model-net-mpi-replay.c 93.8 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 1024
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
#define COL_TAG 1235
#define BAR_TAG 1234
33

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

37
static unsigned long perm_switch_thresh = 8388608;
38

39
/* NOTE: Message tracking works in sequential mode only! */
40
static int debug_cols = 0;
41
static int synthetic_pattern = 1;
42 43 44 45 46
/* 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;
47
static unsigned long long max_gen_data = 1310720;
48
tw_lpid TRACK_LP = -1;
49
int nprocs = 0;
50
static double total_syn_data = 0;
51
static int unmatched = 0;
52
char workload_type[128];
53
char workload_name[128];
54 55 56 57
char workload_file[8192];
char offset_file[8192];
static int wrkld_id;
static int num_net_traces = 0;
58
static int priority_type = 0;
59
static int num_dumpi_traces = 0;
60
static int64_t EAGER_THRESHOLD = 8192;
61

62
static long num_ops = 0;
63
static int upper_threshold = 1048576;
64
static int alloc_spec = 0;
65 66
static tw_stime self_overhead = 10.0;
static tw_stime mean_interval = 100000;
67
static int payload_sz = 1024;
68 69

/* Doing LP IO*/
70
static char * params = NULL;
71
static char lp_io_dir[256] = {'\0'};
72
static char sampling_dir[32] = {'\0'};
73
static char mpi_msg_dir[32] = {'\0'};
74 75 76 77
static lp_io_handle io_handle;
static unsigned int lp_io_use_suffix = 0;
static int do_lp_io = 0;

78 79 80 81
/* variables for loading multiple applications */
char workloads_conf_file[8192];
char alloc_file[8192];
int num_traces_of_job[5];
82 83 84
tw_stime soft_delay_mpi = 2500;
tw_stime nic_delay = 1000;
tw_stime copy_per_byte_eager = 0.55;
85 86 87 88 89
char file_name_of_job[5][8192];

struct codes_jobmap_ctx *jobmap_ctx;
struct codes_jobmap_params_list jobmap_p;

90 91
/* Variables for Cortex Support */
/* Matthieu's additions start */
92
#ifdef ENABLE_CORTEX_PYTHON
93 94 95
static char cortex_file[512] = "\0";
static char cortex_class[512] = "\0";
static char cortex_gen[512] = "\0";
96
#endif
97 98
/* Matthieu's additions end */

99 100
typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
101
typedef unsigned int dumpi_req_id;
102 103

static int net_id = 0;
104
static float noise = 1.0;
105 106 107
static int num_nw_lps = 0, num_mpi_lps = 0;

static int num_syn_clients;
108
static int syn_type = 0;
109

110
FILE * workload_log = NULL;
111
FILE * msg_size_log = NULL;
112 113 114 115
FILE * workload_agg_log = NULL;
FILE * workload_meta_log = NULL;

static uint64_t sample_bytes_written = 0;
116

117 118
unsigned long long num_bytes_sent=0;
unsigned long long num_bytes_recvd=0;
119

120 121
unsigned long long num_syn_bytes_sent = 0;
unsigned long long num_syn_bytes_recvd = 0;
122

123 124 125 126 127 128
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;
129 130 131
static int enable_sampling = 0;
static double sampling_interval = 5000000;
static double sampling_end_time = 3000000000;
132
static int enable_debug = 0;
133

134
/* set group context */
135 136 137 138 139 140 141 142 143 144 145
struct codes_mctx mapping_context;
enum MAPPING_CONTEXTS
{
    GROUP_RATIO=1,
    GROUP_RATIO_REVERSE,
    GROUP_DIRECT,
    GROUP_MODULO,
    GROUP_MODULO_REVERSE,
    UNKNOWN
};
static int map_ctxt = GROUP_MODULO;
146

147
/* MPI_OP_GET_NEXT is for getting next MPI operation when the previous operation completes.
148
* 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.
149 150 151 152 153 154 155
* 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,
156 157 158 159 160 161
    MPI_REND_ARRIVED,
    MPI_REND_ACK_ARRIVED,
    CLI_BCKGND_FIN,
    CLI_BCKGND_ARRIVE,
    CLI_BCKGND_GEN,
    CLI_NBR_FINISH,
162 163
};

