WK_WAITALL
            || op_type == CODES_WK_WAITANY
943 944 945 946 947 948
            || op_type == CODES_WK_WAITSOME)
    {
        int i;
        for(i = 0; i < wait_elem->count; i++)
        {
            if(wait_elem->req_ids[i] == completed_req)
949
            {
950
                wait_elem->num_completed++;
951
                if(wait_elem->num_completed > wait_elem->count)
952
                    printf("\n Num completed %d count %d LP %llu ",
953 954
                            wait_elem->num_completed,
                            wait_elem->count,
955
                            LLU(lp->gid));
956 957
//                if(wait_elem->num_completed > wait_elem->count)
//                    tw_lp_suspend(lp, 1, 0);
958

959
                if(wait_elem->num_completed >= wait_elem->count)
960
                {
961
                    if(enable_debug)
962
                        fprintf(workload_log, "\n(%lf) APP ID %d MPI WAITALL COMPLETED AT %llu ", tw_now(lp), s->app_id, LLU(s->nw_id));
963
                    wait_completed = 1;
964
                }
965

966
                m->fwd.wait_completed = 1;
967
            }
968
        }
969
    }
970
    return wait_completed;
971
}
972

973
/* reverse handler of MPI wait operation */
974
static void codes_exec_mpi_wait_rc(nw_state* s, tw_bf * bf, tw_lp* lp, nw_message * m)
975
{
976
   if(bf->c1)
977
    {
978
        completed_requests * qi = (completed_requests*)rc_stack_pop(s->processed_ops);
979 980
        if(m->fwd.found_match == 0)
        {
981
            qlist_add(&qi->ql, &s->completed_reqs);
982
        }
983 984 985 986 987 988 989 990 991 992 993 994 995
        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++;
           }
996
        }
997
        codes_issue_next_event_rc(lp);
998
        return;
999
    }
1000 1001 1002
         struct pending_waits * wait_op = s->wait_op;
         free(wait_op);
         s->wait_op = NULL;
1003
}
1004

1005
/* execute MPI wait operation */
1006
static void codes_exec_mpi_wait(nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
1007
{
1008
    /* check in the completed receives queue if the request ID has already been completed.*/
1009 1010
                
//    printf("\n Wait posted rank id %d ", s->nw_id);
1011
    assert(!s->wait_op);
1012
    unsigned int req_id = mpi_op->u.wait.req_id;
1013

1014
    struct completed_requests* current = NULL;
1015

1016
    struct qlist_head * ent = NULL;
1017
    int index = 0;
1018 1019 1020 1021 1022
    qlist_for_each(ent, &s->completed_reqs)
    {
        current = qlist_entry(ent, completed_requests, ql);
        if(current->req_id == req_id)
        {
1023
            bf->c1=1;
1024
            qlist_del(&current->ql);
1025
            rc_stack_push(lp, current, free, s->processed_ops);
1026
            codes_issue_next_event(lp);
1027
            m->fwd.found_match = index;
1028
            if(s->nw_id == (tw_lpid)TRACK_LP)
1029 1030 1031
            {
                tw_output(lp, "\n wait matched at post %d ", req_id);
                print_completed_queue(lp, &s->completed_reqs);
1032
            }
1033 1034
            return;
        }
1035
        ++index;
1036
    }
1037

1038 1039 1040 1041 1042
    /*if(s->nw_id == (tw_lpid)TRACK_LP)
    {
        tw_output(lp, "\n wait posted %llu ", req_id);
        print_completed_queue(lp, &s->completed_reqs);
    }*/
1043
    /* If not, add the wait operation in the pending 'waits' list. */
1044
    struct pending_waits* wait_op = (struct pending_waits*)malloc(sizeof(struct pending_waits));
1045 1046 1047
    wait_op->op_type = mpi_op->op_type;
    wait_op->req_ids[0] = req_id;
    wait_op->count = 1;
1048 1049
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
1050
    s->wait_op = wait_op;
1051

1052
    return;
1053 1054
}

1055
static void codes_exec_mpi_wait_all_rc(
1056
        nw_state* s,
1057 1058
        tw_bf * bf,
        nw_message * m,
1059
        tw_lp* lp)
1060
{
1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071
  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--;
    }
  }
1072 1073 1074 1075 1076 1077 1078 1079
  if(s->wait_op)
  {
      struct pending_waits * wait_op = s->wait_op;
      free(wait_op);
      s->wait_op = NULL;
  }
  else
  {
1080
      add_completed_reqs(s, lp, m->fwd.num_matched);
1081 1082 1083
      codes_issue_next_event_rc(lp);
  }
  return;
1084
}
1085

1086
static void codes_exec_mpi_wait_all(
1087
        nw_state* s,
1088 1089
        tw_bf * bf,
        nw_message * m,
1090
        tw_lp* lp,
1091
        struct codes_workload_op * mpi_op)
1092
{
1093
  if(enable_debug)
1094
    fprintf(workload_log, "\n MPI WAITALL POSTED AT %llu ", LLU(s->nw_id));
1095

1096 1097 1098 1099 1100 1101 1102 1103
  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;
1104
        s->mpi_wkld_samples[indx].app_id = s->app_id;
1105 1106 1107 1108 1109 1110
        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)
    {
1111
        struct mpi_workload_sample * tmp = (struct mpi_workload_sample*)calloc((MAX_STATS + s->max_arr_size), sizeof(struct mpi_workload_sample));
1112 1113 1114 1115 1116 1117 1118 1119
        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++;
  }
1120
  int count = mpi_op->u.waits.count;
1121 1122
  /* If the count is not less than max wait reqs then stop */
  assert(count < MAX_WAIT_REQS);
1123