164 165 166 167 168 169
/* 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 */
170 171
    STENCIL = 4, /* sends message to 4 nearby neighbors */
    PERMUTATION = 5
172
};
173 174 175 176
struct mpi_workload_sample
{
    /* Sampling data */
    int nw_id;
177
    int app_id;
178 179 180 181 182
    unsigned long num_sends_sample;
    unsigned long num_bytes_sample;
    unsigned long num_waits_sample;
    double sample_end_time;
};
183 184 185 186 187 188 189
/* 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;
190
    int64_t num_bytes;
191
    int64_t seq_id;
192 193 194 195 196 197 198 199
    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
{
200
	unsigned int req_id;
201
    struct qlist_head ql;
202
    int index;
203 204 205 206 207 208
};

/* for wait operations, store the pending operation and number of completed waits so far. */
struct pending_waits
{
    int op_type;
209
    unsigned int req_ids[MAX_WAIT_REQS];
210
	int num_completed;
211 212
	int count;
    tw_stime start_time;
213 214 215
    struct qlist_head ql;
};

216 217 218 219 220 221
struct msg_size_info
{
    int64_t msg_size;
    int num_msgs;
    tw_stime agg_latency;
    tw_stime avg_latency;
222
    struct qhash_head  hash_link;
223 224
    struct qlist_head ql; 
};
225 226 227 228 229 230 231 232 233 234
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;
235 236
    int app_id;
    int local_rank;
237

238
    int synthetic_pattern;
239 240 241
    int is_finished;
    int neighbor_completed;

242
    struct rc_stack * processed_ops;
243
    struct rc_stack * processed_wait_op;
244
    struct rc_stack * matched_reqs;
245
//    struct rc_stack * indices;
246 247 248 249 250 251 252 253 254 255

    /* 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;

256

257 258
	/* time spent by the LP in executing the app trace*/
	double start_time;
259 260 261 262 263 264 265 266 267

    double col_time;

    double reduce_time;
    int num_reduce;

    double all_reduce_time;
    int num_all_reduce;

268 269 270 271 272
	double elapsed_time;
	/* time spent in compute operations */
	double compute_time;
	/* time spent in message send/isend */
	double send_time;
273 274
    /* max time for synthetic traffic message */
    double max_time;
275 276 277 278 279 280 281 282 283 284
	/* 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;
285

286
    tw_stime cur_interval_end;
287
    
288 289
    /* Pending wait operation */
    struct pending_waits * wait_op;
290

291 292 293 294 295 296
    /* Message size latency information */
    struct qhash_table * msg_sz_table;
    struct qlist_head msg_sz_list;

    /* quick hash for maintaining message latencies */

297 298
    unsigned long long num_bytes_sent;
    unsigned long long num_bytes_recvd;
299

300 301
    unsigned long long syn_data;
    unsigned long long gen_data;
302 303
  
    unsigned long prev_switch;
304
    int saved_perm_dest;
305 306
    unsigned long rc_perm;

307 308 309 310
    /* For sampling data */
    int sampling_indx;
    int max_arr_size;
    struct mpi_workload_sample * mpi_wkld_samples;
311
    char output_buf[512];
312
    char col_stats[64];
313 314 315 316
};

/* data for handling reverse computation.
* saved_matched_req holds the request ID of matched receives/sends for wait operations.
317
* 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.
318 319 320
* 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
{
321
   // forward message handler
322
   int msg_type;
323
   int op_type;
324
   model_net_event_return event_rc;
325
   struct codes_workload_op * mpi_op;
326

327 328 329
   struct
   {
       tw_lpid src_rank;
330
       int dest_rank;
331
       int64_t num_bytes;
332 333 334 335 336
       int num_matched;
       int data_type;
       double sim_start_time;
       // for callbacks - time message was received
       double msg_send_time;
337
       unsigned int req_id;
338
       int matched_req;
339
       int tag;
340
       int app_id;
341 342
       int found_match;
       short wait_completed;
343
       short rend_send;
344 345 346
   } fwd;
   struct
   {
347
       int saved_perm;
348 349 350 351
       double saved_send_time;
       double saved_recv_time;
       double saved_wait_time;
       double saved_delay;
352
       int64_t saved_num_bytes;
353
       int saved_syn_length;
354
       unsigned long saved_prev_switch;
355
   } rc;
356 357
};

358
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);
359 360

static void send_ack_back_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
361 362
/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
363
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op, int is_rend);
364 365
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
366
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
367 368
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
369
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp);
370 371
/* execute the computational delay */
static void codes_exec_comp_delay(
372
        nw_state* s, tw_bf *bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
373 374 375 376 377 378 379 380 381 382 383 384 385 386 387
/* 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 */
388 389 390 391 392 393 394 395
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);
396 397 398 399 400 401 402 403 404 405 406 407 408 409 410
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);

411
/*static void update_message_size_rc(
412 413 414 415
        struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
416
{*/
417
/*TODO: Complete reverse handler */
418
/*    (void)ns;
419 420 421
    (void)lp;
    (void)bf;
    (void)m;
422
}*/
423 424 425 426 427 428 429 430 431 432
/* 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)
{
433 434 435
            (void)bf;
            (void)is_eager;

436 437 438
            struct qhash_head * hash_link = NULL;
            tw_stime msg_init_time = qitem->req_init_time;
        
439
            if(ns->msg_sz_table == NULL)
440 441 442 443 444 445 446 447 448 449
                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)
            {
450
                struct msg_size_info * msg_info = (struct msg_size_info*)malloc(sizeof(struct msg_size_info));
451 452
                msg_info->msg_size = qitem->num_bytes;
                msg_info->num_msgs = 1;
453
                msg_info->agg_latency = tw_now(lp) - msg_init_time;
454
                msg_info->avg_latency = msg_info->agg_latency;
455 456
                assert(ns->msg_sz_table);
                qhash_add(ns->msg_sz_table, &(msg_info->msg_size), &(msg_info->hash_link));
457 458 459 460 461 462 463
                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++;
464
                tmp->agg_latency += tw_now(lp) - msg_init_time;  
465
                tmp->avg_latency = (tmp->agg_latency / tmp->num_msgs);
466
//                printf("\n Msg size %lld aggregate latency %f num messages %d ", qitem->num_bytes, tmp->agg_latency, tmp->num_msgs);
467 468 469 470 471 472 473 474
            }
}
static void notify_background_traffic_rc(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
475 476 477
    (void)ns;
    (void)bf;
    (void)m;
478 479 480 481 482
        
    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); 
483 484 485 486 487 488 489 490
}

static void notify_background_traffic(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
491 492 493
        (void)bf;
        (void)m;

494 495 496 497 498 499 500 501 502 503 504 505 506 507 508
        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);

509
            lprintf("\n Other ranks %d ", num_other_ranks);
510
            tw_stime ts = (1.1 * g_tw_lookahead) + tw_rand_exponential(lp->rng, noise);
511 512 513 514 515 516 517 518 519 520 521
            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);
522
                m_new = (struct nw_message*)tw_event_data(e);
523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555
                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)
    {
556
//        printf("\n All workloads completed, notifying background traffic ");
557 558 559 560 561 562 563 564 565 566 567 568 569
        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;

570
//        printf("\n Local rank %d notifying neighbor %d ", ns->local_rank, ns->local_rank+1);
571
        tw_stime ts = (1.1 * g_tw_lookahead) + tw_rand_exponential(lp->rng, noise);
572 573 574 575 576 577 578 579 580
        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);
581
        m_new = (struct nw_message*)tw_event_data(e); 
582 583 584 585 586 587 588 589 590 591
        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)
{
592 593 594 595
        (void)b;
        (void)msg;
        (void)lp;

596 597 598 599 600 601 602 603 604
        ns->is_finished = 0;
        return;
}
void finish_bckgnd_traffic(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
605 606
        (void)b;
        (void)msg;
607
        ns->is_finished = 1;
608
        lprintf("\n LP %llu completed sending data %lu completed at time %lf ", LLU(lp->gid), ns->gen_data, tw_now(lp));
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
        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;