1124
  int i = 0, num_matched = 0;
1125
  m->fwd.num_matched = 0;
1126

1127
  /*if(lp->gid == TRACK_LP)
1128
  {
1129
      printf("\n MPI Wait all posted ");
1130
      print_waiting_reqs(mpi_op->u.waits.req_ids, count);
1131 1132
      print_completed_queue(lp, &s->completed_reqs);
  }*/
1133
      /* check number of completed irecvs in the completion queue */
1134 1135
  for(i = 0; i < count; i++)
  {
1136
      unsigned int req_id = mpi_op->u.waits.req_ids[i];
1137 1138 1139 1140
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
      qlist_for_each(ent, &s->completed_reqs)
       {
1141
            current = qlist_entry(ent, struct completed_requests, ql);
1142 1143 1144 1145
            if(current->req_id == req_id)
                num_matched++;
       }
  }
1146

1147
  m->fwd.found_match = num_matched;
1148 1149 1150 1151
  if(num_matched == count)
  {
    /* No need to post a MPI Wait all then, issue next event */
      /* Remove all completed requests from the list */
1152 1153 1154
      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);
1155 1156
      s->wait_op = NULL;
      codes_issue_next_event(lp);
1157 1158
  }
  else
1159 1160
  {
      /* If not, add the wait operation in the pending 'waits' list. */
1161
	  struct pending_waits* wait_op = (struct pending_waits*)malloc(sizeof(struct pending_waits));
1162 1163 1164 1165 1166 1167 1168 1169
	  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;
1170
	  wait_op->start_time = tw_now(lp);
1171
      s->wait_op = wait_op;
1172
  }
1173 1174
  return;
}
1175

1176 1177
/* search for a matching mpi operation and remove it from the list.
 * Record the index in the list from where the element got deleted.
1178
 * Index is used for inserting the element once again in the queue for reverse computation. */
1179
static int rm_matching_rcv(nw_state * ns,
1180
        tw_bf * bf,
1181 1182
        nw_message * m,
        tw_lp * lp,
1183
        mpi_msgs_queue * qitem)
1184 1185
{
    int matched = 0;
1186
    int index = 0;
1187
    int is_rend = 0;
1188 1189
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;
1190

1191 1192
    qlist_for_each(ent, &ns->pending_recvs_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
1193 1194
        if(//(qi->num_bytes == qitem->num_bytes)
                //&& 
1195
               ((qi->tag == qitem->tag) || qi->tag == -1)
1196
                && ((qi->source_rank == qitem->source_rank) || qi->source_rank == -1))
1197 1198
        {
            matched = 1;
1199
            qi->num_bytes = qitem->num_bytes;
1200 1201
            break;
        }
1202
        ++index;
1203
    }
1204

1205 1206
    if(matched)
    {
1207 1208 1209 1210 1211 1212 1213
        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
1214 1215
             * the data * (only works in sequential mode)*/
            bf->c10 = 1;
1216 1217
            is_rend = 1;
            send_ack_back(ns, bf, m, lp, qitem, qi->req_id);
1218
        }
1219 1220
        else
        {
1221
            bf->c12 = 1;
1222 1223 1224 1225
            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)
1226
        {
1227
            bf->c9 = 1;
1228 1229 1230 1231
            /*if(ns->nw_id == (tw_lpid)TRACK_LP)
            {
                printf("\n Completed irecv req id %d ", qi->req_id);
            }*/
1232
            update_completed_queue(ns, bf, m, lp, qi->req_id);
1233
        }
1234 1235 1236
        else if(qi->op_type == CODES_WK_RECV && !is_rend)
        {
            bf->c8 = 1;
1237
            codes_issue_next_event(lp);
1238
        }
1239

1240
        qlist_del(&qi->ql);
1241

1242
        rc_stack_push(lp, qi, free, ns->processed_ops);
1243
        return index;
1244 1245 1246 1247
    }
    return -1;
}

1248
static int rm_matching_send(nw_state * ns,
1249 1250 1251
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp, mpi_msgs_queue * qitem)
1252 1253 1254 1255 1256
{
    int matched = 0;
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;

1257
    int index = 0;
1258 1259
    qlist_for_each(ent, &ns->arrival_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
1260 1261 1262
        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)
1263 1264
                && ((qi->source_rank == qitem->source_rank) || qitem->source_rank == -1))
        {
1265
            qitem->num_bytes = qi->num_bytes;
1266 1267 1268
            matched = 1;
            break;
        }
1269
        ++index;
1270 1271 1272 1273
    }

    if(matched)
    {
1274 1275 1276
        if(enable_msg_tracking && (qi->num_bytes < EAGER_THRESHOLD))
            update_message_size(ns, lp, bf, m, qi, 1, 0);
        
1277 1278
        m->fwd.matched_req = qitem->req_id;
        int is_rend = 0;
1279 1280 1281 1282
        if(qitem->num_bytes >= EAGER_THRESHOLD)
        {
            /* Matching receive found, need to notify the sender to transmit
             * the data */
1283
            bf->c10 = 1;
1284 1285
            is_rend = 1;
            send_ack_back(ns, bf, m, lp, qi, qitem->req_id);
1286
        }