639 640 641 642 643 644
    if(bf->c1)
    {
        tw_rand_reverse_unif(lp->rng);
    }
    if(bf->c2)
    {
645
        s->prev_switch = m->rc.saved_prev_switch;
646
        s->saved_perm_dest = m->rc.saved_perm;
647 648
        tw_rand_reverse_unif(lp->rng);
    }
649 650 651 652 653
    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;
654
        s->num_bytes_sent -= payload_sz;
655 656
    }
        tw_rand_reverse_unif(lp->rng);
657
        s->num_sends--;
658

659 660
     if(bf->c5)
         s->is_finished = 0;
661 662 663 664 665 666 667 668 669 670 671 672 673
}

/* 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;
674 675
    int intm_dest_id;
    nw_message remote_m;
676 677 678 679 680 681

    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);

682 683 684 685
    /* Find destination */
    int* dest_svr = NULL; 
    int i, length=0;
    switch(s->synthetic_pattern)
686
    {
687 688
        case UNIFORM:
        {
689
            bf->c1 = 1;
690 691 692 693 694 695 696
            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;
697 698 699

        case PERMUTATION:
        {
700 701
            m->rc.saved_prev_switch = s->prev_switch; //for reverse computation

702 703
            length = 1;
            dest_svr = (int*) calloc(1, sizeof(int));
704
            if(s->gen_data - s->prev_switch >= perm_switch_thresh)
705
            {
706
                // printf("%d - %d >= %d\n",s->gen_data,s->prev_switch,perm_switch_thresh);
707
                bf->c2 = 1;
708
                m->rc.saved_prev_switch = s->prev_switch;
709
                s->prev_switch = s->gen_data; //Amount of data pushed at time when switch initiated
710 711 712
                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;
713 714
                /* TODO: Fix random number generation code */
                m->rc.saved_perm = s->saved_perm_dest;
715
                s->saved_perm_dest = dest_svr[0];
716
                assert(s->saved_perm_dest != s->local_rank);
717 718 719
            }
            else
                dest_svr[0] = s->saved_perm_dest;
720 721

            assert(dest_svr[0] != s->local_rank);
722 723
        }
        break;
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
        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;
773

774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792
    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);
793
            m->event_rc = model_net_event(net_id, "medium", global_dest_id, payload_sz, 0.0, 
794 795 796 797
                    sizeof(nw_message), (const void*)&remote_m, 
                    0, NULL, lp);
            
            s->gen_data += payload_sz;
798
            s->num_bytes_sent += payload_sz;
799 800
            num_syn_bytes_sent += payload_sz; 
        }
801
    }
802 803
    s->num_sends++;

804
    /* New event after MEAN_INTERVAL */  
805
    tw_stime ts = mean_interval  + tw_rand_exponential(lp->rng, noise); 
806 807 808
    tw_event * e;
    nw_message * m_new;
    e = tw_event_new(lp->gid, ts, lp);
809
    m_new = (struct nw_message*)tw_event_data(e);
810 811
    m_new->msg_type = CLI_BCKGND_GEN;
    tw_event_send(e);
812 813 814 815 816 817
    
    if(s->gen_data >= max_gen_data)
    {
        bf->c5 = 1;
        s->is_finished = 1;
    }
818
    free(dest_svr);
819 820 821 822
}

void arrive_syn_tr_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
823 824 825
    (void)bf;
    (void)m;
    (void)lp;
826
//    printf("\n Data arrived %d total data %ld ", m->fwd.num_bytes, s->syn_data);
827
    s->num_recvs--;
828 829 830
    int data = m->fwd.num_bytes;
    s->syn_data -= data;
    num_syn_bytes_recvd -= data;
831 832
    s->num_bytes_recvd -= data;
    s->send_time = m->rc.saved_send_time;
833 834 835
}
void arrive_syn_tr(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
836 837 838
    (void)bf;
    (void)lp;

839
//    printf("\n Data arrived %d total data %ld ", m->fwd.num_bytes, s->syn_data);
840 841 842 843
    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)
844 845 846 847
    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)
848 849 850 851 852 853 854
	{ 
        	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;
	}*/
855
	}
856 857 858 859 860 861
    m->rc.saved_send_time = s->send_time;
    if((tw_now(lp) - m->fwd.sim_start_time) > s->max_time)
        s->max_time = tw_now(lp) - m->fwd.sim_start_time;

    s->send_time += (tw_now(lp) - m->fwd.sim_start_time);
    s->num_recvs++;
862 863
    int data = m->fwd.num_bytes;
    s->syn_data += data;
864
    s->num_bytes_recvd += data;
865 866
    num_syn_bytes_recvd += data;
}
867
/* Debugging functions, may generate unused function warning */
868
/*static void print_waiting_reqs(uint32_t * reqs, int count)
869
{
870
    lprintf("\n Waiting reqs: %d count", count);
871 872
    int i;
    for(i = 0; i < count; i++ )
873
        lprintf(" %d ", reqs[i]);
874
}*/
875 876 877 878 879 880 881 882 883 884 885 886
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);
887
            //printf(" \n Source %d Dest %d bytes %"PRId64" tag %d ", current->source_rank, current->dest_rank, current->num_bytes, current->tag);
888 889
       }
}
890
static void print_completed_queue(tw_lp * lp, struct qlist_head * head)
891
{
892
//    printf("\n Completed queue: ");
893 894
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
895
      tw_output(lp, "\n");
896 897 898
      qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, completed_requests, ql);
899
            tw_output(lp, " %llu ", current->req_id);
900
       }
901
}
902
static int clear_completed_reqs(nw_state * s,
903
        tw_lp * lp,
904
        unsigned int * reqs, int count)
905
{
906 907 908
    (void)s;
    (void)lp;

909
    int i, matched = 0;
910

911 912 913
    for( i = 0; i < count; i++)
    {
      struct qlist_head * ent = NULL;
914 915 916
      struct completed_requests * current = NULL;
      struct completed_requests * prev = NULL;

917
      int index = 0;
918 919
      qlist_for_each(ent, &s->completed_reqs)
       {
920 921
           if(prev)
           {
922
              rc_stack_push(lp, prev, free, s->matched_reqs);
923 924
              prev = NULL;
           }
925
            
926 927
           current = qlist_entry(ent, completed_requests, ql);
           current->index = index; 
928 929
            if(current->req_id == reqs[i])
            {
930
                ++matched;
931
                qlist_del(&current->ql);
932
                prev = current;
933
            }
934
            ++index;
935
       }
936 937

      if(prev)
938 939 940 941
      {
         rc_stack_push(lp, prev, free, s->matched_reqs);
         prev = NULL;
      }
942
    }
943
    return matched;
944
}
945
static void add_completed_reqs(nw_state * s,
946 947
        tw_lp * lp,
        int count)
948
{
949
    (void)lp;
950
    for(int i = 0; i < count; i++)
951
    {
952 953
       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.
954
       qlist_add(&req->ql, &s->completed_reqs);
955
    }//end for
956
}
957

958 959 960 961 962 963
/* 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);
}

964
static int notify_posted_wait(nw_state* s,
965
        tw_bf * bf, nw_message * m, tw_lp * lp,
966
        unsigned int completed_req)
967
{
968 969
    (void)bf;

970 971
    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;
972

973
    m->fwd.wait_completed = 0;
974

975 976
    if(!wait_elem)
        return 0;
977

978
    int op_type = wait_elem->op_type;
979

980 981 982
    if(op_type == CODES_WK_WAIT &&
            (wait_elem->req_ids[0] == completed_req))
    {
983
            m->fwd.wait_completed = 1;
984 985
            wait_completed = 1;
    }
986 987
    else if(op_type == CODES_WK_WAITALL
            || op_type == CODES_WK_WAITANY
988 989 990 991 992 993
            || op_type == CODES_WK_WAITSOME)
    {
        int i;
        for(i = 0; i < wait_elem->count; i++)
        {
            if(wait_elem->req_ids[i] == completed_req)
994
            {
995
                wait_elem->num_completed++;
996
                if(wait_elem->num_completed > wait_elem->count)
997
                    printf("\n Num completed %d count %d LP %llu ",
998 999
                            wait_elem->num_completed,
                            wait_elem->count,
1000
                            LLU(lp->gid));
1001 1002
//                if(wait_elem->num_completed > wait_elem->count)
//                    tw_lp_suspend(lp, 1, 0);
1003

1004
                if(wait_elem->num_completed >= wait_elem->count)
1005
                {
1006
                    if(enable_debug)
1007
                        fprintf(workload_log, "\n(%lf) APP ID %d MPI WAITALL COMPLETED AT %llu ", tw_now(lp), s->app_id, LLU(s->nw_id));
1008
                    wait_completed = 1;
1009
                }
1010

1011
                m->fwd.wait_completed = 1;
1012
            }
1013
        }
1014
    }
1015
    return wait_completed;
1016
}
1017

1018
/* reverse handler of MPI wait operation */
1019
static void codes_exec_mpi_wait_rc(nw_state* s, tw_bf * bf, tw_lp* lp, nw_message * m)
1020
{
1021
   if(bf->c1)
1022
    {
1023
        completed_requests * qi = (completed_requests*)rc_stack_pop(s->processed_ops);
1024 1025
        if(m->fwd.found_match == 0)
        {
1026
            qlist_add(&qi->ql, &s->completed_reqs);
1027
        }
1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040
        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++;
           }
1041
        }
1042
        codes_issue_next_event_rc(lp);
1043
        return;
1044
    }
1045 1046 1047
         struct pending_waits * wait_op = s->wait_op;
         free(wait_op);
         s->wait_op = NULL;
1048
}
1049

1050
/* execute MPI wait operation */
1051
static void codes_exec_mpi_wait(nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
1052
{
1053
    /* check in the completed receives queue if the request ID has already been completed.*/
1054 1055
                
//    printf("\n Wait posted rank id %d ", s->nw_id);
1056
    assert(!s->wait_op);
1057
    unsigned int req_id = mpi_op->u.wait.req_id;
1058

1059
    struct completed_requests* current = NULL;
1060

1061
    struct qlist_head * ent = NULL;
1062
    int index = 0;
1063 1064 1065 1066 1067
    qlist_for_each(ent, &s->completed_reqs)
    {
        current = qlist_entry(ent, completed_requests, ql);
        if(current->req_id == req_id)
        {
1068
            bf->c1=1;
1069
            qlist_del(&current->ql);
1070
            rc_stack_push(lp, current, free, s->processed_ops);
1071
            codes_issue_next_event(lp);
1072
            m->fwd.found_match = index;
1073
            if(s->nw_id == (tw_lpid)TRACK_LP)
1074 1075 1076
            {
                tw_output(lp, "\n wait matched at post %d ", req_id);
                print_completed_queue(lp, &s->completed_reqs);
1077
            }
1078 1079
            return;
        }
1080
        ++index;
1081
    }
1082

1083 1084 1085 1086 1087
    /*if(s->nw_id == (tw_lpid)TRACK_LP)
    {
        tw_output(lp, "\n wait posted %llu ", req_id);
        print_completed_queue(lp, &s->completed_reqs);
    }*/
1088
    /* If not, add the wait operation in the pending 'waits' list. */
1089
    struct pending_waits* wait_op = (struct pending_waits*)malloc(sizeof(struct pending_waits));
1090 1091 1092
    wait_op->op_type = mpi_op->op_type;
    wait_op->req_ids[0] = req_id;
    wait_op->count = 1;
1093 1094
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
1095
    s->wait_op = wait_op;
1096

1097
    return;
1098 1099
}

1100
static void codes_exec_mpi_wait_all_rc(
1101
        nw_state* s,
1102 1103
        tw_bf * bf,
        nw_message * m,
1104
        tw_lp* lp)
1105
{
1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116
  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--;
    }
  }
1117 1118 1119 1120 1121 1122 1123 1124
  if(s->wait_op)
  {
      struct pending_waits * wait_op = s->wait_op;
      free(wait_op);
      s->wait_op = NULL;
  }
  else
  {
1125
      add_completed_reqs(s, lp, m->fwd.num_matched);
1126 1127 1128
      codes_issue_next_event_rc(lp);
  }
  return;
1129
}
1130

1131
static void codes_exec_mpi_wait_all(
1132
        nw_state* s,
1133 1134
        tw_bf * bf,
        nw_message * m,
1135
        tw_lp* lp,
1136
        struct codes_workload_op * mpi_op)
1137
{
1138
  if(enable_debug)
1139
    fprintf(workload_log, "\n MPI WAITALL POSTED AT %llu ", LLU(s->nw_id));
1140

1141 1142 1143 1144 1145 1146 1147 1148
  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;
1149
        s->mpi_wkld_samples[indx].app_id = s->app_id;
1150 1151 1152 1153 1154 1155
        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)
    {
1156
        struct mpi_workload_sample * tmp = (struct mpi_workload_sample*)calloc((MAX_STATS + s->max_arr_size), sizeof(struct mpi_workload_sample));
1157 1158 1159 1160 1161 1162 1163 1164
        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++;
  }
1165
  int count = mpi_op->u.waits.count;
1166 1167
  /* If the count is not less than max wait reqs then stop */
  assert(count < MAX_WAIT_REQS);
1168

1169
  int i = 0, num_matched = 0;
1170
  m->fwd.num_matched = 0;
1171

1172
  /*if(lp->gid == TRACK_LP)
1173
  {
1174
      printf("\n MPI Wait all posted ");
1175
      print_waiting_reqs(mpi_op->u.waits.req_ids, count);
1176 1177
      print_completed_queue(lp, &s->completed_reqs);
  }*/
1178
      /* check number of completed irecvs in the completion queue */
1179 1180
  for(i = 0; i < count; i++)
  {
1181
      unsigned int req_id = mpi_op->u.waits.req_ids[i];
1182 1183 1184 1185
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
      qlist_for_each(ent, &s->completed_reqs)
       {
1186
            current = qlist_entry(ent, struct completed_requests, ql);
1187 1188 1189 1190
            if(current->req_id == req_id)
                num_matched++;
       }
  }
1191

1192
  m->fwd.found_match = num_matched;
1193 1194 1195 1196
  if(num_matched == count)
  {
    /* No need to post a MPI Wait all then, issue next event */
      /* Remove all completed requests from the list */
1197 1198 1199
      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);
1200 1201
      s->wait_op = NULL;
      codes_issue_next_event(lp);
1202 1203
  }
  else
1204 1205
  {
      /* If not, add the wait operation in the pending 'waits' list. */
1206
	  struct pending_waits* wait_op = (struct pending_waits*)malloc(sizeof(struct pending_waits));
1207 1208 1209 1210 1211 1212 1213 1214
	  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;
1215
	  wait_op->start_time = tw_now(lp);
1216
      s->wait_op = wait_op;
1217
  }
1218 1219
  return;
}
1220

1221 1222
/* search for a matching mpi operation and remove it from the list.
 * Record the index in the list from where the element got deleted.
1223
 * Index is used for inserting the element once again in the queue for reverse computation. */
1224
static int rm_matching_rcv(nw_state * ns,
1225
        tw_bf * bf,
1226 1227
        nw_message * m,
        tw_lp * lp,
1228
        mpi_msgs_queue * qitem)
1229 1230
{
    int matched = 0;
1231
    int index = 0;
1232
    int is_rend = 0;
1233 1234
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;
1235

1236 1237
    qlist_for_each(ent, &ns->pending_recvs_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
1238 1239
        if(//(qi->num_bytes == qitem->num_bytes)
                //&& 
1240
               ((qi->tag == qitem->tag) || qi->tag == -1)
1241
                && ((qi->source_rank == qitem->source_rank) || qi->source_rank == -1))
1242 1243
        {
            matched = 1;
1244
            qi->num_bytes = qitem->num_bytes;
1245 1246
            break;
        }
1247
        ++index;
1248
    }
1249

1250 1251
    if(matched)
    {
1252 1253 1254 1255 1256 1257 1258
        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
1259 1260
             * the data * (only works in sequential mode)*/
            bf->c10 = 1;
1261 1262
            is_rend = 1;
            send_ack_back(ns, bf, m, lp, qitem, qi->req_id);
Misbah Mubarak's avatar